ããã¯äœ
- è€æ°ã®ç¹èš±çã®æç« ããç¹å®ã®æ¯éå£ã«ãããäºãã®é¡äŒŒåºŠããå ã«ïŒæ¬¡å 空éäžã«é 眮ãïŒã€ã³ã¿ã©ã¯ãã£ãã«é²èŠ§ã»æ¢çŽ¢çããŒã¿åæãªã©è¡ãã³ãŒãã§ãïŒ
- ã¯ã©ã¹ã¿ãŒè§£æã«ããå šäœæè¡åå俯ç°ã«å ãïŒåå¥ç¢ºèªïŒæ€çŽ¢ïŒæè¡èŠçŽ çãããã¯ãŒã¯åïŒç¹èš±çæïŒæ§ã ãªåæãè¡ãã³ãŒããšèšã£ãŠãè¯ãã§ãããïŒ
- éèŠãªç¹åŸŽã¯3ç¹ïŒ
-- 俯ç°å³ã§ãç¹èš±éã®çžå¯Ÿçãªäœçœ®ããææ¡ãã
-- ãããã¯ãŒã¯ã§ãèŠçŽ éã®å ·äœçãªã€ãªããããæç¢ºãšãã
-- ç¹èš±çæã§ãååšããªãããŒã¿ã®è£å®ãããã
åºç€ç ç©¶ã®ç¹èš±ãªã©ã«åºã¥ãææãªå¿çšéçºã®å ·äœçäŸç€ºãšãã®çšŒãæ¹ãªã©ã®ããšãšãªãæªæ¥äºæž¬ãªã©è¡ããŸã
ïŒæè¡ã¹ããŒãžããšã«ç°ãªãææ³ãçšããŸãã
èèœæè¡ã¹ããŒãžãæè¡å éã¹ããŒãžã§ã¯ã俯ç°å³ãã確èªå¯èœãªæ§é ã«åºã¥ããã¯ã€ãã¹ããŒã¹ã»ãã«ãŒãªãŒã·ã£ã³ã®çºçã®ç¢ºèªãšãããã¯ãŒã¯ã®ãšããžãããŒãã®çæç¢ºèªãè¡ããåºç€ç¹èš±ãªã©ãæœåºããM&Aæ¢çŽ¢ããã³ãã£ãŒã®æè¡ã®ç«ã¡äžããæ¢çŽ¢ãæè³å æ¢çŽ¢ãè¡ããŸãã
æè¡åæ»æã§ã¯ã俯ç°å³ç¢ºèªãšãããã¯ãŒã¯ç¢ºèªãè¡ããææãªå¿çšç¹èš±ãªã©ãæœåºããæè¡ã·ããã®ãŠã©ããã³ã°ããããæ¢çŽ¢ã倧åŠãªã©ã«ããæè¡ææºäŒæ¥åè£æ¢çŽ¢ãè¡ããŸãã
æè¡è¡°éæã§ã¯ã俯ç°å³ç¢ºèªãšãããã¯ãŒã¯ç¢ºèªãè¡ããã³ã¢ãã£ãã£ã«ããæè¡ã®ç¡æå³åãšåºé¡äŸ¡å€ã®åªå€±ãèŠæ¥µããäžèŠæè¡ã®éžå¥ãè¡ããæè¡éçºãããµãŒãã¹ãªã©ãžã®ç§»è¡å€æã奿è¡ãžã®æè³åæ¿å€æãªã©ãè¡ããŸãã
ããããã®ææã«ãããŠããã¯ã€ãã¹ããŒã¹ãã¯ã©ã¹ã¿ã®å¢çããšããžãªã©ãç¹å®ã®åº§æšã«äŸ¡å€ãèŠåºãããªãã座æšãçšããç¹èš±çæãè¡ããç®çé åã§ã䟡å€ã®ããã¢ã€ãã£ã¢ãã®å ·äœåãªã©ãè¡ããŸããïŒïŒ - ããŠã¹ãªãŒããŒã§ããããã®ç¹èš±ã®æŠèŠãèŠãããšãã§ãïŒGoogle patentã§è©²åœã®ç¹èš±ãéãããšãã§ããŸãïŒ
- å·¥å€«æ¬¡ç¬¬ã§æ§ã ãªç®çã»çšéã«çšããããšãã§ããã§ãããïŒ
- æç« ã¯ãã¯ãã«ãšããŠæ±ããŸããæ§ã ãªæç« ãã¯ãã«ã®çææ¹æ³ããããŸãããåºæ¬çã«ã¯ãCompositional Distributional Semantics (CDS)ïŒæã®çµ±èªæ§é ãèæ ®ããèªã®æå³ãã¯ãã«ãšãªãããã«, æç« ãã¯ãã«ãçæããŠããŸããtransformerã屿çãªååžä»®èª¬ãattentionã§æ¡åŒµããŠããã€ã¡ãŒãžãšãããªãã°ãæ¬ææ³ã¯æ¯éå£ã«éå®ããååžä»®èª¬ããattentionã§ã¯é¿ãããã䞻芳çãªæèãã§ããã ã䜿ããã«å®¢èгçãªæèã®ã¿ããèšç®èšèªåŠïŒçµ±èªãtfidfãªã©çµ±èšïŒãšããæŒç¹¹ãå©çšããããšã§æ¡åŒµããŠããã€ã¡ãŒãžã§ãã
远èšã¡ã¢
-
远èšãç¹°ãè¿ããŠããæŽåæ§ãåããŠãããéåžžã«èªã¿ã«ããç¶æ ã§ãïŒè¿æ¥äžã«åæŽçããŸãïŒ
-
æè¡ã ãã§ã¯ãªãæ¹æ³è«ãéèŠãšãªããŸããåŸãåãã«æ€èšŒããååãã«äºæž¬ããããšãéèŠãšãªãã§ããããçŸåšæ€èšŒäžã§ãã
-
ãé¡ãããã ããŸããšå©ãããŸããåŸãåãæ€èšŒã§ã¯ã©ããããããã£ããã®ãæ£æçã«éžãã§èšãããããããªã屿§ããããšããã§ãã
-
å人çã«ã¯ãäŒç€Ÿã®æ¹éã«è¿œåŸããçµæãåºãã ãã®ãæ©å«äŒºãã®ãããªããŒã¿åæã«éåžžã«å€§ããªå±æ©æãèŠããŠãããããŒã¿ãµã€ãšã³ã¹ã¯ãŸãçŸå Žã®ãã®ã§ããã¹ããä»äžåºŠèµãåãçŽãã¹ãããšèããŠãããŸãã
-
äºæ¥ã®æ¹åãæ±ºããã«ãç¹èš±åæã®ã¿ã§ã¯ãããªãããšã¯èªæã§ãããã¯ã€ãã¹ããŒã¹ã¯åºç€ç¹èš±ãåãããšããç®æãæ¹åã®æç¢ºåã¯ã§ããŸããããã€ãºã ããã§ããããã®ç¯å²ã®äžã§å ·äœçã«æéãæå®ããŠæåããã¹ããŒã¹ãã©ãã«ãªããã¯èŠããŸããã人éã®åããèŠãŠç¢ºåºŠé«ãäºæž¬ããæ¹æ³ããããŸãããç¹èš±ã®èª²é¡ãGPTãªã©ã§åæ§æããããã«ããã°ã©ãŠã³ãã®æ å ±ã課é¡ã«ä»äžãããã®èª²é¡ã®äŒžé·ã確èªãããã€ãºãæžããææ³ããããŸããçµ±åçã«å€æãããšè¯ããšããããã®æ§é åãè¡ã£ãŠããŸãã
-
ã§ãããããšãçŸç¶ãŸãšãããšæ¬¡ã®éãïŒæ€èšäžïŒ
æ¢åã³ã¢ãã£ãã£åéã«ãããæå確çã®é«ãæ¹åã®ç€ºå
ãæ¢ååéç¶æåãïŒãŽã貯ãããïŒäºæ¥èª²é¡ã®æšç§»ãšç«ã¡äžããã®ç¹å®
ãæžïŒããèŠçŽ ããããã¯ãŒã¯æç³»ååæ
ãæ€èšäžïŒãååãåæã»äºæž¬
æ°èŠåéåå
¥ã«ãããéèŠç¹èš±ã®è¿
éæœåº
ãæè³ã»ã¹ã¿ãŒãã¢ããåãïŒéèŠãªã€ãããŒã·ã§ã³ã®ç¹å®ïŒç¥è²¡ãªã¹ã¯ã®ç¹å®
ãæžïŒããç¹èš±åŒçšããããã¯ãŒã¯ã®åæ
ãæžã»æ€èšŒäžïŒãåŸãåãåæ
é¢é£äŒæ¥åååæ
ããããã£ã³ã°ããªãæ¹åãïŒåãå
¥ããã¹ãæ¹åïŒç«¶æ¥åè£ã®ç¹å®ïŒãã©ã³ã圱é¿å
ãæžïŒã泚åé åã®å¯èŠå
ä¿ææè¡ã®æŽ»çš
ãä¿ææè¡ã®å¿çšå
ã®æ¢çŽ¢
ãæžïŒãClarivateã®ææ³
ãæ€èšäžïŒã課é¡ãå©çšããææ³ã¯å¯èœ
ãæžïŒãå
·äœçãªããŒã¯ãŒããçšããããŠããªãé¡äŒŒé åã§å¥çšéãæ¢ããå¯èœïŒããŒã¯ãŒãå«ãŸãªãïŒæŠå¿µæ€çŽ¢ã§å
·äœç¹å®ãããŠããªãåéãæœåºïŒã
å
šåéã«ãããä»åŸäŒžé·ããåéã®ç¹å®
ã察象å€ïŒ
â»ãªããçŸåšã¯ãæè¡çºå±ãšè¡°éã®ãµã€ã¯ã«ã®è»žã«ãããŠããã®ã©ãã«æ³šç®æè¡ãäŒæ¥ãäœçœ®ããŠãããé 眮ããé 眮ããšã«ã©ã®ãããªå¯Ÿå¿ãæå¹ãã®ãã³ãã¬ãŒãã¥ãããããŠããããã³ãã¬ãŒããå®æãæ¬¡ç¬¬ã軞ãå¢ããæ©æ¢°åŠç¿ã«èœãšã蟌ãããšã暡玢ããŠããŸããå ·äœçã«ã¯ãã³ã¢ãã£ãã£ããã®å¯ŸåŠæ³ã®äºæž¬ã·ã¹ãã ãäœã£ãŠããŸãã
- æç« ã®é¡äŒŒãšã¯ãäžçŸ©çã«æ±ºãŸããã®ã§ã¯ãããŸãããå€å
çãªãã®ã§ãããèŠç¹ã«ãã£ãŠã©ã®ãããªé¡äŒŒãå¿
èŠãèŠå®ãããŸãã
æ¬ææ³ã§ã¯ãããããã®ç¹èš±çã®æç« ã¯ããã®ç®çãèŠç¹ã«å¿ããä»»æã®è€æ°ã®ææ®µã§ãã¯ãã«åãããŸãïŒäœæããããã¯ãã«ã¯ç®çã«å¿ãåæ§æãããŸãïŒ
ïŒãã¯ãã«ã«å¯Ÿãç®çã«å¿ãæ§ã ãªèä»ããè£æ£ãè¡ããŸãïŒããäžè¬åãããããªãCLIPããåŸããã¯ãã«ã«ããèä»ããããŸãïŒããç¹èš±åé¡ãèæ ®ããããªãCPCåé¡åé¡ãArcFaceãªã©è·é¢åŠç¿ãå¯Ÿç §åŠç¿ã«ããè£æ£ããŸãïŒãã®ä»debiasingãããå Žåã«ã¯ä»»æã®äžç«åã«ããæåœ±ãåçåã«ããè·é¢è£æ£ãè¡ãè£æ£ããŸãïŒããç¹å®ã®ãã¡ã€ã³ãèæ ®ããããªããã®ãã¡ã€ã³ã§åŠç¿ããã¢ãã«ããåŸããã¯ãã«ã«ããèä»ããããŸãïŒããããªç¹èš±é åãæœåºããããªãåŒçšã°ã©ããšæç« ãã¯ãã«ãçšãGraphSAGEããåŸããã¯ãã«ãå©çšããŸãïŒCNN, RNN, transformerãªã©ã¢ã«ãŽãªãºã ã®å©ç¹ã𿬠ç¹ãèæ ®ããŠæ¡çšããŸãïŒå»ççšèªãªãJaMIEãMedNER-Jãæµçšãããªã©ããããŸãïŒãã®ä»ãäœããã®å šäœæé©åãããããšã奜ãŸããã§ãïŒ
è£æ£åèïŒ
ããã·ã³ãã«ã«ã¯æ¬¡ã®éã
def neutralize(word, bias, vec_map):
e = vec_map[word]
e_biascomponent = (np.dot(e,bias)/(np.linalg.norm(bias)**2)) * bias
e_debiased = e - e_biascomponent
return e_debiased
ïŒ
ã»ã
ã»æ°Žæšæ ã岡åŽçŽèг
ãåã蟌ã¿è¡šçŸã®æå³é©å¿ã«ããç¥èããŒã¹èªçŸ©ææ§æ§è§£æ¶
ãhttps://www.anlp.jp/proceedings/annual_meeting/2023/pdf_dir/C3-1.pdf
ã»OpenAI ã® Embeddings API ã¯ã€ã±ãŠãã®ããå®éçã«èª¿ã¹ãŠã¿ã
ãhttps://qiita.com/akeyhero/items/ce371bfed64399027c23
ã»Discovering Universal Geometry in Embeddings with ICA
ãhttps://arxiv.org/abs/2305.13175
ãåçš®èšèªãç»åã¢ãã«ã®åã蟌ã¿ã«å
±éããæå³ã®ç¬ç«æåãå¯èŠåãªã©
ã»Nonparametric Bayesian Deep Visualization
ãhttp://chasen.org/~daiti-m/paper/ishizuka21npdv.pdf
ã»è€æ°ã®ãã¯ãã«ãå¿
èŠãâå
ã®é«æ¬¡å
æ
å ±ããã©ã®æ
å ±ãåãåºããããšããèãæ¹ã¯ãäŸãã°ã次å
å§çž®ã«æ³šç®ããã°ã次ã®ããã«ãåæãããã®ã§ãããã
ã»Dynamic visualization of high-dimensional data
ãhttps://www.nature.com/articles/s43588-022-00380-4
ãhttps://www.biorxiv.org/content/10.1101/2022.05.27.493785v1
ã»Deep TDA: "Why you should use Topological Data Analysis over t-SNE or UMAP?"
https://datarefiner.com/feed/why-tda
ã»Semantics in High-Dimensional Space
https://www.frontiersin.org/articles/10.3389/frai.2021.698809/full
ã»é«æ¬¡å
空éã®æå³è«
ã»é«æ¬¡ç©ºéã®ç©ºçœã«ã€ããŠïŒå¹³åããé¡äŒŒãèŠãã®ã§ã¯ãªãïŒçºæææžçæãšããçç±ã®äžã€
ã»1ãã座æš-èšèªã¢ãã«ãäœæãçºæææžçæããå ŽåãšïŒGPT4ãªã©é«åºŠãªèšèªã¢ãã«ã«åº§æšã¢ããã¿ãã€ããŠçºæææžçæããå Žåãšã§ïŒäœãç°ãªãããšã«ãªããè峿·±ãïŒæ¯éå£ä»¥å€ããæè¡åžžèãæã¡èŸŒã¿è£å®ãã€ã€çæãããçºææç« ã¯ãã倧ããªäŸ¡å€ãæã€ã®ãïŒãããšãæè¡åžžèãã®ããšãªããªã©ããŠããå°ããªäŸ¡å€ãšãªãã®ã
ã»æ¬¡ã¯ãšãŠãè峿·±ãïŒ
ãã»æå³è«ç空éã¯ãã°ããŒãã«ãªæ§é ã§ã¯ãªãããã®ããŒã«ã«ãªæ§é ãšè¿é£ã®èгç¹ããèŠããšè峿·±ããã®ã§ãã
⢠æå³è«ç空éã«å¯Ÿããæäœãããã®ã°ããŒãã«æ§é ã«åœ±é¿ãäžãããããªæå³è«ç空éã®æäœã倿ã¯ãæçã«ãªãå¯èœæ§ã¯éåžžã«äœããèšç®ã®ç¹ã§ç¡é§ã§ããå¯èœæ§ãéåžžã«é«ãã§ããã
ã»Automatic Data Curation for Self-Supervised Learning: A Clustering-Based Approach
https://arxiv.org/abs/2405.15613
ããŒã¿äžæ§åã«k-meansã
ã»NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models
https://arxiv.org/abs/2405.17428
ä»ã²ãšã€æ§èœãäœãã£ãLLMãdecodee embeddingã®æ¹è¯ã
transformerããŒã¹ã®embeddingã¯encodeeãdecoderãçŽ çŽãããªãå°è±¡ã§ãããåæ¹æ¹è¯ã詊ã¿ããç¶ããŠããŸãã
ã»æåŸ
è¬æŒ2: ä»äºãã€ã¿å
ç (æ
¶å¿çŸ©å¡Ÿå€§åŠ)
ã¿ã€ãã«: èšèªç¿åŸã«èšå·æ¥å°ã¯å¿
èŠã
èšèªåŠçåŠäŒ30åšå¹Žèšå¿µã·ã³ããžãŠã
ç¥èãéžæããæ©èœã§ããç¥æµãã¢ããã¯ã·ã§ã³ïŒ
ç¥èéžæã¢ãã«ãšãçµã¿åãããããšã§åµé æ§ã®èª¬æãã§ããããšãèžãŸããã«é¢çœããªã
ã»ãžã§ãã¹ã®è¿°ã¹ãç¥æ§ã1982
https://m.youtube.com/watch?si=lzbF1DfhLzeRzI4V&v=ymbD_a-G1IQ&feature=youtu.be
ã 倿§ãªçµéšãšåµé æ§ïŒæªæ¥ã®ããã«èããã¹ãããš
ç¥æ§ãšèŠç¹
è¬æŒè
ã¯ãç¥æ§ãšã¯ããºãŒã ã¢ãŠããããŠå
šäœåãèŠãèœåã§ããã现éšã«åãããã«å€§å±ãææ¡ããåã ãšè¿°ã¹ãŸããè³¢ã人ã
ã¯ãä»äººãæ°ä»ããªããããªæçœãªè§£æ±ºçãèŠã€ããããšãã§ãããã®ããã«ææçœªæªæãæããããšããããšèª¬æããŠããŸãã
ã€ãããŒã·ã§ã³ãšçµéšã®å€æ§æ§
ã€ãããŒã·ã§ã³ãçã¿åºãã«ã¯ãç°ãªãçµéšãçµã³ã€ããèœåãå¿
èŠã§ããäžè¬çãªéïŒé«æ ¡ãã倧åŠãžé²åŠãããªã©ïŒãæ©ãã ãã§ãªãã詩人ãšããŠæ°å¹Žéãéãããããç°ãªãæåãèæ¯ãæã€äººã
ãšäº€æµããããšã§ãæ°ããèŠç¹ãåŸãããšãã§ããŸãããŠã©ã«ãã»ãã£ãºããŒãäžåºŠLSDã䜿çšããããã¡ã³ã¿ãžã¢ãã®ã¢ã€ãã¢ãåŸããšããéžè©±ã玹ä»ããã倿§ãªçµéšãåµé æ§ãåºæ¿ããããšã匷調ãããŠããŸããã
å人çã¯ãã®ãšãããšæããŸããã
æããã€ãããŒã·ã§ã³ã®çããåºãŠãããšæããŸãã
çŸåšã§ã¯ãæå³ç©ºéãšããæŒç®å¯èœãªç©ºéããããããã®é«å次å
空éãã髿¬¡å
ã®ãŸãŸãèŠã俯ç°çªããäœããŸãã
é©åãªæè¡ã䜿ãã°ã人ã®èšæ¶ã«é Œããããªãã£ã1982ãããåçŸããã客芳çã«ãã€ãããŒã·ã§ã³ãèµ·ããããšãã§ããã¯ãã§ãã
å人çã«ã¯ã俯ç°ããŒã«ãåé¡ããŒã«ãããããããèŠããåããŒã«ãšããŠäœ¿ã£ãŠãããã¡ã¯ãã©ãã«ããªããã®ã§ã¯ãšæããŠããŸãã
ã»è«ççæèã«ã€ããŠã¯åå§äººãçŸä»£äººãéãã¯ãªããšãèšãããŠãããšãããçµæãšããŠã¯å·®ãããããã§ãããªãããšèšãããã°ãå人çã«ã¯ãç¥æ§ã®åé¡ã ãšèªèããŠããŸãã
ããžã«ã«ã·ã³ãã³ã°ã¯è«ççæèãšMECEãããªããŸãããåå§äººã«ã§ããŠããªãã®ã¯MECEã§ãããã
人éã¯éåžžïŒã€ããç©äºãææ¡ã§ããªããšèšãããŠãããšãããæŒããªãããããªãã®MECEã人éã®èªç¶ãªèœåã«åŸãã°ïŒã€çšåºŠããæ€èšããªããªãã§ããããããããçŸå®ã®èŠçŽ ã¯æ°çŸãè¶
ããŸãã人éã®èªç¶ãªèœåã§ã¯MECEãéæã§ããªãã®ã§ããåå§äººã®éçã§ãã
ã©ãããŠããããåå§äººã¯èšé²ææ®µãæãããç¶²çŸ
çãªæ€èšãã§ããŸããã§ããããããã¹ã¿ãŒããšããŠãæåãçºæãèšé²ããèšé²ãä¿ç®¡ããããã«ãªããèšé²ãæŽçãç¶²çŸ
å³ãšããŠäœæããã®æ§é ããç¥èãåŸãŠã髿¬¡å
ç§åŠã®èãæ¹ãæ¡çšããããè€éãªMECEãæ±ããããã«ãªã£ãŠãããçŸä»£äººã®å€æã«è³ã£ãŠããããã§ãã
ç¥æ§ãšã¯ç¶²çŸ
俯ç°èœåã§ãããšãããMECEãšã¯ããã°ãç¥æ§ãè¡šãææšãšãã§ããã§ãããã
èšé²ãããªãå£é äŒéã¬ãã«ã§ã¯çŸä»£äººã§ãåå§äººãšåãçšåºŠã®ç¥æ§ããæãŠããç¶²çŸ
俯ç°ã§ããæè¡ã䜿ãã°çŸä»£äººã®ç¥æ§ãæãŠãããã«ãªãããšèšã£ãŠããã§ãããã
ããã§ç޹ä»ãããããªç¶²çŸ
俯ç°å³ã¯ãåå§äººã®å€æããéããããé«åºŠãªç¥æ§ãæã€ããã®ãç¥æ§åäžããŒã«ãšãã£ãŠãé¢çœãããããŸããã
ã»é«æ¬¡å
ããŒã¿ã»ããã«æœãå¹Ÿäœæ§é ãšæ·±å±€åŠç¿ : ãã®è§£æãšå€§èªç±åºŠååŠç³»ãžã®å¿çš
æ¬æŠ éœäž
https://repository.dl.itc.u-tokyo.ac.jp/records/48134
ã»Zipf çœè²åïŒã¿ã€ããšããŒã¯ã³ã®åºå¥ãããããè¯è³ªãªåã蟌ã¿ç©ºéãšæå€±é¢æ°
暪äºã祥
https://speakerdeck.com/eumesy/zipfian-whitening
å人çã«ã¯ïŒçœè²åçããªãæ§é ãå€å°ã®ã¢ã«ãŽãªãºã ã®å·®ã¯ããããšããŒã¿ã«å¯Ÿãæ£ããæ§é ãšããããšããïŒçœè²åçãè¡ããšç䌌çãªã¯ã©ã¹ã¿ãªã³ã°ãè¡ãããããšã«ããæ§é ããåçŽåããããæ¢åãã¿ã¹ã¯ã«é©åçãšãªãïŒãšããã€ã¡ãŒãžãæã£ãŠããŸãïŒãã®ããç¹å®ã®ç®çãç¡ãå Žåã«ã¯çœè²åçã¯å±éºãªã®ã§ã¯ãšïŒïŒæ§é ã®ããã¿ã®åå ãç ç©¶ããŠã¯ããŸããïŒ
ã¿ã€ãïŒåèªã®çš®é¡ïŒãšããŒã¯ã³ïŒåèªã®æ°ïŒã®éããèæ
®ããåã蟌ã¿ç©ºéã®è£æ£ã¯ïŒã¡ããã©æ¬èšäºã®tfidf etc embeddings cluster reconstructing visã§ãã£ãŠããããšã§ããïŒè峿·±ãã£ãïŒæå³ãšé »åºŠïŒå€ããã°ããããã§ã¯ãªãïŒäœé »åºŠã®äžéšãéèŠïŒïŒã¿ã€ãã®è£æ£ãè¡ã£ãïŒã§ãšãããåã蟌ã¿ç©ºéã¯ãã¯ãè¯ãã®ã ããïŒ
çœè²åçãè¡ããšïœãšèšèŒãããïŒè³æãšèªåã®èãã¯åãæ¹åã§ããããã«èŠããïŒ
word2vecã¯äœé »åºŠèªãèæ
®ããããäžæãããææ°ã®åã蟌ã¿ã¢ãã«ããè¶³ãç®ã®çµæãè¯ãããšãããïŒçµéšçã«ããã®éãã§ããããªã®ã§ããŸã ã«word2vecããŒã¹ã§ã®ãã¯ãã«ã䜿ã£ãŠããŸããïŒ
察ç
§åŠç¿ãäœé »åºŠèªãèæ
®ããïŒ
é »åºŠèæ
®ã®çœè²åã
ãã¡ã€ã³æ¯éå£ã«ç¹åããã»ããæ§èœãäžããã£ãŠã®ãtfidf etc embeddings cluster reconstructing visã§ãã£ãŠããããšã§ããïŒãã£ãŠããŠããæ¹åã¯æ£ããããã§ãïŒ
æã®é »åºŠã«ã€ããŠã¯ïŒæãšã¯åèªã®ã¿ã€ãã®äžåœ¢æ
ã§ããïŒãæå³ãšãã空éãã«ããåèªãšåãã§ãããŠåäžã®æ¬¡å
æ°ã§è¡šçŸãããã¹ããã®ïŒã€ãŸãæãšåèªã¯èŠåããå¿
èŠã¯ãªãïŒãšèªèããŠãããŸããïŒããå°ãèããŠã¿ãå¿
èŠããããããããŸããïŒ
ã»Discovering State Variables Hidden in Experimental Data
https://arxiv.org/abs/2112.10755
ç»åããã®å
éšæ¬¡å
ã®äºæž¬ãšé ãç¶æ
倿°ã®å©çš
æ¯éå£ããšã®èª¬ææ¬¡å
ã®åºåãçµã¿èŸŒãã§ãããšè§£éã®ããŒã¹ãã§ãé¢çœããã
ã»èšèªã¢ãã«ã®å
éšæ©åºïŒè§£æãšè§£é
Benjamin Heinzerling æšªäº ç¥¥ å°æ æé
https://speakerdeck.com/eumesy/analysis_and_interpretation_of_language_models

äžå³ã¯ã€ãŸãããã®ããŒãžã§è©Šè¡ããŠããããšã§ãã

äžå³ã§ã¯ãµããããã¯ãŒã¯ãäœã£ãŠããæ¥ç¶ããŠããŸãã
ãããã¯ãŒã¯çè«ã§è¿°ã¹ããããããã«æ¥ç¶ãããã¯ãŒã¯ã¯æ§èœãäœãããµããããã¯ãŒã¯éã®æ¥ç¶ãšãã圢ã®ãããã¯ãŒã¯ã®æ§èœã¯é«ããããšãããã€ã§ããã
ïŒãªãå人çã«ã¯ãéšçœ²éã®ãµã€ãã解決ãããããšã«ããæ¥çžŸåäžã¯ããµããããã¯ãŒã¯ã圢æãããŠãããã€ã©ããŸã§ãã£ãŠããµããããã¯ãŒã¯ãç Žå£ãããã»ã©ã®ãã«çµåãšã¯ãªããªãå€§äŒæ¥ã§ã¯ããŸãåãæ°ã¯ããŸããããµããããã¯ãŒã¯ïŒå°éåãäžååãªãšãããã«çµåã«ãªããããäžå°äŒæ¥ã§ã¯éã«ãéšçœ²éã®å£æ ¹ãç¡ããã°å°éæ§ã®äœäžãšæ¥çžŸäœäžã«ã€ãªããã®ã ãšæããŠããŸãããŸãã€ãŸããå°èŠæš¡èšèªã¢ãã«ã§ã¯ããã¯ãããªãã®ã ãããªããšãïŒ
ã»The Geometry of Categorical and Hierarchical Concepts in Large Language Models
https://arxiv.org/abs/2406.01506
ãç·åœ¢è¡šçŸä»®èª¬ïŒLinear Representation HypothesisïŒã ã«åºã¥ãã
ãã€ããªæŠå¿µïŒäŸ: ç·æ§â女æ§ïŒã¯ãLLMã®è¡šçŸç©ºéå
ã®çŽç·æ¹åãšããŠè¡šçŸããããéå±€çé¢ä¿ïŒäŸ: åç© â åºä¹³é¡ â ç¬ïŒã¯ã衚çŸç©ºéå
ã§ãçŽäº€ãããé¢ä¿ãšããŠãšã³ã³ãŒãããããã«ããŽãªçæŠå¿µïŒäŸ: {åºä¹³é¡ãé³¥é¡ãç¬è«é¡ãéé¡}ïŒã¯ãå€é¢äœïŒpolytopesïŒãšããŠè¡šçŸããã
ãšããç¥èŠã瀺ãããŠããã
ã·ã³ãã«ã«X軞ãéå±€ã®äž»èŠçŽ ãy軞ãåäžéå±€ã®çš®é¡ã®äž»èŠçŽ ããšãªã£ããšçè§£ããŠããã®ãïŒãtfidf cluster visã§ç¹èš±ãæ¯éå£ãšããéã«ã¯ãXè»žãæ©æ¢°ã»çç©ã®ãã€ããªã®äž»èŠçŽ ãyè»žãæ©æ¢°ã®è€éããçç©ã®è€éããã§å³ç€ºã§ããããšãå€ãã
ããã¯æ¬¡ã®ç€ºåããããšããããã
ã»ç¬¬äžã®äž»èŠçŽ ã«ã€ããŠæ°å€ã墿žããããšã«ãããéå±€ã倿ŽããçµæãåŸãããïŒXåº§æšæäœã§é局倿Žãã§ããïŒ
ã»ç¬¬äºã®äž»èŠçŽ ã«ã€ããŠæ°å€ã墿žããããšã«ããããã®éå±€ã«ãããè€éãã倿ŽããçµæãåŸãããïŒyåº§æšæäœã§è€éã倿Žãã§ããïŒ
ã»x, y ,z軞ãããªãæ§é ãæ¢çŽ¢ããããšã«ãããç¹å®ã®çµæãåŸããããïŒäŸãã°ãç¹å®è»žãåºå®ããéã«èŠãããåéããã«ãŒãªãŒã·ã£ã³ã³ã®æ°èŠäºæ¥ã衚ãããªã©ïŒ
ã»åèªåæ£è¡šçŸã®ãªã幟äœçé
眮ãšåèªå
±èµ·è¡åã® ãã€å
éšæ§é ã®åæ
åç°æåŒã
https://dspace.jaist.ac.jp/dspace/bitstream/10119/18344/5/paper.pdf
(1) 幟äœåŠçé
眮ã®çºèŠ
åèªãã¯ãã«ã¯ãæå³é¢ä¿ã«å¿ããŠç¹å®ã®å¹Ÿäœæ§é ïŒå¹³è¡åèŸºåœ¢ãæ£åé¢äœãäºè§åœ¢ãªã©ïŒã圢æããã
ç¹ã«ã顿šé¢ä¿ïŒKing - Man + Woman â QueenïŒã¯ãå¹³è¡åèŸºåœ¢ã®æ§è³ªãæºãããŠããã
(2) æ°åŠçè§£æ
åèªå
±èµ·è¡åã¯ãã³ãœã«ç©ãçšããŠè¡šçŸå¯èœã§ãããèšèªã®ååžæ§é ãç·åœ¢ä»£æ°çã«èšè¿°ã§ããããšã瀺ããã
å¹³è¡å蟺圢ã圢æããããã®å¿
èŠå忡件ãå°åºããåèªåæ£è¡šçŸãæã€ç·åœ¢æ§ã®èµ·æºãæ°åŠçã«èª¬æã
ãåŸæ¥ã®ãã¯ãã«ç©ºéã¢ãã«ïŒVSMïŒãååžä»®èª¬ïŒdistributional hypothesisïŒãè¶
ãããããæ§é çãªã¢ãããŒãã
å人çã«èå³ãããæ¹åã¯ããã
ããã«ãããããããQA圢åŒã®çæAIã§ã¯ãã©ãçãããããæåããæ°èŠäºæ¥ãªã©ã®ãIPã©ã³ãã¹ã±ãŒãã«é¢ãããŽãŒã«ã«ãã©ãçãããšèããŠãããïŒäŸãã°ãæåããæ°èŠäºæ¥ãšããæå³ã¯ãã·ã³ãã«ã«ã¯ãç¹å®ã®è»žãèšå®ããããã§èŠã€ããåéåœ¢ã®æ§é ãšããŠæœåºå¯èœã ãšä»®å®ãæ€èšŒããŠãããšãããå°ãªããšã詊ããæ°äŸã«ãããŠã¯æ£ãããïŒ
ã»éšå空éã®æ¬äŒŒçŽäº€æ§ã«ããTransformer èšèªã¢ãã«ã®å
éšè¡šçŸã®è§£é
https://www.anlp.jp/proceedings/annual_meeting/2025/pdf_dir/P2-12.pdf
ã»ç¬ç«æååæã«ããäºååŠç¿æžã¿å€èšèªã¢ãã«ã®å±€ã暪æããåèªåã蟌ã¿è¡šçŸã®åæ
https://www.anlp.jp/proceedings/annual_meeting/2025/pdf_dir/P2-15.pdf
ICAã«ããè§£æã¯è峿·±ããšãã
ã»TrendScape 1.0: èšèªã¢ãã«ã®æœåšç©ºéäžã®æŠå¿µæ¢çŽ¢
https://www.anlp.jp/proceedings/annual_meeting/2025/pdf_dir/P2-25.pdf
ïŒåã蟌ã¿è¡šçŸã®ãããã¯ãŒã¯åã¯ä»¥äžã®ããã«tfidf cluster visã§ãè¡ã£ãŠããããããçããæ¹æ³ã§ããªããšæããçžéç¹ã¯ïŒ
ãææ¡ããã¢ãããŒã㯠5 ã€ã®äž»èŠãªã¹ãããããæ§æãããïŒ
ãŸãïŒå
¥åã¯æœåšç©ºéã«ãããåã蟌ã¿è¡šçŸãšæŽåããïŒæ ŒçŽãããïŒ
次ã«ïŒæ¢çŽ¢ç©ºéã®ãµã³ããªã³ã°ãè¡ããïŒäŸãã°æŠå¿µ A ããæŠå¿µ B ãžã®ãã¯ãã«ç©ºéã®é¢ä¿ãæœåºããïŒããããããã¯ãŒã¯ã«æ§é åãããïŒ
ç¶ããŠïŒæ¢çŽ¢ç©ºéãåãµã³ããªã³ã°ããªããïŒãã¹ã®åšèŸºãæ¢çŽ¢ããïŒ
æåŸã«ïŒç©ºéå
ã®é¢ä¿ãåæ§ç¯ããïŒå¯èŠåãããïŒã
ãTrendscape æ³ã¯ïŒåèªåã蟌ã¿ãåºã«äœåéã®æŠå¿µçãªãéçããå¯èŠåããæ¬ç ç©¶ã®æ žå¿çææ³ã§ããïŒãã®é©çšã¯ä»¥äžã® 4 ã€ã®äž»èŠã¹ãããã§æ§æãããïŒã
ã1ãæ¢çŽ¢ç©ºéã®æ§ç¯ïŒäºã€ã®äœåïŒsource ãš targetïŒéã®æŠå¿µçæ©æž¡ããšãªãåèªçŸ€ãç¹å®ããããïŒâword2vec-google-news-300â ã®âcloser_thanâ ã¡ãœãããçšããŠãµã³ããªã³ã°ãè¡ã£ãïŒãã®ã¡ãœããã¯ïŒsource ãš target ã®äž¡æ¹ã«ãè¿ããåèªãå¹ççã«æœåºããïŒããã«ïŒèšç®å¹çãšå€æ§æ§ã®ãã©ã³ã¹ãåãããïŒãµã³ããªã³ã°ããåèªçŸ€ããã©ã³ãã ã«500 èªãéžæãïŒåææ¢çŽ¢ç©ºéãšããïŒã
2ã€ã®äœåã®äž»èŠå
±éïŒåèªãåãåºããŠããã§æ¯èŒãããšãå
±éããªãåèªãæœåºããã®ãïŒ
ã2. ãããã¯ãŒã¯æ§é åïŒæœåºãããåèªçŸ€ã®é¢ä¿æ§ãæ§é åããããïŒscikit-learn ã©ã€ãã©ãªã®AgglomerativeClustering ãçšããŠéå±€çã¯ã©ã¹ã¿ãªã³ã°ãå®è¡ããïŒã¯ã©ã¹ã¿ãªã³ã°ã®çµæã«åºã¥ããŠåèªéã«ãšããžãäœæãïŒãã®éã¿ãã³ãµã€ã³é¡äŒŒåºŠã® 2 ä¹ãšããŠèšå®ããïŒãã®ææ³ã«ããïŒæå³çã«è¿ãåèªå士ã匷ãçµã³ã€ãããããã¯ãŒã¯æ§é ã圢æãããã
åèªããŒã¹ïŒã¯ã©ã¹ã¿ã§é¡äŒŒåºŠãåŸãŠåèªããŒãéãšããžãšéã¿ã«ïŒ
ã3. çµè·¯æ¢çŽ¢ïŒæ§ç¯ããããããã¯ãŒã¯äžã§ïŒsourceãã target ãžã®æççµè·¯ãæ¢çŽ¢ããïŒãã®éçšã§ã¯ïŒnetworkx ã©ã€ãã©ãªã®æççµè·¯ã¢ã«ãŽãªãºã ãåºæ¬ãšãã€ã€ïŒæå³çãªé¢é£æ§ãããéèŠããããïŒç¬èªã®ãã¥ãŒãªã¹ãã£ãã¯é¢æ°ãå®è£
ããïŒå
·äœçã«ã¯ïŒåã¹ãããã§ top-kïŒk=3ïŒã®åè£ãèæ
®ãã確ççãªæ¢çŽ¢ãè¡ãïŒå±æçãªæé©è§£ã«é¥ããªã¹ã¯ã軜æžããïŒã
äœåéã§å
±éããŠããªãåèªããŒããšå
±éããŠããåèªããŒãå«ããçµè·¯æ¢çŽ¢ïŒ
ã4. çµæã®çµ±åïŒæ¢çŽ¢ã®å®å®æ§ãšç¶²çŸ
æ§ãé«ããããïŒäžèšã®ããã»ã¹ãè€æ°åç¹°ãè¿ãïŒåŸãããçµæãçµ±åããïŒæçµçãªãããã¯ãŒã¯ã¯ïŒpyvis ã©ã€ãã©ãªãçšããŠå¯èŠåãïŒæŠå¿µéã®é¢ä¿æ§ãçŽæçã«çè§£ã§ãããã工倫ããïŒã
ããããã®ã¹ããããéããŠïŒäœåéã®æŠå¿µçãªæ¶ãæ©ãšãªãåèªã®ç³»åãç¹å®ãïŒãã®é¢ä¿æ§ãèŠèŠåããããšã«æåããïŒã
ããŒããå
šãããã¯ãŒã¯ãäœã£ãŠãäœåãšã®è¿ãã§ãããã¯ãŒã¯å圢æããææ³ãããã¡ãªã®ããªãäœåéã®é¢ä¿ãæç¢ºã«èгå¯ããã«ã¯ããŒãã®è²åãã§å€æã§ãããã¡ãã®ã»ããèŠããããããã ãåŸä»ãåŠçã¯ããã«ãããã ãªãããŒããäœåãšèŠçŽ ã®é¡äŒŒæ§ãåŸä»ãã§èšç®ããããŒãã®è²ãèšå®ããã»ããããè¯ãããª
ã»ãã³ãã£ãŒäŒæ¥ã«ãããã«ãŒãªãŒã·ã£ã³æŠç¥ã®å®æœã¯å¹æãããã
https://cir.nii.ac.jp/crid/1050006275874652800
確ãã«ãã«ãŒã»ãªãŒã·ã£ã³ãçºèŠãå®è¡ããã°è¶³ããåé¡ã§ã¯ãªãã
ã»ãã«ãŒã»ãªãŒã·ã£ã³æŠç¥ã«ããåžå Žåµé ã«åããç°åžžæ€ç¥æè¡ãçšããç¹èš±æŠç¥å€åã®ç¥èç²åŸ
https://www.google.com/url?sa=t&source=web&rct=j&opi=89978449&url=https://www.jstage.jst.go.jp/article/pjsai/JSAI2023/0/JSAI2023_2L4GS302/_pdf&ved=2ahUKEwiYpIyNp5eMAxWzm68BHaZgBe0QFnoECCYQAQ&usg=AOvVaw0RpvRjPRljI7d8aXElE265
Local Outlier Factor(LOF) ã®äœ¿çšã«ããç°åžžå€æ€åºãšæ°èŠæ§ãæè¡ã·ããã®æ€åºãé¢çœãããå人çã«ã¯ãã®ãããªã¹ã«ã©ãŒã§ãªãããã髿¬¡ã®å¹Ÿäœæ§é ãæã£ãŠæ€åºããå¿
èŠããããšã¯æãã
ã»Training Large Language Models to Reason in a Continuous Latent Space.
https://arxiv.org/abs/2412.06769
ãããããAIã¯ããªãèšèã§èããå¿
èŠãããã®ãïŒã
ãŸãã«ã
å人çãªèŠè§£
ïŒå人çã«ã¯ïŒããŒã¿ããŒã¹ã¯ããŒã¿ãå
¥æãããã®ïŒåæã»è§£æã»æ¢çŽ¢ããŒã«ã¯èªãäœããã®ïŒãšèããŠãããŸãïŒ
èªç¶èšèªåéã§ã¯ç¹ã«ïŒData analystã§ãã£ãŠãïŒData scientistã®ããã«ãããŒã¿ããæªç¥ãã¢ãã«åãçè§£ããæ¹æ³ãèªãçã¿åºãããªããã°ïŒanalystã®ç®çã§ãããããŒã¿ããã€ã³ãµã€ããåŸãŠçããåºããããšãã§ããªãïŒãšèããŠããŸãïŒ
ïŒèªç¶èšèªåéã§ã¯Data analystãšData scientistã¯ã»ãŒåããšèªèããŠããŸãïŒïŒïŒåžžã«æ§é ãäºåã«æ±ºãããŒã¿ããŒã¹ãªã©ã«æŽçããŠããããã®ã§ã¯ãªãïŒãæ¢çŽ¢ãŸãã¯åææã«ãªã¢ã«ã¿ã€ã ã«æ§é åããå¿
èŠãããããšèªèããŠããïŒãã®ããã«ã¯ãªã¢ã«ã¿ã€ã ã«ã¢ãã«ãéžæã䜿ããŸãã¯ã¢ãã«ãäœãå¿
èŠãããïŒãšããã°ããã§ãããã?ãïŒ
çŸæç¹ã«ãããŠïŒãå¿
èŠã«è¶³ããã ãã®èªç¶èšèªåŠçæè¡ãåããæ¢çŽ¢ããŒã«ãã¯ååšããŠãããïŒå€å
žçãªããã¹ããã€ãã³ã°ä»¥äžã®æ¢çŽ¢ãè¡ããããªãèªãããã°ã©ã ãçµã¿äœãã»ããªãïŒãšèªèããŠããŸãïŒ
â»ãïŒçºå£²ã«é¢ãèŠå¶ãå³ããã®ã§ïŒéçºè
èªèº«ã«ãã䜿ããªãéå
·ãäžçã«ã¯å±±ã»ã©ããããšã¯çŽåŸããæèŠã§ããã
ã
䜿çšäŸïŒïŒããŒã¯ãŒãæ€çŽ¢ã§ãå顿€çŽ¢ã§ããªãïŒç¬¬äžã®æ€çŽ¢ææ³ãšããŠ
次ã®ãããªèšäºãèŠã€ãããšããŸãïŒ
ãã€ãã®DXïŒååºå
éšãšç¹èš±ã«èŠãéãã®ç§å¯ïŒç«¶åä»ç€Ÿã®å¯Ÿæçã¯ïŒ
https://xtrend.nikkei.com/atcl/contents/18/00288/00002/
ãã€ãã®ååºã·ã¥ãŒãºïŒã©ã®ãããªç¹èš±ãããã®ã§ããããïŒ
æ€çŽ¢ããŠã¿ãŸãããïŒ
J-PlatPatã§æ€çŽ¢ïŒåºé¡äººïŒ"ãã€ã" AND ããŒã¯ãŒãïŒ"ã·ã¥ãŒãº"ã0ä»¶
ãã€ããã·ã¥ãŒãºã«é¢ããç¹èš±ãåºé¡ããŠããªããšã¯èãã«ããã§ãïŒã·ã¥ãŒãºãšããçšèªã䜿ã£ãŠããªãã®ã§ãããïŒããã³ãäžèŠïŒ
â
tfidf embeddings/cluster visã䜿ã£ãŠã¿ãïŒ
ä»»æã®æ€çŽ¢ããŒã«ã§æ€çŽ¢ïŒåºé¡äººïŒåå¯ãããïŒïŒ"ãã€ã"
ïŒæ€çŽ¢æŒããå€ããããŸããããã§ã¯èæ
®ããŸããïŒãªãïŒåå¯ããããŸãåäœããŠããªãããã§ãïŒ
ããŠã³ããŒã
â
ããŠã³ããŒããããã¡ã€ã«ãæå®ãïŒã³ãŒãå®è¡
â
5åã»ã©ç«ãšéã³ã€ã€åŸ
ã€
â
衚瀺ãããã°ã©ãã確èª
ãã
ãã»å³äžã«æ¿ç³»ã®ã¯ã©ã¹ã¿
ãã»å·Šãšäžã«ãŽã«ãç³»ã®ã¯ã©ã¹ã¿
ãã»å³ã«éä¿¡ç³»ã®ã¯ã©ã¹ã¿
ãã»äžå€®ã«ãæ©èœã»ææã»æ§é ç³»ãã®ã¯ã©ã¹ã¿
ãããããšç¢ºèªãããïŒ
èšäºãããããšãã€ãã®ååºã·ã¥ãŒãºã¯ãæ©èœçã§ãã¬ãŒãå
¥ãæ§é ãã§ããããã§ãïŒ
該åœããç¹èš±ã¯ïŒè©²åœããããªã¯ã©ã¹ã¿ãèæ
®ãããšãäžå€®å°ã
äžå¯ãä»è¿ãã«ååšããŠããã§ããããïŒ
â
å¥ãŠã€ã³ããŠã§éããŠããã€ã³ã¿ã©ã¯ãã£ãHTMLãããããŸãïŒ
ãã
æåŸ
éãã®ç¹èš±ãïŒæåŸ
éãã®äžå€®äžå¯ãä»è¿ã§èŠã€ãããŸããïŒ
â
tab, enterã
âGoogle Patentã§ç®çã®å
¬å ±ãéãïŒåå¥ç¢ºèªãè¡ããŸãïŒ
ãã®ããã«ïŒããŒã¯ãŒããåé¡ã䜿ãããšãæ€çŽ¢ããããšãã§ããŸãïŒ
ïŒãã®ä»ïŒããŒã¯ãŒãã§ã®æ£èŠæ€çŽ¢ãå顿€çŽ¢ïŒç¹åŸŽæ€çŽ¢ãªã©ïŒpandasã§å¯èœãªæ€çŽ¢ã¯å
šãŠã§ããŸãïŒïŒ
ã
ïŒäžèšèšèŒã«ãããããªïŒãããŒã¯ãŒãã§ãåé¡ã§ãé£ããæ€çŽ¢ããããéã«ã¯ãã®ææ³ãåããŠããã§ãããïŒ
[ç¥è²¡å¡Ÿ]ç·Žç¿åé¡ã·ãªãŒãºã®çš®æãããšæ€çŽ¢ã®é£æåºŠ
https://note.com/sakaimisato/n/n512d162618b1
ïŒåºæ¬çã«çžŠè»žæšªè»žã«æå³ã¯ãããŸãããïŒ
ãåŒã³æ°Žãã䜿ãã°ïŒãæ¹åã«æå³ããä»äžã§ããŸã
軞ã«èªç¥ã§ããæå³ãããããããšã«ããïŒä»»æã®æ¹åã®æ€çŽ¢ããããããªããŸãïŒ
ããæ¹ã¯åçŽïŒç®çã®æ¯éå£ã«å¯Ÿãã軞ãšãããåé¡ïŒãªã©ïŒã®ãããŒéåããå ãåŠçãïŒãããŒãé衚瀺ãšããã ãã§ãïŒ
äŸãã°ïŒããäŒæ¥ã®è£œå€ç¹èš±ã®ãã¡ãããã«é¢ãã補å€ç¹èš±ãæœåºãããå Žåã«ã¯ïŒãããã®ãããŒéåãå ãåŠçããŸãïŒãããšïŒè£œå€ç¹èš±ã®ãã¡ãããã«é¢ããç¹èš±ã®ã¿ãä»ããé¢ããŠè¡šç€ºããããããªããŸãïŒãã®æ¹åã¯ããããç€ºãæ¹åã§ããïŒãã®è»žäžã«ã¯ãããã®éãã衚çŸãããŠããŸãïŒ

ãããŒéåã«ã¯ïŒå人çã«ã¯ç¹èš±åé¡ã䜿ãããšãå€ãã§ãïŒMeSHãéºäŒåãªã©ä»ã®åé¡ã§ãè¯ãã§ãããïŒ
å
šãé¢é£ã®ãªããããŒæ¯éå£ãçšããã°ïŒåºãé¢é£ããç¹èš±çŸ€ãäžã€ã«ãŸãšãŸãããããªããŸãïŒãã¥ãŒã¹èšäºã®åšãã«é¢é£ç¹èš±ãéãããå Žåã¯ïŒãã®ææ³ãæ¡çšããŠãè¯ãã§ãããïŒ
ïŒæç€ºçã«1軞ãä»»æã®è»žãšããããšãã§ããŸã
次ã¯ïŒx軞ãæç³»åãšããäŸã§ãïŒ

éå»ããã®åŒçšãªã®ãïŒå¥åéããã®åŒçšãªã®ãïŒãªã©èŠããããªããŸãïŒ
ç¹å®ã®äŒæ¥ã®ç¹èš±ãæ¯éå£ãšããå Žåã«ã¯ïŒæ¬¡ã®ãããªããšããããã§ãããïŒ
被åŒçšæ°ãå°ãªããšãïŒåŒçšç·ãç¹ãã£ãŠããæãå€ãç¹èš±ã¯ïŒãã®äŒæ¥ã«ãšã£ãŠéèŠãªç¹èš±ã§ãããïŒ
ä»åéã«å€ãã®åŒçšç·ã䌞ã°ããŠããç¹èš±ã¯ïŒãã®äŒæ¥ã®ã³ã¢æè¡ã«è¿ããã®ãšãªã£ãŠããã§ãããïŒ

SDIã®æ¯éå£ã«ã€ããŠé©çšããã°ïŒååºé¡ãå€ãâåå²ãå€ãéèŠã§ãããç¹èš±ïŒãæœåºããããšãã§ããŸãïŒ
åªå
æ¥ä»ãã®å
¬éïŒéåžžåºé¡çã®å
¬éïŒæ¥æ¬ç§»è¡ããPCTã®å
¬éïŒç»é²å
¬å ±ã®å
¬éïŒã®ã¯ã£ãããã垯ãèŠããŸããïŒ
æ¯éå£ãç»é²å
¬å ±ã«éå®ããã°ïŒéèŠã§ãããïŒæ©ææš©å©åããç¹èš±ãèŠãããããªããŸãïŒ
åŒçšã®è§£éã«ã€ããŠã¯æ§ã
ãªè³æãããããŸããããæ¬¡ãèªãã§ããã°è¯ãã§ããã
Papers and patents are becoming less disruptive over time
https://www.nature.com/articles/s41586-022-05543-x

â»åº§æšã®ã³ã³ãããŒã«ã«ã¯æ¬¡ã®ãããªããæ¹ããããŸãã
äŸãã°è³ã®äœçœ®ãšããããŒã¯ãŒããçžé¢ãããããšãã
ãã®ããŒã¯ãŒãã®æ¯éå£ã®åº§æšãåŸãŠè³ã®äœçœ®ãšåãããã«ä¿®æ£ãæ°ããªåº§æšãåŸãããããç¹°ãè¿ããè³ã®äœçœ®ãšããŒã¯ãŒããçžé¢ãã座æšãäœããã³ã¬ã«åºã¥ãå³ç€ºããã°ããããã§ãã
åžè²©ããŒã«ã䜿ããªãå Žåãã®ãããªå·¥å€«ã¯ããããã§ããã
ïŒåŠç察象ã¯ç¹èš±ã«éããŸããïŒåäžæ§ã®ããé©åºŠã«é·æã®ããã¹ã矀ã§ããã°äœã§ãè¯ãã§ãïŒ
äŸãã°
æç®ãšç¹èš±ãåæã«è¡šç€ºããäŸïŒ

ããŒãºèª¿æ»èœåã¯å€§äŒæ¥ã»ã©é«ãæå©ãšãªããŸãïŒãããæã¡ç Žãã«ã¯æ¢çŽ¢ã»è§£æïŒãå€åããã»ã¹ã®è§£æãããã®ããŒãºææ¡ãå¿
èŠã§ãããïŒç¹èš±ã»æç®ã»ãã¥ãŒã¹ãæ··åããããšã«ããïŒãã®è§£æã¯ãã容æã«ãªãã§ãããïŒ
ïŒç¹èš±ãšæç®ã§ã¯äœ¿ãããŠããçšèªã®è§£å床ãç°ãªããŸãã®ã§ïŒããŸãçµ±åã§ããªãããšããããŸãïŒé¢é£ã®äœãåŒã³æ°Žãçšãã¯ã©ã¹ã¿ãå§çž®ããïŒCLIPãéãäžè¬åããïŒæç« çæã¢ãã«ãä»ããïŒãããã¯ã¢ãã«ãçšãç²åºŠãèãããïŒæ±çšã¢ãã«ãéãïŒç䌌察ç
§åŠç¿ã«ããè£æ£ããïŒãããã¯ãŒã¯åãäžå¿æ§ã®åé¡ã«è¿éããïŒæ¬ææ³ã®ã¿ã§ããé³åã®å€§ãããå°ããããïŒãªã©è¡ããšè¯ãã§ãããïŒïŒãã¡ã€ã³ç¹åããçµæã¯ã¢ãã®åœ±ãæ ããïŒãã®åœ±ã¯ã¢ãèªäœã§ã¯ãªãïŒè€æ°ã®åœ±ãçšããŠã¢ãã®åœ¢ãåæ§æããïŒç¹èš±ã¯é·ã䌞ã³ã圱ã§ããããšãå€ããã¥ãŒã¹ã¯ãããéå¿ã«è¿ã圱ã§ããããšãå€ããšããã©ãè¿ã¥ãããïŒãã£ãŠããããšã¯ãŸããããªãšããã§ãïŒïŒ
å°èª¬ã®äººæ°åºŠãšå°èª¬éã®é¡äŒŒåºŠã®é¢ä¿ã®äŸïŒ

èªã¿ããå°èª¬ã®å
容ãã€ã¡ãŒãžãïŒãã®ã€ã¡ãŒãžãã©ã®ãããã¯ã®çµã¿åãããããªããèãïŒãããã¯ã俯ç°å³äžã®ã©ã®ã¯ã©ã¹ã¿ãšé©åããã確èªãïŒãã®ã¯ã©ã¹ã¿ã®åšèŸºã«ãã人æ°åºŠã®é«ãå°èª¬ãéžã³ïŒèªãïŒãªã©ãé¢çœãã§ãããïŒããŒã¯ãŒãæ€çŽ¢ã§èªã¿ããå°èª¬ãèŠã€ããããšã¯ïŒäººæ°äœåã®ãããå€ãåéã§ã¯ç¹ã«ïŒããªãé£ããã§ããããã»ã»ã»
æ¥æ¬èªNLPã³ã³ãïŒåå èšé²ïŒ å°èª¬å®¶ã«ãªãã ãã¯ãæ°äºæž¬ ~â䌞ã³ãâã¿ã€ãã«ãšã¯ïŒ~
https://qiita.com/kzuzuo/items/82ef280f9e6962eb00b2
ç®çãèŠç¹ã«å¿ããä»»æã®è€æ°ã®ææ®µã§ã®ãã¯ãã«åã®äŸïŒ
ããããã®ãã¯ãã«åææ³ã玹ä»ããŠãããŸãïŒ
åºæ¬ã¯ïŒå
ã®ãã€ãã®äŸã瀺ãã«åœãã䜿çšãããã¯ãã«ã§ãïŒ
以äžïŒtfidf embeddings cluster vecãšåŒã¶ããšã«ããŸãïŒ
ãããã¯ã¢ãã«ã§ãã¯ãã«ãäœãããšãã§ããŸãïŒ
以äžïŒtopic embeddings vecãšåŒã¶ããšã«ããŸãïŒ
BERT ãã¯ãã«ã¯âŠçœ®ããŠãããŠã
text-embedding-ada-002ã§ãã¯ãã«ãäœãããšãã§ããŸãã
以äžãtext-embedding-ada-002 vecãšããŸãã
ã»ãã€ãã®äŸãtfidf embeddings cluster vecïŒäœ¿çšåèªæ°èª¿æŽïŒã

䜿çšåèªæ°ã倿Žããããã¯ã©ã¹ã¿ãå§çž®ãããŠããïŒ
ïŒå¯èŠåã¢ã«ãŽãªãºã ã®ãã©ã¡ãŒã¿ãŒã¯å€æŽããŠããªãïŒ
ãããããŠã§ã¢ã®éåèœåæç¥ã·ã¹ãã ãç¹èš±ã¯ïŒéä¿¡ç³»ã®ã¯ã©ã¹ã¿ã®è¿ãã«é
眮ãããŠããïŒ
ãã»ã³ãµãŒã·ã¹ãã ãåããå±¥ç©ãç¹èš±ãšãïŒå±¥ç©ã«åããããïŒéåæç¥ã·ã¹ãã ãç¹èš±ã¯ïŒé¢ããŠããïŒ
not衚çŸã®æç¡ã¯åãã¯ã©ã¹ã¿ã«é
眮ãããïŒ
çºæãããããå¥ã«é
眮ãã€ã€ïŒåŠå®ã现ããªè¡šçŸã¯ç¡èŠãåãã¯ã©ã¹ã¿ã«ãŸãšããŸãïŒé »åºŠãèæ
®ããã®ã§ïŒææ«ã®çºæã®äž»äœã泚ç®ããããããªã£ãŠããŸãïŒïŒè€æ°ã®è«æ±é
ãèšèŒãããŠããã°ïŒææ«ã®çºæã®äž»äœã衚ãåèªã¯èªåçã«å€ããªããŸãïŒ
ç¹èš±æç« ãå³ç€ºããã«tfidf embeddings cluster vecã¯ç¬¬äžéžæãšããŠçæ³çã§ãããïŒ
ïŒäžå®ä»¥äžåèªæ°ãå°ãªãæç« ãå³äžã®éé¢ããã¯ã©ã¹ã¿ã«é
眮ããŠããããïŒå
šäœãšããŠã¯æªãã§ããïŒ
ã»ãã€ãã®äŸãtopic embeddings vec

ãããã¯ã¢ãã«ã䜿ããšïŒã¯ã©ã¹ã¿ãå§çž®ãããããïŒ
å§çž®ãããããŠããå
šäœæ§é ãå€±ãæ°å³ã§ã¯ããïŒãã©ã¡ãŒã¿æ¬¡ç¬¬ã ãïŒïŒ
åè¿°éãïŒç¹èš±ãšæç®ãåæã«ç¢ºèªããéãªã©æçšãªå ŽåãããïŒ
åŸè¿°ããïŒã¯ã©ã¹ã¿ã®å¢çã¯èªèãé£ãïŒ
ãããããŠã§ã¢ã®éåèœåæç¥ã·ã¹ãã ãç¹èš±ã¯ïŒå±¥ç©ã»æ©èœæ§é ç³»ã®ã¯ã©ã¹ã¿ã®è¿ãã«é
眮ãããŠããïŒ
ãã»ã³ãµãŒã·ã¹ãã ãåããå±¥ç©ãç¹èš±ãšãïŒå±¥ç©ã«åããããïŒéåæç¥ã·ã¹ãã ãç¹èš±ã¯ïŒåãå Žæã«é
眮ãããïŒããŽã«ãããŒã«ãç¹èš±ãšããŽã«ãããŒã«ã補é ããæ¹æ³ãç¹èš±ã¯åãå Žæã«é
眮ãããïŒ
åãç¹èš±æç« ã«å«ãŸãããçºæçŸ€ãåãã¯ã©ã¹ã¿ã«ãŸãšããŸãïŒæç« å
ã®åèªã¯ç䟡å€ã«èšç®ãããã®ã§ïŒææ«ã®çºæã®äž»äœã泚ç®ãããããšã¯ç¹ã«ãããŸããïŒ
ïŒå
ã®ãã¯ãã«ã¯æç« ã®åæ§æãããã€ã¡ãŒãžã§ããïŒãããã¯ã¢ãã«ã¯æç« ã®ãããã¯ã远å ããŸãïŒ
ããã§ã¯LDAãçšããŸããïŒLDAã¯ã¹ããŒã¹ã¢ãã«ãšãã€ãºæšå®ã®çµã¿åããã®ãããªãã®ã§ãïŒ
ã»ãã€ãã®äŸãtfidf embeddings cluster vec ãš topic embeddings vec ãåæ
åæãããšãã©ã³ã¹ãè¯ãå Žåããã

ãããããŠã§ã¢ã®éåèœåæç¥ã·ã¹ãã ãç¹èš±ã¯ïŒå±¥ç©ã»æ©èœæ§é ç³»ã®ã¯ã©ã¹ã¿ã®è¿ãã«é
眮ãããŠãããïŒãããã¯ã¢ãã«ããããéä¿¡ç³»ã®ã¯ã©ã¹ã¿ã®è¿ãã«é
眮ãããŠããïŒ
â»ãããã¯ã¢ãã«ããã§ç¹åŸŽèŠçŽ ãåãåºããã»ããïŒåæ£è¡šçŸãèæ
®ã§ããŠããã®ããªïŒãšæããªãããªãïŒ
Discovering Universal Geometry in Embeddings with ICA
https://arxiv.org/abs/2305.13175
https://nlp-colloquium-jp.github.io/schedule/2023-12-20_momose-oyama/
ããã£ãŠæå³ã倧ãããšããã倧ããªãã«ã ã®èŠçŽ ãåãåºãã®ãšã©ã®çšåºŠã®å·®ããããã®ã§ããããã»ã»ã»å転ãããŠè§£éå¯èœæ§ãåäžããã察å¿ãããå°ãããŠè¯ãããªãšæãã
ïŒ1次å
å¢ãããšãããªæã

ããïŒãã®ã¯ã©ã¹ã¿ã®é
眮ãé£ããã®ãïŒãã®ç¹ãç¬ç«æ§ã匷ãã®ãïŒãªã©ããããŸãïŒ
2次å
ã·ãã£ãšã«ãããŠãããå°ãæ
å ±ãã»ããããã®ã¯ã©ã¹ã¿ã®éãªãã«æå³ãããã®ããªãã®ãããšæãããã3次å
ããããã§ç¢ºãããã®ãè¯ãã§ãããïŒ
ã»ãã€ãã®äŸãtext-embedding-ada-002 vec
https://openai.com/blog/new-and-improved-embedding-model
ïŒçºæã®äž»äœã«æ³šç®ãããã©ãããªã©ãä»»æã«ã³ã³ãããŒã«åºæ¥ãªãïŒããçšåºŠäžäœæŠå¿µã§ã®é¡äŒŒãäžã€ã®ã¯ã©ã¹ã¿ã«å§çž®ãããïŒåºç¯ãªããã¹ãã«å¯ŸãäºååŠç¿ããŠããããç¹å®ã®æ¯éå£ãæå®ããŠè§£æããéã«é¡äŒŒãããã¯ãã«ãšãªãããããåæ¹åã§ãªãã¢ã«ãŽãªãºã ã§ã¯åããåºãããïŒãªã©ã®çç±ã«ããïŒããŸã䜿ããŸããïŒBERTãã¯ãã«ãåãã§ããïŒãã ãïŒå°æ¥çã«ã¯ããããŸããïŒ
ïŒOpenAIâs Embedding Model With Vector Database
https://betterprogramming.pub/openais-embedding-model-with-vector-database-b69014f04433
â»PLaMoãããŒã¹ã«ããããã¹ãåã蟌ã¿ã¢ãã«ã®éçº
https://tech.preferred.jp/ja/blog/plamo-based-embedding/
ã
䜿çšäŸïŒ.ïŒïŒåèªã¬ãã«ãŸãã¯æç« ã¬ãã«ã®æŠå¿µæ€çŽ¢ã-æŠå¿µã®åŒãç®ã»æŠå¿µã®éã¿ä»ãæ€çŽ¢- -æç« 矀ã«å¯ŸããChatGPT APIãçšãã質çå¿ç-
ãã¯ãã«ãçšããæ€çŽ¢ããå¯èœã§ãïŒããã§ã¯æŠå¿µæ€çŽ¢ãšåŒã¶ããšã«ããŸãïŒ
æŠå¿µã¯åèªãã¯ãã«ã§ãããïŒæç« ãã¯ãã«ã§ãè¯ãã§ãïŒ
æŠå¿µãåŒãããšãã§ããŸãïŒ
æŠå¿µã«éã¿ãä»äžããããšãã§ããŸãïŒ
Brain Computer Interfaceã«é¢ããæ¯éå£ã«å¯ŸãïŒ
['brain','electrod','chip']ãã®ããŒã¯ãŒããæŠå¿µãšããŠäžãïŒ
['eeg']ãã®æŠå¿µãåŒãïŒ
æŠå¿µæ€çŽ¢ããçµæã¯ãããªæãïŒ
sim_vecãæŠå¿µãšã®é¡äŒŒåºŠã§ãïŒ

æ¯éå£äžã®ç¹å®ã®æç« ãæŠå¿µãšããŠäžãïŒ
ãã®æç« æŠå¿µã«å¯ŸãïŒç¹å®ã®ããŒã¯ãŒãæŠå¿µãè¶³ããŸãã¯åŒãïŒ
æŠå¿µæ€çŽ¢ããããšãã§ããŸãïŒ
ããç¹èš±ãããŒã¹ãšããŠïŒããå°ãã ãç¹å®ã®æ¹åãèæ
®ããç¹èš±ãæ¢ãããïŒãšããå Žåã«äœ¿ããšè¯ãã§ãããïŒ
䜿çšäŸïŒã§çºèŠããã·ã¥ãŒãºç¹èš±ã«å¯ŸãïŒè»žã«çŸããŠããªã现ããæ¹åïŒäŸãã°ã·ã¥ãŒãºã®éãã«ã€ããŠå å³ããç¹èš±ã¯ãªãããªïŒãšããå Žåãªã©ïŒ
æŠç¥ã®æ¢çŽ¢æãªã©ãã䜿ããŸããïŒ
ïŒããåŸå矀ã«ä»ãå¹³åãåãïŒãã®ç¹èš±çŸ€ã®æŠå¿µãšãã®ç¹èš±çŸ€ã®æŠå¿µã®å·®ã¯ãæŠç¥æŠå¿µãã瀺ãã§ã¯ãã®ç¹èš±ã«ãã®æŠç¥æŠå¿µãè¶³ãããïŒïŒãŸãã¯ãã®ç¹èš±çŸ€ã®æŠå¿µãšãã®ç¹èš±çŸ€ã®æŠå¿µã®å·®ã¯ãåéæŠå¿µãã瀺ãã§ã¯ãã®ç¹èš±ãããã®åéæŠå¿µãåŒãå¥ã®åéæŠå¿µãè¶³ãããïŒïŒåãæ¹åã詊è¡ããŠããäŒç€Ÿãå人ãèŠã€ãååããïŒãªã©æ€èšãå®è·µããŠã¿ãã®ãé¢çœããšããã§ãïŒïŒ
ïŒæŒç®çã§çæããæŠå¿µããdecoderãçšãæç« çæãããããšããã§ããïŒæ¬²ããç¹èš±ãæŠå¿µã§è¡šçŸã§ããŠãããªãïŒãã®æç« ã¯çæã§ããŸãïŒäŸãã°ïŒã空çœé åïŒæªç¥é åïŒæ°èŠåéã®æŠå¿µãæŒç®çã«ãã圢æãïŒãã®ç©ºçœé åçã«ãããŠäŸãã°ã©ã®ãããªç¹èš±ãåºé¡ãã¹ããäŸç€ºã§ããããã§ããâ䜿çšäŸïŒïŒïŒãžïŒ
æŠå¿µããšã®éã¿ãèæ
®ããŠæ€çŽ¢ããããšãå¯èœã§ãïŒ
1 * 'brain' + 2 * 'electrod' + 1 * 'chip'
ã£ãŠæãã§ããïŒ
ãªãïŒããããAIãè¡ã£ãŠããèšç®ã¯ïŒãããããŒã¿ã«åºã¥ã现ååããŠçµã¿çŽãããã®ãšèããŠãè¯ãã§ãããïŒéã¿ä»ãããã¡ããšèšå®ãïŒæ€çŽ¢ã§ãããªãïŒããããAIããæ€çŽ¢çµæã¯è¯ããã®ãšãªãåŸãã§ããã
ïŒåãããã¯ãã«ç©ºéãæ±ãtext-to-imageã§ãéã¿ä»ãã¯æçšã§ãïŒ
0.5 * 'a girl' + 0.4 * 'winter' + 0.1 * 'summer'ãïŒå
¥åéšåãå°ã
æ¹å€ããstable diffusionã«æŠå¿µãšããŠäžããå Žåã«çæãããç»åã¯ïŒ

女æ§ãäžå¿ãšããŠïŒéªæ¯è²ã®ãªãå€ã®ããã«å
ã¯åŒ·ãçãæ¡ãå²ããŠããïŒå¬æ¥å
ã£ãŠãšãã§ããããïŒéã¿ä»ãã©ããç»åãçæãããããšæããŸãïŒ
åèªã«éã¿ä»ããããããšã«ããïŒæç« ãã®ãã®ãããèŠç¹ãå¶åŸ¡ãããããªãã§ãããïŒåªæã®ããã«ïŒæç« ããã®èŠç¹æœåºããã©ãã¯ããã¯ã¹ãšèªç¥ãããã¢ã«ãŽãªãºã ã«ä»»ããŠããŠã¯ïŒèŠç¹ã®å¶åŸ¡ã¯ãé£ãã§ãïŒäŸãã°ïŒããã¢ã«ãŽãªãºã ã¯ïŒæé ã®åèªãéèŠãããïŒè²ãç¡èŠãããªã©ç¥ãããŠããŸãïŒïŒ
â»FRONTEOã®ãã¯ãã«å ç®ã®èãæ¹ã¯ïŒéåžžã«ã·ã³ãã«ã ãã©é¢çœããã§ããïŒ
å ç®ã«ããè©äŸ¡å€ãã©ãã ãå€åããã確èªãïŒãã®å ç®ããå åã®éèŠæ§ãè©äŸ¡ããã®ãè¯ãã§ãããïŒ
æç« æ¯éå£ã«å¯ŸããChatGPT APIãçšãã質çå¿ç
æ€çŽ¢ã«ããäžäœïœä»¶ã®ç¹èš±ãæœåºãããã®ç¯å²ã«å¯ŸããŠRetrievalQAãçšãChatGPTã«ãã質çå¿çãè¡ãããšãã§ããããã«ãªããŸããã
https://python.langchain.com/en/latest/modules/chains/index_examples/vector_db_qa.html
ãX瀟ã®ç¹èš±ã§ïœã«é¢ããç¹èš±ã100ä»¶æ±ãããããã«å¯ŸãïŒïŒåé¡ã詳现ã§ãªãã£ããå€åé¡ã«ããã£ãŠãããïŒããŒã¯ãŒããè€æ°ã®è¡šçŸã§è¡šããããªã©ã®äºæ
ã«ããïŒåé¡ãããŒã¯ãŒãã§çµã蟌ã¿é£ãäºæããè€æ°ã®èšèªãæ±ãã¥ããèšèªã§èšèŒãããŠããäºé
ïŒãïŒè³ªçå¿çã«ããæ±ããã
ãtfidf visã匷åã«åæ£ä»®èª¬ãå©çšããŠãããåèªãééã£ãŠäœ¿ãããŠããŸãçç¥ããŠããåçŸçé«ãæ€çŽ¢ããããã®ã«å¯ŸããChatGPTã䜿ãããšã§ãããŸã§åŒ·åãªä»®å®ãããããåæ£ä»®èª¬ã¯å©çšããããã®æèã§ãã®åèªã§ããå Žåã«ã¯ãã¯ããã®ããã«è§£éããã粟床é«ãæ€çŽ¢ã§ããæå³çã«çžè£çã«ã§ããã
ãªã©ïŒãã°ãããã§ããïŒ
æè»æ€çŽ¢ãšåŒã³ãŸããããã
次ã®ããã«åŒã°ããããã«ããªã£ãŠããããã§ã
RAG(Retrieval-augmented generation)
èªåãæããæ¬²ããã£ãïŒè§£ååºŠå€æïŒããããèšãæããå«ãã æ€çŽ¢ãã§ããïŒãšããããŸãïŒ


æç« æ¯éå£ã«å¯ŸããChatGPT APIãçšãã質çå¿çã¯ïŒãŸã æ¹è¯äžã§ããïŒäŸãã°ãããªæãã®çãã³ãŒãã§éæã§ããŸãïŒ
import shutil
from langchain.document_loaders import TextLoader
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.vectorstores import Chroma
from langchain.text_splitter import CharacterTextSplitter
import glob
from langchain.prompts.chat import ChatPromptTemplate, SystemMessagePromptTemplate,
umanMessagePromptTemplate
from langchain.llms import OpenAIChat
from langchain.chains import RetrievalQAWithSourcesChain
def df2ChatGPT(dfxxx_w, return_source_documents=0):
df = dfxxx_w.copy()
df = df.reset_index(drop=True)
df_text = 'åºé¡æ¥ïŒ'+df['åºé¡æ¥'].astype(str) +' åºé¡äººïŒ'+ df['åºé¡äºº'] +' å
容ïŒ'+ df['åç§°'] +' ' \
+ df['èŠçŽ'] + df['è«æ±ã®ç¯å²'].str[:100]
filespath = 'tmp_chatgptretrieval_texts'
try:
os.mkdir(filespath)
except:
shutil.rmtree(filespath)
os.mkdir(filespath)
for i in range(len(df_text)):
tmp_df_text = df_text.iloc[i]
with open(filespath + f'/chat{i}.txt', mode='w', encoding='UTF-8') as f:
f.write(tmp_df_text)
files = glob.glob(filespath+'/' + '**/*.txt', recursive=True)
text_all = []
for i, file in tqdm(enumerate(files)):
loader = TextLoader(file, encoding='UTF-8')
documents = loader.load()
text_splitter = CharacterTextSplitter(chunk_size=400, chunk_overlap=0)
texts = text_splitter.split_documents(documents)
text_all += texts
embeddings = OpenAIEmbeddings()
docsearch = Chroma.from_documents(text_all, embeddings)
system_template="""Use the following pieces of context to answer the users question.
If you don't know the answer, just say that you don't know, don't try to make up an answer.
*ç®æ¡æžãã§åçããŠãã ããïŒ
Answer in japanease:
{summaries}
"""
messages = [
SystemMessagePromptTemplate.from_template(system_template),
HumanMessagePromptTemplate.from_template("{question}")
]
prompt = ChatPromptTemplate.from_messages(messages)
chain_type_kwargs = {"prompt": prompt}
qa = RetrievalQAWithSourcesChain.from_chain_type(
OpenAIChat(temperature=0),
chain_type="stuff",
retriever=docsearch.as_retriever(search_kwargs={"k": 9}),
chain_type_kwargs=chain_type_kwargs,
reduce_k_below_max_tokens=True
)
return qa
seikiword = '.*è.*'
qa = df2ChatGPT(dfxxx[ dfxxx['èŠçŽ'].str.contains(seikiword, na = False)
| dfxxx['è«æ±ã®ç¯å²'].str.contains(seikiword, na = False)
| dfxxx['åç§°'].str.contains(seikiword, na = False)
])
print(qa({"question": 'èã«é¢é£ãã課é¡ã瀺ããŠãã ãã'})['answer'])
GPT-4ã«èª²é¡ãæãããïŒãã®èª²é¡ãç¹è𱿝éå£ã«çŸããŠãããChatGPT APIã§ç¢ºèªãïŒçŸããŠããªã課é¡ã§ããã°ïŒæ°ããããšã§ãããã€æŠç¥çãªç¹èš±ãã§ããã課é¡ã ãšã¿ãªã
ãªã©ãå¯èœæ§ãããããã§ãïŒ(以äžã¯ãäžèšèª²é¡ã®äžã§ïŒãããŸã§ã«å
¬éãããç¹èš±èšèŒã®èª²é¡ã«å«ãŸããŠããªããã®ã¯ãããŸããããšèããŠããŸããïŒãããChatGPT APIã«çœ®ãæãïŒååšããç¹èš±ã®äžã«ãããã©ãããèãããã§ãïŒ)
https://eager-duck-e33.notion.site/Issues-with-Footwear-0c3a0a5a0613410b8ca51a49d70de68b
- å±¥ç©ã«é¢ãïŒäžè¬çãªèª²é¡ãïŒïŒãæãã
ããµã€ãºã®äžäžèŽ: å±¥ç©ã®ãµã€ãºãåã ã®è¶³ã«åããªãå ŽåããããŸããæšæºåããããµã€ãºè¡šèšã§ãããã©ã³ããã¢ãã«ã«ãã£ãŠåŸ®åŠãªéããããããšãäžè¬çã§ãã
è¶³ã®å¥åº·åé¡: é©åã§ãªãå±¥ç©ã¯ãè¶³ã®å¥åº·åé¡ïŒäŸ: è¶³ã®çã¿ã骚æãæ°Žã¶ãããã¿ã³ãå€åæ¯è¶Ÿãªã©ïŒãåŒãèµ·ããå¯èœæ§ããããŸãã
èä¹ æ§: å±¥ç©ã®èä¹ æ§ãäœããšãçæéã§å£åãç Žæãçºçããé »ç¹ã«è²·ãæ¿ããå¿ èŠããããŸãã
äŸ¡æ Œ: å±¥ç©ã®äŸ¡æ Œã¯ããã©ã³ããçŽ æã補æ³ã«ãã£ãŠå€§ããç°ãªããäºç®ã«åãããŠéžã¶ããšãé£ããå ŽåããããŸãã
ç°å¢ãžã®åœ±é¿: å±¥ç©ã®è£œé ã廿£ã«ããç°å¢ãžã®è² è·ãæžå¿µãããŠããããµã¹ãããã«ãªéžæãæ±ããããŠããŸãã
ãã¶ã€ã³ãšæ©èœæ§ã®ãã©ã³ã¹: çŸãããšå¿«é©ããå Œãåããå±¥ç©ãèŠã€ããããšã¯ãæã«å°é£ã§ãã
ã·ãŒãºã³ãçšéã®é©åæ§: å±¥ç©ãè³Œå ¥ããéãå£ç¯ãçšéã«å¿ããŠé©åãªã¿ã€ããéžã¶å¿ èŠããããŸããããã«ã¯ã鲿°Žæ§ãéæ°æ§ãä¿æž©æ§ãªã©ã®ç¹æ§ãé¢äžããŸãã
補åã®å質管ç: å±¥ç©ã®è£œé éçšã§ã®å質管çãäžååã ãšãäžè¯åãåžå Žã«åºåãå¯èœæ§ããããŸãã
ã¡ã³ããã³ã¹: å±¥ç©ã«ãã£ãŠã¯ãé©åãªæå ¥ããä¿®çãå¿ èŠã§ããããããã®æéãã³ã¹ããåé¡ã«ãªãããšããããŸãã
åç©ã®çŠç¥: åç©ç±æ¥ã®çŽ æïŒäŸ: ç®é©ïŒã䜿çšããå±¥ç©ã¯ãåç©ã®çŠç¥ããšã·ã«ã«ãªèª²é¡ãèæ ®ããå¿ èŠããããŸããã
- å±¥ç©ã«é¢ãïŒæºããããŠããªã課é¡ïŒæºããããŠããªãããŒãºãïŒïŒãæãã
ãå®å šãªãµã€ãºã«ã¹ã¿ãã€ãº: æ¶è²»è ã®è¶³ã®åœ¢ç¶ããµã€ãºã«å®å šã«åãããã«ã¹ã¿ãã€ãºãªãã·ã§ã³ããå€ãã®ãã©ã³ãã§ã¯ãŸã æäŸãããŠããŸããã
æ±çšæ§ã®é«ãå±¥ç©: 1è¶³ã§æ§ã ãªã·ãŒã³ãç®çã«å¯Ÿå¿ã§ãããæ¬åœã«æ±çšæ§ã®é«ãå±¥ç©ãåžå Žã«ã¯ãŸã ååæäŸãããŠããŸããã
ãµã¹ãããã«ãªçŽ æãšè£œæ³: ç°å¢ã«åªããçŽ æã補æ³ãçšããå±¥ç©ã¯å¢ããŠããŸããããŸã ãã¹ãŠã®ãã©ã³ãã補åã§æ¡çšãããŠããããã§ã¯ãããŸããã
ã»ã³ãµãŒæèŒã«ããå¥åº·ç®¡ç: å±¥ç©ã«çµã¿èŸŒãŸããã»ã³ãµãŒæè¡ã«ãã£ãŠãè¶³ã®å¥åº·ç¶æ ãæ©è¡ã®è³ªãã¢ãã¿ãªã³ã°ãã補åãäžè¬çã§ãªãã
èªåèª¿æŽæ©èœ: ççšè ã®è¶³ã«åãããŠèªåçã«ãã£ããæãç· ãä»ãã調æŽããæ©èœããã»ãšãã©ã®å±¥ç©ã«ã¯ãŸã æèŒãããŠããŸããã
軜éåãšèä¹ æ§ã®äž¡ç«: 軜éã§ãããªããèä¹ æ§ã«åªããå±¥ç©ã¯ãåžå Žã§ãŸã ååã«æäŸãããŠããªãå ŽåããããŸãã
äžè²«ãããšã·ã«ã«åºæº: åç©çŠç¥ãåŽåè ã®æš©å©ã«é æ ®ãããšã·ã«ã«ãªåºæºã培åºçã«å®è·µãããã©ã³ãã補åãããŸã ååã«æ®åããŠããŸããã
æ®åäŸ¡æ Œåž¯ã®é«å質履ç©: äœäŸ¡æ Œåž¯ã§ãé«å質ãªå±¥ç©ãæäŸãããã©ã³ãã補åãããŸã åžå Žã«ã¯ååã«ååšããªãå ŽåããããŸãã
æèã»é²èæ©èœã®åäž: å±¥ç©ã«ãããæèã»é²èæ©èœã¯ãäŸç¶ãšããŠæ¹åã®äœå°ããããšèããããŸãã
ç¡æã»è¿ éãªè©Šçã»äº€æãµãŒãã¹: ãªã³ã©ã€ã³ã·ã§ããã³ã°ãäžè¬çã«ãªãäžãç¡æã§è¿ éãªè©Šçã»äº€æãµãŒãã¹ãæäŸããäŒæ¥ã¯ãŸã éãããŠããŸããã
- äžèšèª²é¡ã®äžã§ïŒãããŸã§ã«å ¬éãããç¹èš±èšèŒã®èª²é¡ã«å«ãŸããŠããªããã®ã¯ãããŸããïŒã€ãŸãïŒæ°ããã³ããšãªããããã®ã¯ãããŸããïŒ
ãç§ã®ç¥èã¯2021幎9æãŸã§ã®ãã®ã§ãããææ°ã®ç¹èš±ã«ã€ããŠã¯ææ¡ããŠããŸããããã ãããããŸã§ã®ç¹èš±ã«ãããŠã¯ãããã€ãã®èª²é¡ã«åãçµãã§ããäŸããããŸããäŸãã°ãå±¥ç©ã®ãµã€ãºã«ã¹ã¿ãã€ãºãèªåèª¿æŽæ©èœã«é¢ããç¹èš±ãååšããŸãããããã以äžã®ãããªèª²é¡ã«ã€ããŠã¯ãç¹èš±ããŸã ååã«åãçµãã§ããªãå¯èœæ§ããããŸãã
æ±çšæ§ã®é«ãå±¥ç©: 1è¶³ã§æ§ã ãªã·ãŒã³ãç®çã«å¯Ÿå¿ã§ãããæ¬åœã«æ±çšæ§ã®é«ãå±¥ç©ã®ç¹èš±ã¯ãŸã å°ãªããããããŸããã
ã»ã³ãµãŒæèŒã«ããå¥åº·ç®¡ç: å±¥ç©ã«çµã¿èŸŒãŸããã»ã³ãµãŒæè¡ã«ãã£ãŠãè¶³ã®å¥åº·ç¶æ ãæ©è¡ã®è³ªãã¢ãã¿ãªã³ã°ãã補åã«é¢ããç¹èš±ã¯ããŸã ååã«éçºãããŠããªãå¯èœæ§ããããŸãã
ãµã¹ãããã«ãªçŽ æãšè£œæ³: ç°å¢ã«åªããçŽ æã補æ³ãçšããå±¥ç©ã«é¢ããç¹èš±ã¯å¢ããŠããŸããããŸã ãã¹ãŠã®ãã©ã³ãã補åã§æ¡çšãããŠããããã§ã¯ãªããæ°ããçŽ æã補æ³ã«é¢ããç¹èš±ã¯ãŸã éçºã®äœå°ãããã§ãããã
äžè²«ãããšã·ã«ã«åºæº: åç©çŠç¥ãåŽåè ã®æš©å©ã«é æ ®ãããšã·ã«ã«ãªåºæºã培åºçã«å®è·µãããã©ã³ãã補åã«é¢ããç¹èš±ã¯ããŸã ååã«æ®åããŠããªãå¯èœæ§ããããŸãã
æèã»é²èæ©èœã®åäž: å±¥ç©ã«ãããæèã»é²èæ©èœã«é¢ããæ°ããæè¡ãçŽ æã®ç¹èš±ã¯ããŸã éçºã®äœå°ãããã§ãããã
ç¹èš±ã®ååŸã«é¢ããŠã¯ãç¶æ³ãæ¥ã å€åããŠãããããææ°ã®æ å ±ã調æ»ããããšãéèŠã§ããç¹èš±ããŒã¿ããŒã¹ãå ¬çæ©é¢ã®ãŠã§ããµã€ããå©çšããŠãææ°ã®ç¹èš±æ å ±ãå ¥æããããšããå§ãããŸããã
ã»ChatGPTã®ç¥èäžè¶³ãè£ãããã«ãgoogleæ€çŽ¢ãwikipediaãªã©ãçµã¿åãããŠãè¯ãã§ããïŒ
ä¹±æŽãªäŸã§ããæ€çŽ¢äŸã¯ä»¥äžã®éãïŒ

ïŒGoogle Driveã«ä¿åããŠãã倧éã®æç®PDFãã¡ã€ã«ã«ã€ããŠïŒãã®ç¥èã®ç¯å²å
ãïŒChatGPTã§è³ªåå¿çã§ããããã«ãã§ããŸãïŒDuckDBãæ°žç¶åããŸããã®ã§ïŒä»åŸã¯è³ªåããã£ããå
¥åããã ãã§ãïŒïŒ
䟿å©ã§ãããæ¬åœã«ïŒGoogle Driveã®æ€çŽ¢æ©èœã¯ããŸãã¡ã§ãããããã»ã»ã»ã
è²»çšã¯embeddingãã©ã®ã¢ãã«ã§è¡ããã«äŸåããã§ãããïŒOpenAIEmbeddings()ã€ãŸãtext-embedding-ada-002ãçšããŠ7500åã»ã©ã§ããïŒåŸæ¥äŸ¡æ Œæ¹å®ããããŸãããåé¡ã»ã©ã§æžãã¯ãã§ããïŒïŒ
å®äžãããšããã«ã¯ãå
¥åããŒã¯ã³æ°ã倧ãããã€ãã¡ã€ã³ãç°ãªããããªãã¢ãã«ãå¿
èŠã§ããâŠ
ãã䜿ããªãããªãšæã£ãŠããæ¬¡ããã¯ãã«æœåºã®ããã«çšããŠãè¯ããããããŸããã
äžé衚çŸã®åãåºãæ¹ã«ã€ããŠã¯ïŒã€ããªã©åèã«ãªãã§ããã
Explore T5
https://www.kaggle.com/code/debarshichanda/explore-t5
ChatGPTã«ãããåã®ãã¯ãã«ã«ããæœåºãããã«ããã¯ã«ãªãããã§ããChatGPTã§èšãæããèæ ®ããæ€çŽ¢ããããã®ã§ãã®åã®æœåºã§èšãæããé€ãããŠã¯ãããããããŸãããã¬ãŒãã³ã·ã¥ã¿ã€ã³è·é¢ã¯ããŸãã¡ã§ããããâŠã©ããããã®ã§ããããã
â»Llama 2 + multilingual-e5-large ã§ LangChain ã®RetrievalQAã詊ã
https://note.com/npaka/n/n3164e8b24539
embeddingãã©ã®ã¢ãã«ã§è¡ããã«ã€ããŠãä»ãªãe5ãè¯ãã§ããããã
å
¥åããŒã¯ã³æ°ã¯514ã§ãã
â»embeddingãrinnaãe5ãçšãè¡ããPDFã®RQAãããäŸã
alfredplpl/RetrievalQA.py
https://gist.github.com/alfredplpl/57a6338bce8a00de9c9d95bbf1a6d06d
â»Kaggle - LLM Science Exam
Use LLMs to answer difficult science questions
https://www.kaggle.com/competitions/kaggle-llm-science-exam/leaderboard
å€ååããŠã®RAGã³ã³ã
LLMã®fine tuningãããembedingã®ã¢ã³ãµã³ãã«ã®ã»ããè¯ãã£ãæ§åã
èªåããã¯ãã«åæãã¢ã³ãµã³ãã«ãå€çšããŸãããããªå°è±¡ã¯ãããŸããã第äžéžæã¯ãã¯ããã¡ãããªã
â»JapaneseEmbeddingEval
https://github.com/oshizo/JapaneseEmbeddingEval
åã蟌ã¿è¡šçŸãšç¹å®ã®è©äŸ¡ææ³ã§ã®è©äŸ¡
ã»ChatGPT plugin ãé¢çœãã§ããïŒ
次ãªã©è峿·±ã䜿ã£ãŠããŸã
webpilot
webæ€çŽ¢

askyourPDF
PDFèªã¿åã

scholorAI
æç®æ€çŽ¢ïŒ äœæ
ããªã³ã¯ãæç®ã¿ã€ãã«ãæ£ãããªãããšãã»ã»ã»

kraftful
æ¥çæšæºã®ãã¹ããã©ã¯ãã£ã¹ãæ¢çŽ¢ã補åããã¥ã¡ã³ããŒã·ã§ã³ãäœæïŒãšã®ããšïŒ

yabbie
ããŒã±ãã調æ»

zapier
æ§ã
ãªããŒã«ãšã®ãªã³ã¯
Gmail:

notion:

Googledrive:
ãªã©
speechki
é³å£°ã«å€æ

noteable
ããã°ã©ã ãå®è¡
äŸãã°ïŒwebäžã®ãèäœæš©æ³ãå
šæãnoteableäžã§å
¥æãïŒåå²ãïŒpandas DataFrameã«å
¥åããŠãããŠããïŒ

ãã¯ãã«ã«å€æããŠãããšRetrivslQAåçã®åŠçãã§ãããªã©ïŒ

ïŒãã¯ãã«ãçšããChatGPTãçšããŠpandas DataFrameå
éšã®æååãªã©ã«çŽæ¥ã¢ã¯ã»ã¹ããããšãã§ããã°ãã倧éã®æç« ãäžæ°ã«ããæè»ã«æ€çŽ¢ãåŠçãã§ããããã«ãªããïŒä»ã®ãšããã§ããŠããªãïŒïŒ
ç»åã«ã€ããŠã¯ïŒpytorchãimagenetãã€ãã£ããïŒ

ChatGPTãäœã£ãã·ããªãªïŒïŒïŒã«é©åãã確çãïŒCLIPã§æ±ãããïŒ



https://qiita.com/kzuzuo/items/e35e0c0535f0d32b135e
ç»åããäºæž¬ããç©äœã«å¯ŸããŠã·ãã¥ãšãŒã·ã§ã³ãäºæž¬ããæç« ãçæã確çãåºã確çã®åŸã£ãŠããã«ã·ãã¥ãšãŒã·ã§ã³ãäºæž¬ããæç« ãçæãâŠãç¹°ãè¿ãã°ã
ç»åã説æããé·æããç°¡åãªç»åèªèã¢ãã«ãšèšèªã¢ãã«ããã§ããããšã
ã»ChatGPT makes materials research much more efficient
https://engineering.wisc.edu/news/chatgpt-makes-materials-research-much-more-efficient/
æããšã«æ§é åã«å¿
èŠãªããŒã¿ãGPTã§å€æãã衚ã«ãŸãšãããMIãªã©ã§ç¹èš±ããåŠç¿ããŒã¿ãéããå Žåãªã©é€éå®ãããã§ãã
ã»ChatGPTãšNoteableã«ããç§åŠæè¡æ
å ±åæ
https://speakerdeck.com/hayataka88/chatgpttonoteableniyoruke-xue-ji-shu-qing-bao-fen-xi
ã»When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories
https://arxiv.org/abs/2212.10511
ãèšèªã¢ãã«ã人æ°ã®ããäºç©ã«ã€ããŠã®æ
å ±ã¯ããèŠããŠããããã©ãããŸã人æ°ã®ãªãäºç©ã«ã€ããŠã®æ
å ±ã¯èŠåŽããŠèŠããŠããããšãçºèŠããŸãããã
ãèšèªã¢ãã«ãæ
å ±ãèŠããèœåãæ¹åããããã«ãå€éšã®æ
å ±æºïŒéãã©ã¡ããªãã¯ãªã¡ã¢ãªãšåŒã°ããŸãïŒããæ
å ±ãååŸããæ¹æ³ã詊ããŸããããããããã®æ¹æ³ãå®å
šãªè§£æ±ºçã§ã¯ãªããç¹ã«äººæ°ã®ããäºç©ã«ã€ããŠã®è³ªåã§ã¯ãååŸããæ
å ±ã誀解ãæãå¯èœæ§ãããããšãããããŸãããã
ã質åã®å
容ã«å¿ããŠãèšèªã¢ãã«èªèº«ã®èšæ¶ã䜿ãããå€éšã®æ
å ±æºããæ
å ±ãååŸããããé©å¿çã«åãæ¿ããæ¹æ³ãããã«ãããèšèªã¢ãã«ã®ããã©ãŒãã³ã¹ãåäžããããã«ã³ã³ãã¥ãŒã¿ãå¿
èŠãšããåŠçèœåïŒæšè«ã³ã¹ãïŒãåæžãããŸãããã
â»èšèªã¢ãã«ã«å¯ŸãæãæåŸ
ããŠããã®ã¯ãæ§é åãããŠããªãããŒã¿ã®ãæ¢åã®ã¢ãã«ãå
¥åãšããŠäœ¿ããçšåºŠã®æ§é åãwebäžã®èšäºãè«æãPCãgoogle drive å
ãã¡ã€ã«ã®æ§é åãªã©ïŒ
æç®ããç¹å®ã®ããããã£ã®èŠçŽ ãæœåºãæ§é åããããšã¯ã§ããŠããïŒããšã¯ïŒå®å®ããŠãããã§ããããã«ããããšïŒããå°ã詊ããŠã¿ãïŒ
ïŒãæã¯çããã¯ä¿èª¬ãïŒãŒãçæä¿¡ä»°ããå± ãïŒçæã®ã¬ããªãã¯ãšé·æã®ããžãã¯ãåãæ»ãããã«
https://readingmonkey.blog.fc2.com/blog-entry-609.html
é·æã®ããžãã¯ã¯ïŒçµ±èªãŸã§ããããšãç°¡åãªæ°åŒãšåèªãçšããããšã«ããïŒããçšåºŠæç¢ºã«è¡šçŸã§ããã§ãããïŒ
ïŒãã¥ãŒã©ã«æ°åŒãœã«ããŒã«ãããéäžçµæã®è¿œè·¡ãšæäœ
https://github.com/cl-tohoku/trace-manipulate
ãã¯ãã«æŒç®ã®æå³
䜿çšäŸïŒ.ïŒïŒæŠå¿µããã®æç« ã®èªåçæã-åºæ¬ç¹èš±ã®èªåçæ-
䜿çšäŸïŒïŒïŒã®æŒç®çã§çæããæŠå¿µããæç« çæãã§ããŸãïŒ
ããŒã¿ããã¬ããžã®èšèªåïŒData-to-TextïŒvec2seqã§ããïŒç©ºéäžã®åº§æšããæç« ãçæãã翻蚳ã¢ãã«ã®ãããªãã®ã§ãïŒ
ForesightïŒæªæ¥æŽå¯ïŒæè¡ã®äžã€ãšèšã£ãŠãè¯ãã®ã§ãããïŒ
äŸãã°ïŒã空çœé åïŒæªç¥é åïŒæ°èŠåéã®æŠå¿µãæŒç®çã«ãã圢æãïŒãã®ç©ºçœé åçã«ãããŠã©ã®ãããªç¹èš±ãåºé¡ãã¹ããäŸç€ºã§ãããããã§ãïŒå€§ããªç©ºçœãçãããŸã衚çŸåã調æŽã§ããã°ïŒãã«ãŒã»ãªãŒã·ã£ã³ãåãéãããããããªãåºæ¬ç¹èš±ã®èªåçæãã§ãããããšã«ãªããŸãïŒ
ãªãæç« çæãããã®ã§ããããïŒãåºæ¬çãªèãæ¹ã¯æ¬¡ã®éãã§ãïŒ
ã»ãã¯ãã«ç©ºéã«ããã宿
ã¯èªåœãããªããšããïŒèªåœããªã空éã«ãäŸãã°['brain','electrod','chip']ãšããè€æ°èªåœã§è¡šçŸãããç¹ãããããã§ãïŒãã®ç¹ã¯vec2seqã«ããæç« ã§è¡šãããšãã§ããŸãïŒ
ïŒèªåã¯åæã«ããã®ãããªãã¯ãã«ç©ºéã¯ãèªåœã®ãŽãŒã¹ããã§ã§ããŠããããªã©ïœè¡šçŸããŸããããã§è¿°ã¹ãŠããã®ã¯ããŽãŒã¹ãã®åœ±ã«ãæå³ããããã£ãŠæãã§ããïŒ
ã»['brain','electrod','chip']ã§æŠå¿µã圢æããå Žåã¯ïŒ'brain','electrod','chip'ããå«ãæç« ãçæãããããã§ãããïŒæç« çæã®æå³ã¯ããŸããªããããããŸãããïŒãã ïŒäœ¿çšäŸïŒïŒïŒã§ç€ºãããããªçŸ€ããæ±ãããæŠç¥æŠå¿µãã®å Žåã¯ã©ãã§ãããïŒã空çœãã®å Žåã¯ã©ãã§ãããïŒåèªããæŠå¿µãæ±ããŠããªãã®ã§ã©ã®åèªããçæããããããããŸãããïŒåšèŸºããèªåœãéããŠãããããããªããããããŸããïŒãããªãšãã«æç« çæã¯åœ¹ã«ç«ã€ã§ãããïŒ
ã»æç« çæã«ããïŒãæŠå¿µã瀺ãäž»é¡ãã ãã§ãªãïŒãæ¯éå£å
šäœã瀺ãæè¡åžžèã芳ç¹ãã衚çŸããããïŒãšæåŸ
ã§ããŸãïŒ
ãïŒäžçã®ãã¹ãŠã®æç« ããã空çœããæ±ããã®ã§ã¯ãªãïŒæ¯éå£ããã空çœããæ±ããã®ã§ïŒãæ¯éå£å
šäœããäºåååžãšããŸãïŒããããäºååŠç¿ã¢ãã«ã¯äžçãã¹ãŠã®æç« ã®ç©ºçœãåŠãã§ãããšèãããïŒãã®ç©ºçœã¯äžçãã¹ãŠã®æç« ã®èгç¹ããåŸããããã®ãšãªããŸãïŒéå°ãªèгç¹ãã³ã³ãããŒã«ã§ããªã芳ç¹ãåé€ãããïŒåœæ¥è
ãç¡èŠããäžè¬è«ãšãã芳ç¹ãåé€ãããïŒæ¬ç®çã«ã¯æ²¿ããªãã§ãããïŒïŒã¢ããã¿ãŒç³ãçšãïŒæ¯éå£å
šäœã®ååžã«è¿ã¥ãããªã©ïŒæ¬ç®çã§äœ¿ããããã«ããææ®µããªãããªãã§ããïŒ
ãïŒå¯èœãªéãçŽ çŽã«ããã¯ãã«ç©ºéå
šäœãèžçããŸãã
ãïŒååžå
æ±åãšååžå€æ±åããããšããïŒæ¯éå£ãä»»æã«èšå®ããããšã«ããååžå
æ±åã®ã¿ãèµ·ãååžå€æ±åãèµ·ããªãïŒã¡ãªããã享åããŸãïŒ
ãïŒChatGPTãªã©å€§èŠæš¡èšèªã¢ãã«ã«ããçæãããç¹èš±æç« ã¯ååžå€æ±åãšãªããŸãïŒååžå
æ±åãããšãããã®ã¢ãã«ãšã¯ç®çãç°ãªããŸãïŒåŸè¿°ããŸããïŒååžå
æ±åããåŸïŒååžå€æ±åã§è£è¶³ããããšã¯ããã§ãïŒ
ãïŒèŠèŠãšäŒŒããããªãã®ã§ãïŒãšå®æã«é¡äŒŒãè¿°ã¹ããšæãããŸããã©ïŒç¶²èãç³ã¿èŸŒã¿ç©ºéç¹åŸŽéåimg2vecïŒæŠå¿µã®å
¥åïŒ âèŠç¥çµ âå€åŽèç¶äœ(6, P, M)ã»ç¹åŸŽå å·¥ã»å€æ(泚æä»äžïŒã»å·Šå³æ··åçæïŒ)vec2vec âèŠèŠé(ã³ã©ã åäœ)ã«ãããèŠèŠåœ¢ævec2seqãâ髿¬¡é åããã®å
¥åïŒæ¯éå£ã®ä»»æèšå®ïŒïŒ
ãæŽæ°äžïŒãç¶²èã»èŠç¥çµé害ã«ãã倱æããå埩ããæè¡ã«é¢ãããå人çåŠæ³ã¡ã¢
ãhttps://qiita.com/drafts
ã»çæã®ããšãšãªãååžèªäœã¯æ£ãããšç¢ºä¿¡ãšãããæ£ãããšæ³å®ã§ããã®ã§ïŒåååµé çãšèªèããããšãã§ãïŒåãæã®ã€ã³ãµã€ããåºæ¿ãããŸãïŒ
ã»åŸè¿°ãããããã¯ãŒã¯åã«ãããŠã¯ã空çœãå°æ°ã¯èŠããããã§ããç¹çŸ€ã§ã¯ç©ºçœã®äžèº«ãããããããã§ãïŒããã解決ããããã«ã空çœé åã®ç¹èš±çæã«æå³ããããŸãã
ïŒè£ãåŠçãåãããŒã¿ããã§ããã£ãŠã®ã¯ã¡ãã£ãšé¢çœãã§ãããããããèªç¶èšèªåŠçã§ã¯åæãšãµã€ãšã³ã¹ãåããšè¿°ã¹ãçç±ã®äžã€ã§ãïŒ
ã»ãæªæ¥ãèªãããšããããããIPã©ã³ãã¹ã±ãŒãã®äŸ¡å€ã§ãããšããŠãä»åŸæš©å©ãåãããåºç€ç¹èš±ãæç€ºããããšã¯ããã®ç®çã«çŽæ¥æ²¿ããŸãã
ïŒåºé¡ç¹èš±ãæç€ºããææ³ã¯ãããããããŸããïŒããããã¯åã«åãç¹èš±æç®ããããã©ããã®æ€çŽ¢ã®ããã§ãïŒæ¬²ããç¹èš±ã¯ïŒæš©å©ã¯åãããæå³ã®ãªããã®ãšããªãããåãªãåãçºæã®ãªããã®ïŒã€ãŸãåãªãç¹ïŒãããïŒä»ç€Ÿåªäœãšãªããç¯å²ãã®ãããåºæ¬ç¹èš±ããšãªãåŸããã®ïŒã§ãïŒæç« çæææ³ã¯ç©ºçœãçšããããšã«ããïŒç¯å²ã®ããåºæ¬ç¹èš±ã®çæãå¿åã§ããã§ãããïŒïŒç©ºçœã¯å¯åºŠã®åé¡ã§ããã®ã§ä»¶æ°ã«äŸåãé£ãïŒç©ºçœã¯ç¹èš±æç®ã®ã¿ããæ€åºãããšããŠãæè¡å
šäœã®ç©ºçœã§ããããšä»®å®ã§ããïŒãšããç¹ãè¯ããšããã§ãïŒ
ã»è¿œèšïŒãç¹èš±çæã«ã¯èšèªã¢ãã«ãçšããŸããïŒTransformerã®ã¿ãçšãæç« çæããå Žåã®åŒ±ç¹ãè£ãããã«ïŒæŠå¿µãã¯ãã«ããã¡ã€ã³ç¥èãæã€äºåååžãšããŠçšããïŒãšçè§£ãšããŠãè¯ãããã§ãïŒ
- 確çã¢ããªã³ã°ã®ã¢ãããŒãã®éã:
ãã€ãºçã¢ãããŒãã§ã¯ãäºåååžãšå°€åºŠãçµã¿åãããŠäºåŸååžãèšç®ããããã«åºã¥ããŠæ¬¡ã®åèªãéžæãããŸãããã®ã¢ãããŒãã¯ãäºåç¥èïŒäºåååžïŒãšèŠ³æž¬ããŒã¿ïŒå°€åºŠïŒãæç€ºçã«çµã¿åãããŠã¢ãã«ãæŽæ°ããŸãã
äžæ¹ãTransformerã¢ãã«ã§ã¯ã倧éã®åŠç¿ããŒã¿ã䜿ã£ãŠç¢ºçååžãåŠç¿ãããã®ååžã«åŸã£ãŠæ¬¡ã®åèªãéžæããŸãããã®ã¢ãããŒãã¯ãäºåç¥èãšèŠ³æž¬ããŒã¿ãæé»çã«åŠç¿ããŒã¿ã«çµã¿èŸŒãŸããŠãããã¢ãã«ã¯èªåçã«æèãæ§é ãåŠç¿ããŸãããã¡ã€ã³ç¥èãéèŠã§ãããäºåååžãå©çšãããå Žåã¯ããã€ãºçã¢ãããŒããé©åã§ãã
å€§èŠæš¡ãªããŒã¿ã»ããããå¹ççã«åŠç¿ããè€éãªæèãæãããå Žåã¯ãTransformerã¢ãã«ãé©åã§ãã
ãã€ãºçã¢ãããŒãã¯ããã¡ã€ã³ç¥èã«åºã¥ããŠæ°ããã¢ã€ãã¢ãææ¡ã§ããŸãããå€§èŠæš¡ãªããŒã¿ã»ããã«å¯ŸããŠã¯å¹ççã§ã¯ãªãå ŽåããããŸããäžæ¹ãTransformerã¢ãã«ã¯ãåŠç¿ããŒã¿ã«åºã¥ããããçã«æ°èŠãªæç« ãçæããããšãé£ããå ŽåããããŸãã
ãã®ãããæ°èŠæ§ã®ããæç« ãçæããéã«ã¯ããã€ãºçã¢ãããŒããšTransformerã¢ãã«ãçµã¿åãããããšã§ãããããã®ã¢ãããŒãã®å©ç¹ã掻çšããããšãæãŸããã§ããäŸãã°ããã€ãºçã¢ãããŒãã§äºåååžãèšå®ããTransformerã¢ãã«ã§å¹æçã«åŠç¿ã»çæãããšãã£ãæ¹æ³ãèããããŸãã
GPT-4ããã«ã¯æå¿ãããããªãã§ããã»ã»ã»
åŸè¿°ããMultimodal Chain-of-Thought Reasoning in Language Modelsã§ã䌌ãïŒèå¯ããªãããŠããããã«ãæããŸãïŒ
ã
ãåèãŸã§ïŒæµããèšèŒããŠãããŸãïŒ
ïŒç¹èš±çæã®å®äŸã¯èšèŒãã«ããã§ããïŒåŸã»ã©åŸãåãæ€èšŒã§ãããã®ã远èšããŠãããŸãïŒ
ïŒãããã°ã©ã ã®äœæ
ã»image captioningãåãvec2seqã®ã¢ã«ãŽãªãºã ãå©çšããŠããã®ã§æµçšããŸãïŒCNN-LSTMãªã©æ§ã
ãªææ³ããããŸããïŒæ¬èª²é¡ã§ã¯CNN-transformer decoderãè¯ãããã§ãïŒ
ïŒVit-GPT2 https://huggingface.co/nlpconnect/vit-gpt2-image-captioning ããããŸãããïŒã¢ãã«æ§é ãªã©ã«æãå
¥ããå¿
èŠããªããšåæã§ããã°ãã¡ãã§ãè¯ãã£ããããããŸããïŒ
ïŒãåŠç¿ã«å¿
èŠãªããŒã¿éã»ã¢ãã«å®¹éã»ã¡ã¢ãªã»åŠç¿æéã»ãã€ããŒãã©ã¡ãŒã¿ãŒã®ãããä»ã
ã»èªåã®äœ¿ç𿹿³ã§ã¯ãã¯ãã«ç©ºéãæ¯åäœæããããšã«ãªããŸãã®ã§ïŒæ¯ååŠç¿ããããšã«ãªããŸãïŒåŠç¿æéã®ç¢ºèªã¯éèŠã§ãïŒ
ã»ã»åŠç¿ã«å¿
èŠãªããŒã¿éã¯6000ä»¶ã§å¿
èŠã«è¶³ããããã§ãïŒåŠç¿æéã¯cpuã§6æéã»ã©ã§è¶³ããã§ããããïŒã»ãïŒèš±å®¹ç¯å²å
ã§ãïŒ
ïŒè»¢ç§»ã®å¯èœæ§ãdata augmentationã®å¯èœæ§ã¯åŸã»ã©æ€èšããŸãïŒ
ïŒã³ãã¯kaggleãªã©åç
§ãããšè¯ãã§ããã äŸãã°ãã®èŸºã§ãããã
Bristol-Myers Squibb â Molecular Translation
https://www.kaggle.com/competitions/bms-molecular-translation/discussion
äŸãã°ïŒ
https://www.kaggle.com/competitions/bms-molecular-translation/discussion/243932
ãAfter swhiched to transformer from LSTM, I've tried many conbination of patameters.
And then found these insights;
ãImage size is most importatnt, bigger is better.
ãIn my intuition, aspect ratio is useful information, but just resizing to fixed image size works best.
ãIn encoder, Adding positional encoding to only query and key is better than adding to encoder output directly. It's same as DETR does.
ãAs my model got to predict very well, most of trainig data got so easy one, so I thought focal loss works here. And it actually did. It was later verified by my teammate after team mergeing.
ãAfter team merging, my teammate shared how to deal with train/test difference, or CV/LB gap. I've approached to the gap by denoise/noise method(shared above) and pseudo labeling. both worked.
ãBatch size is 128 and train 50 epochs. It takes to train models about 10 days. Without TPU, it would be a month or moreâŠ
ãOther small detail can be found in the notebook.ã
ïŒã衚çŸèœåã®ç¢ºèª
ã»ç¹èš±ãçæããã®ã§ãããïŒmemoryããé©åã«k,vã«é¢ããäž»é¡ãåãç¶ãã§ããããšïŒtgtããé©åã«qã«é¢ããåžžèã芳ç¹ãåãç¶ãã§ããããšïŒããããã©ã³ã¹ã«ããé©åãªåµé çãªæç« ãçæãããããšïŒã¯éèŠã§ãïŒææ³äžã®æ£ããã¯äºã®æ¬¡ã§è¯ãïŒïŒããã確èªããŸãïŒèŠäºïŒã¢ãã«ã®ã¢ã«ãŽãªãºã ã調æŽããŸãïŒ
ã»ã»ãããã¹ããã®ãå«ãŸãªãã衚çŸãé©åã«çæã§ãããããã§ãïŒ

ïŒç»å±±ã§ã¯ããã¯ããã¯ãèè² ã£ãŠããããšãåžžèã§ããã®ã«èè² ã£ãŠããªãïŒãã®éåžžèãææã§ããŠããå¯èœæ§ïŒããªãŒã¯ã©ã€ãã³ã°ã§ã¯ããã¯ããã¯ãèè² ããªãããšãåžžèã§ã¯ããã§ãããããã©ã©ã®èгç¹ã§èŠãŠãããã«äŸåããã§ãããïŒ
ïŒåé¡ãå€ãããã§ãããïŒpretrained decoderãžã®åãæ¿ããæ€èšããŠãè¯ãã§ãããïŒãã ãïŒå®æã«éã¿ãæµçšãããšïŒããããæç« ã¯çæããããå¿
èŠãªæç« ã§ã¯ãªããïŒhallucinationãªã©ïŒç®çãéæã§ããªããªããããããŸããã®ã§æ³šæãå¿
èŠã§ãïŒ(å€§èŠæš¡èšèªã¢ãã«ã䜿ãããšæ¥œãªãã§ãããã©ãããhttps://huggingface.co/docs/hub/adapter-transformers )
ïŒãmemoryæ
å ±éãåæžããå Žåã®è¡šçŸèœåã®ç¢ºèª
ã»CNNã®åºåã¯(7, 7, 512)ã§ããããã(49, 512)ã«å€æãdecoderã®memoryãšããŠããŸãïŒtfidf/mbart/mt5/CLIP embeddings cluster reconstructing visã®æŠå¿µã®åºåã¯å¯å€ã§ãã(1, 300)ã»ã©ã§ãïŒç¹èš±ãçæããå ŽåïŒäœ¿ããæ
å ±éã¯ïŒãããå°ãªããªããŸãïŒç¹ã«äœçœ®æ
å ±ã¯å€±ãããŠããŸãïŒ(49, 512)ã(1, 512)ãšãããªã©å€æãè¡ãïŒïŒãšåçã®è¡šçŸãå¯èœã確èªããŸãïŒãŸã(1, 200)ã»ã©ã«å転ããïŒåçã®è¡šçŸãå¯èœå確èªããŸãïŒ
ã»å¿
èŠã«å¿ãïŒæ
å ±ãå¢ãããŸãïŒ
ã»ã»äºæ³ä»¥äžã«ã¢ãã«ãã³ã³ãã¯ãã«ããå¿
èŠãããã¡ãã£ãšé©ããŠããŸãïŒ
å¯èœæ§ã®å¶çŽãããªããšé ãæ¯ãçµã£ãŠåŠç¿ããªããã®ããã§ã¡ãã£ãšé¢çœããšããã§ãïŒ
ã»ã»memoryå
¥åã«ä»ãïŒäœçœ®æ
å ±ãé€å»ããã»ããlossã¯äœããªããŸãããã»ã»ã»ãªãã ãããã»ã»ã»ïŒ
äœçœ®æ
å ±ã倱ãããç»åã®äžå¿ãšæç« ã®äž»äœãããããããªãããšæã£ãã®ã§ããïŒ25,512ãããã«ç»åãšæç« ã§å
±éããäž»äœãçŸãããããªã£ãŠãããšäºæ³ããïŒäœçœ®æ
å ±ããããšåŠç¿ããããã ãããšã®äºæ³ïŒãã®ãããªåŸåã¯æç¢ºã«ã¯ç¢ºèªã§ããŠããŸããïŒïŒæèæŽçã¡ã¢ïŒç»å49ã®ãã¡48ã岩ïŒ1ã人ãšããïŒå¹³åããã°äººã®æ
å ±ã¯1/49ãšãªãïŒãã®æ
å ±ãäž»ãšããŠçæãããæç« ã¯å²©ãäž»äœãšãªãããšãèªç¶ã ããïŒããããããªã£ãŠããªãïŒä»®èª¬ïŒ: memoryåŽïŒç»åã®äººéšåã«ãã岩ããéãéã¿ããã1/49*xã§ãã£ãïŒãããtransformer encoder ãšã¯ç°ãªãïŒCNNã¯ããã§ããã¢ã«ãŽãªãºã ã§ã¯ã¯ãªãïŒä»®èª¬ïŒïŒqåŽïŒæç« ã®åŠç¿ã«ãã岩ãã人ãäž»äœãšãªãããããšåŠç¿ããïŒ1/49ãšãªã£ã人æ
å ±ã§ãåãåºãããããªã£ãŠããïŒä»®èª¬ïŒã劥åœã ãããïŒãããšïŒããšç»åã«ãããŠã¯ïŒmemoryåŽã®äœçœ®æ
å ±ã¯äž»äœãã©ãã«çŸãããããããªã以äžããã»ã©éèŠã§ã¯ãªãïŒäž»äœã¯äžå€®ã«çŸãããããåžžã«ããã§ããªãïŒïŒç¿»èš³ã§ã¯memoryåŽã®äž»äœã®äœçœ®ã¯ã»ãŒåºå®ã§ããéèŠïŒïŒqåŽã®åŠç¿æç« ããäž»äœãšãªãããããã®ãåãåºãã®ã¿ã«ããã»ããæŽçãããããïŒãšãªã£ãŠããã®ãããããªãïŒïŒ
ã»ã»memoryå
¥åã«ä»ãïŒå転ïŒãã¥ã¢ã³ã¹ïŒãããŠããk,vã«å€æããã»ããlossã¯äœããªãããã§ãïŒ
k,vã®å¿
èŠãªéšåã«decoderããã®éç®ã«ããäœçœ®æ
å ±ãä»äžãçŽããŠããïŒå転éšåã«äœçœ®æ
å ±ãåŠç¿ãããŠããïŒã»ã»ã»ã®ã§ããããïŒïŒå€åæç®ãããã§ãããïŒæ¢ããŠã¿ãŸãïŒïŒïŒè¿œå ããå±€ã«ãããŠencoderã®ç¬ç«æåãæœåºããdecoderã®äž»é¡ãšäžèŽããæåã«attentionãåœãŠãããããŠããã®ã§ããã?ïŒïŒICAã®ã®ããã«èªç±åºŠãäžãç¬ç«æåãæœåºããå Žåã¯ã©ããªãã§ããããïŒ
ã¢ãã«ã¯ãã£ãšãããªæãã§ãïŒ
倿°åãç»åç³»ã®åç§°ãšãªã£ãŠããããšæ±ãã§ããååŒããŠãã ããïŒ
class vec2seqModel(nn.Module):
def __init__(self, n_head, n_decoder_layer, vocab_size, embedding_size,
dim_feedforward=2048, dropout=0.1, dtype=None):
super(vec2seqModel, self).__init__()
self.pos_encoder = PositionalEncoding(embedding_size, 0.1)
self.pat_encoder = PatentClassEncoding(embedding_size)
self.TransformerDecoderLayer = nn.TransformerDecoderLayer_custom(d_model=embedding_size, nhead=n_head, dim_feedforward=dim_feedforward, dropout=dropout)
self.TransformerDecoder = nn.TransformerDecoder(decoder_layer=self.TransformerDecoderLayer, num_layers=n_decoder_layer)
self.embedding_size = embedding_size
self.embedding = nn.Embedding(vocab_size , embedding_size)
self.last_linear_layer = nn.Linear(embedding_size, vocab_size)
self.memorydimconvertliner = nn.Linear(200, embedding_size)
self.init_weights()
def init_weights(self):
initrange = 0.1
self.embedding.weight.data.uniform_(-initrange, initrange)
self.last_linear_layer.bias.data.zero_()
self.last_linear_layer.weight.data.uniform_(-initrange, initrange)
self.memorydimconvertliner.bias.data.zero_()
self.memorydimconvertliner.weight.data.uniform_(-initrange, initrange)
def generate_Mask(self, size, decoder_inp):
decoder_input_mask = (torch.triu(torch.ones(size, size)) == 1).transpose(0, 1)
decoder_input_mask = decoder_input_mask.float().masked_fill(decoder_input_mask == 0, float('-inf')).masked_fill(decoder_input_mask == 1, float(0.0))
decoder_input_pad_mask = decoder_inp.float().masked_fill(decoder_inp == 0, float(0.0)).masked_fill(decoder_inp > 0, float(1.0))
decoder_input_pad_mask_bool = decoder_inp == 0
return decoder_input_mask, decoder_input_pad_mask, decoder_input_pad_mask_bool
def forward(self, encoded_image, decoder_inp):
encoded_image = encoded_image.permute(1,0,2)
encoded_image = self.memorydimconvertliner(encoded_image)#å€åå±€ãè¶³ããªã
decoder_inp_embed = self.embedding(decoder_inp)* math.sqrt(self.embedding_size)
decoder_inp_embed = self.pos_encoder(decoder_inp_embed)
decoder_inp_embed = self.pat_encoder(decoder_inp_embed)
decoder_inp_embed = decoder_inp_embed.permute(1,0,2)
decoder_input_mask, decoder_input_pad_mask, decoder_input_pad_mask_bool = self.generate_Mask(decoder_inp.size(1), decoder_inp)
decoder_input_mask = decoder_input_mask.to(device)
decoder_input_pad_mask = decoder_input_pad_mask.to(device)
decoder_input_pad_mask_bool = decoder_input_pad_mask_bool.to(device)
decoder_output = self.TransformerDecoder(tgt=decoder_inp_embed, memory=encoded_image,tgt_mask=decoder_input_mask, tgt_key_padding_mask=decoder_input_pad_mask_bool)
final_output = self.last_linear_layer(decoder_output)
return final_output, decoder_input_pad_mask
ïŒãæ¥æ¬èªã§åçã確èª
ã»æ¥æ¬èªã§åçã確èªããŸã
ïŒïŒãšçµ±åããŠãè¯ãã§ããã
ïŒãmemoryãïŒtfidf/mbart/mt5/CLIP embeddings cluster reconstructing visã®æŠå¿µã®ãã¯ãã«ã«çœ®ãæã確èª
ã»å®éã®åäœã確èªããŸã
ã»å¿
èŠã«å¿ãïŒæ
å ±ãå¢ãããŸãïŒ
ã»ç»åè±æã§å¯èœã§ããç¹èš±æç« æ¥æ¬èªã§å¯èœã§ãªãéšåã«ã€ããŠã¯ïŒçŽäº€å¯èœãªçç±ãæ¢çŽ¢ã§ããã¯ãã§ãïŒçç±ãèŠã€ã解決ãç¹°ãè¿ããŸãïŒ
ã»ã»åäœç¢ºèªæçæäŸ
Predicted caption : æ¬ é瀺 çºæ ã¯ ïŒ ã¢ã ããŒã æ§ é害 ïŒ ã¢ã ããŒã äžå
š åã³ é害 ã® æ²»ç åã³ äºé² å㯠äºé² åã³ ãã® ãã ãª çŸæ£ ã åŠçœ® ã ïŒ ãã® äœ¿çš æ¹æ³ ã«é¢ãã æ¹æ³ ã æäŸ ã ã ïŒ ãã® ãã ã« æäŸ ãã ïŒïœïŒåŸåçç¥ïŒ
ãããäœé »åºŠèªãéèŠããŠãè¯ãããã§ãïŒååŠçã¯è©°ããå¿
èŠãããã§ãããïŒæç« ã¯ããå°ãé·ããšè¯ãã§ããïŒäœé »åºŠèªãåååãå
¥ãã150åèŠçŽãšãªããšçæ³ã§ããããïŒ
ã»ã»lossã«ã¯å·¥å€«ã®äœå°ãããããã§ãïŒ
ã€ã³ãµã€ããåŸãã«å¿
èŠãªåèªã¯æ±çšèªã§ããããäœé »åºŠèªã§ããããïŒåãæã®ç¥èãåèµ·ããã®ã¯ã©ã¡ãã§ããããïŒãŸãïŒBRIOã®ããã«å質è©äŸ¡ãå ããŠãè¯ããããããŸããïŒ
ãåèªçµ±èšãæå€±é¢æ°ã«åãå
¥ããæ·±å±€åŠç¿ã«ãã倿§ãªéè«å¯Ÿè©±çæ
ãhttps://www.jstage.jst.go.jp/article/tjsai/37/2/37_37-2_G-L62/_pdf
ã»ã»å€§ããããbatchãçšããªãããšãéèŠã§ããïŒ
ãã€ãºã«å¯Ÿããé 匷æ§ã¯ããã»ã©éèŠã§ãªããŸãããçšåºŠã®äœé »åºŠèªãæè¿ãããªãïŒç¢ºãã«ãã®éããšãªããããã§ããããïŒ
ãgoogle: A playbook for systematically maximizing the performance of deep learning models.
ãhttps://github.com/google-research/tuning_playbook#why-shouldnt-the-batch-size-be-tuned-to-directly-improve-validation-set-performance
ããïŒåæãšããŠããããµã€ãºäŸåæ§ã¯ãªããã¹ãã ãšããŠïŒããããµã€ãºãå°ãããšïŒãµã³ãã«ã®åæ£ã«ããåŠç¿ã¢ã«ãŽãªãºã ã«å€ãã®ãã€ãºãæ··å
¥ãïŒãã®ãã€ãºãæ£ååã®å¹æãããããå¯èœæ§ããããŸãïŒã
ããªãã»ã©ãããå°ãä»ã®æ£ååææ®µãèŠçŽããŠãèŠããã
ã»ã»å°éããŒã¿ã«åºã¥ããšïŒè€è£œã«ããé«å質ã«èŠããããšããããŸãïŒ
ãã®å©ç¹ã𿬠ç¹ãèæ
®ãïŒãŸãïŒhuman, train, train dev, dev, testã®ã©ãã«å·®ãããã確èªã察å¿ãã¹ããªã®ã¯ã¢ãã«ã®è€éããªã®ãæ£ååãªã®ãè©äŸ¡ãªã®ãçã
èãïŒé²ããŠããå¿
èŠããããŸãïŒ
ïŒç¹èš±æç« ã®ãã¹ãŠãåŠç¿ããŠããŠã¯ïŒèªåã®æ³å®ããéçšã«ãããŠïŒæéããããããããšããããããããŸããïŒåœãªæç« ã®éžæã»åæžïŒæè¡åžžèã芳ç¹ã®è¡šçŸãå¶éããïŒãïŒè»¢ç§»ïŒãã¡ã€ã³å€ã®æè¡åžžèã芳ç¹ã®è¡šçŸã蚱容ããïŒãæ€èšããã»ããè¯ãããã§ãïŒ
ïŒããç»åèŠçŽ ã¯ïŒãã®èª¬ææã«ååšããããåèªèŠçŽ ã«å¯Ÿå¿ããŠããã¯ãã§ãïŒããç¹èš±æç« ããäœã£ããã¯ãã«ã¯ïŒåãç¹èš±æç« ã«ååšããããåèªèŠçŽ ãå«ãã§ããã¯ãã§ãïŒäžæ¹ïŒç¹èš±æç« ãããã®åèªèŠçŽ ãé€å»ããŠãïŒéåæãªãæç« ãäœãããšã¯å¯èœã§ãïŒãããŠè©²åœããåèªèŠçŽ ã¯ããããŸããã®ã§ïŒé©åœãªæç« ã®éžæã»åæžãããã«ãããŠïŒé€å»ããŠããŸãããšããããŸãïŒãã®éåæã®ãªãæç« ã«ãããŠã¯ïŒæç¢ºãªå¯Ÿå¿é¢ä¿ã¯åŽ©ããŠããŸãïŒãã®éåæãªãæç« ã«ãããŠåèªèŠçŽ ãå圢æãããããšããã°ïŒæ®ãããè€æ°ã®èŠçŽ ãçµã¿åãããïŒæ¯èŒçé£ããåŠç¿ãå¿
èŠãšãªãã§ãããïŒãã®éåæã®ãªãæç« ãåŠç¿ã«çšããå ŽåïŒã©ã®ãããªå·¥å€«ãå¿
èŠã§ããããïŒ
ïŒãç®çã«é©åããããã«ïŒèª¿æŽã»æ¹åã»æ©èœè¿œå
ã»ç®çã«é©åããããã«ïŒå皮調æŽã»æ©èœè¿œå ãè¡ããŸãïŒ
ã»ãäœé »åºŠèªãèæ
®ããŠçæããŠæ¬²ãããïŒç¹å®ã®äœé »åºŠèªã®ã¿ãéèŠããŠã»ããããªãïŒå€æ§ãªçæãã§ãã€ã€ïŒæã®æ§é ã¯æ§é åãããŠããïŒã§ããã°åšèŸºã®ç©ºçœç¯å²ãã©ã³ãã ãã«åãã倿§æ§ã調æŽã§ããŠããŠæ¬²ããããšããè©äŸ¡ææšãäœãïŒæé©ãªæ¡ä»¶ãæ¢çŽ¢ããŸãïŒ
è©äŸ¡ææšããé©åœã«ã§ããã°ïŒoptunaãwandbã®sweepã§æ¢çŽ¢ã§ããŸããã»ã»ã»
ChatGPTãInstructGPTã®æµããèŠãéãïŒãã¥ãŒã¹ããªãã¯ãªè©äŸ¡ææšã§ã¯ãªãïŒDiscriminatorã®ãããªå¥ã¢ãã«ã«ããè©äŸ¡ãçå£ã«èæ
®ããŠãè¯ãããã§ãã»ã»ã»
ã»ç¹èš±åé¡ãå ããŠïŒãã粟床ãé«ããããïŒpromptãšããŠä»»æã«ç¹èš±åé¡ãæç€ºããã®æ¹åãããã«èæ
®ããç¹èš±çæãã§ããããããšïŒå©çšå¯èœæ§ãåäžãããã§ãïŒ
ã»ããããžãŒãçšãïŒç©ºçœé åã®åºæ¬ç¹èš±ã®çæã詊ã¿ãŸãïŒ
ãhomcloudïŒ
ãããŒã·ã¹ãã³ããã¢ããžãŒ(persistent homology)ã«ããšã¥ããããŒã¿è§£æãœãããŠã§ã¢ã§ãïŒ
ãã3次å
ç¹éåããŒã¿(ãã€ã³ãã¯ã©ãŠã)ã®è§£æ
ããhttps://homcloud.dev/py-tutorial/pointcloud.html
次å
å§çž®ãè¡ã£ãŠãããããžãŒã®ç¹åŸŽã¯æ®ããäŒãã®ã§ããçšåºŠäœæ¬¡å
ã§èšç®ããŠãè¶³ããã§ãããã
ã»ã»èª¿æŽåçæäŸ
Predicted caption : æå¹ æå ã 嫿 ã ãªã æ³šå° å€ ã§ ãã£ ã æ³šå° å€ ïŒ æå¹ æå ã num ïŒ
嫿 ã ㊠ãªã 該 ã«ã«ã·ããã³ æå ã num ïŒ threnum num ïœ threnum ïŒ
ã® ç¯å² ã« ãã 該ïœïŒåŸåçç¥ïŒ
ãæå¹æåãå«ãŸãªããšããŠããäºæ
ãšæå¹æåã远å ããããšãšãªã£ãæçŸ©ã¯äžäœäœãªã®ãã»ã»ã»æ¬æ°ãªçºæ³ïŒã§ãããïŒïŒãšããããå«ãŸãªã衚çŸãåçŸã§ããããšã«ã¯æºè¶³ã§ã¯ãããŸãïŒåšèŸºèªåœã«å±éºãªæåãããå«ãŸãªããçæããäºæ
ããã£ãïŒæè¡åžžèãèæ
®ããïŒçæã§ãããšå¬ããã§ããïŒ
äœé »åºŠèªãšããŠã«ã«ã·ããã³ãæ¡çšãããŠããŸããïŒ
æ°å€ã¯ãã¡ããã¡ãã§ããããã¯æåããåœãŠã«ããŠããŸããïŒ
ã»ã»ããããã®çæã¯ã§ããã®ã§ããïŒåŠç¿æéãäºå®ããªãŒããŒããŠããŸããŸããïŒ
CPUã§æŒãéãããšããŠããã®ã§è²ã
工倫ããªããã°ãªãããªããªãé£ãããšããã§ãïŒ
æ§èœãããŸãèœãšããåŠç¿æéãåæžããã«ã¯ããã«ã©ãã§ãããâŠé£ãããã§ãã転移ã§ãããïŒè»¢ç§»ããå ŽåïŒæ€èšŒãã©ããããé£ããã®ã§ãããïŒèªåã䜿ãç¯å²ãªãåããããªãã¡ã€ã³ã®æ¯éå£ãçšããããšã«ãªãã§ããããã転移æç¡ã§å·®ã¯æ®ã©ç¡ãã§ãããããã©ïŒãããã¹ããã©ããã»ã»ã»
ã»ã»ç¹èš±åé¡äºæž¬ãå«ããçæäŸ
['<A61M>', '<A61J>', '<non>', '<non>', '<SEP>', 'ã·ãªã³ãž', 'å€', 'ç', 'ã®', 'å
端', 'éš', 'ã¯', 'ä»', 'ã®', 'å»ç', 'çšå
·', 'ã®', ïœ
å顿
å ±ãããã°ããé©åã«åŠç¿ã§ãçæãããããã®ã§ã¯ãšæããšåæã«ïŒçææã«åé¡ãpromptãšããŠäžããã°ïŒä»»æã®äœé »åºŠèªãçæã§ãããããªãã®ã§ã¯ãšæåŸ
ããŠããŸãïŒãŸãïŒããæŒç¹¹çãªçµæãå€ãäžããããšã«ããïŒãããããªãã®ç¯å²ããåŠç¿ïŒæç€ºã§ããããã«ããªããšæåŸ
ããŠããŸãïŒ
ãªãïŒã¡ã€ã³ã°ã«ãŒãåäœã§ãè¯ãããã§ãïŒ
ïŒself-askã®ãããªææ³ãšã¿ãªããŠãè¯ãã®ãããããŸãããïŒ
ã»ã»åé¡ã¯ããŒã¯ã³ãšããŠäžããŠããïŒåé¡ããŒã¯ã³ã«ã¯ç¹å®ã®encodingãæœããŠããŸãïŒïŒtransformerã«éèšèªæ
å ±ãäžããææ³ã¯ãããããããŸããïŒããŒã¯ã³ãšããŠäžããŠã倧差ãªããšã®æç®ãå€ãããã§ããã®ã§ïŒãã®ææ³ãæ¡çšããŸããïŒãŸãç¹èš±ã®å顿
å ±ã¯ïŒåèªãšçŽæ¥çžé¢ããïŒããã¹ããšåè³ªã®æ
å ±ãæã€æ
å ±ã§ããïŒãšã¿ãªããŠãåé¡ã¯ãªããšããããïŒããŒã¯ã³ãšããŠæ±ã£ãŠãåé¡ãªãã§ãããïŒïŒ
ïŒæ£èŠè¡šçŸã埮åå¯èœã«åã蟌ã¿è¡šçŸåãäžããïŒãªã©ããããŸã
ïŒè»¢ç§»ãdata augmentationã®å¯èœæ§ãæ€èšããŸãïŒæåãåãå
¥ããŠã¿ãŸãïŒ
ïŒãã«ãã¢ãŒãã«åã«ããæ¹åã¯å¯èœã§ããããïŒç¹èš±åé¡ãåãå
¥ããŠã¿ãã®ãé¢çœããã§ãïŒ
ïŒè€æ°ã®æãçæãdiscriminatorã§ãªã©ã³ãã³ã°ããã®ãæã§ãïŒ
ïŒæŠå¿µãã¯ãã«ãšçæãã¯ãã«ã®çµã¿åãããã·ã£ããã«ãïŒãã®çµã¿åãããæ£ãããäºæž¬ããdiscriminatorãäœãïŒçææããªã©ã³ãã³ã°ããã®ãæã§ãïŒ
ïŒäœçœ®æ
å ±ã¯ãã倧ãããã£ã³ã»ã«ãããã§ããã»ã»ã»å
¥åã®å·¥å€«ããããšè¯ãã§ããããïŒ
ïŒmemorydimconvertlinerãéããåŸã®æœåšç©ºéã«ã€ããŠïŒãããã©ã®ãããªæ§é ãšãªã£ãŠããã®ãèŠãŠã¿ãã®ãé¢çœããã§ãïŒæ¯éå£ã«å¯Ÿãäž»é¡ãæŽçããããã¯ãã«ã«ãªã£ãŠãããšæãã®ã§ãããã»ã»ã»
ïŒé«éåã»å°ã¡ã¢ãªã¯å€§ããªèª²é¡ã§ãïŒæ¬¡ã¯äœ¿ããã§ããããïŒ
ãsparce transformer
ãhttps://arxiv.org/abs/1904.10509
ãããçšåºŠæ·±ãå±€ãåæãšããŠããã§ããããïŒå°èŠæš¡ã§ã¯ã¡ãªãããå°ãªããšãæããŸããïŒå€å°ã§ãã»ã»ã»
ïŒäºæž¬ã«å¯Ÿãã確信床ããããã®ãåºåããŠãè¯ããããããŸãããåŠç¿ã«ã¯äœ¿ããªãã£ããããã«æç« ãšããç¹ãå®åšããŠããç®æãšã空çœç®æãšã§ãäºæž¬ã«ãã¬ã®çšåºŠã«éããããã®ãè峿·±ããšããã§ãã空çœã¯äžå¯èœã§ãããããã空çœã§ããããšãããã§ãããããã®å Žåã¯ãã¬ã倧ãããªãã®ã§ã¯ãªãã§ãããã?
ãTransformerã¢ãã«ãžã®MC Dropoutã®é©çš
ãhttps://www.ai-shift.co.jp/techblog/2518
ïŒãã€ãºäºåŸååžã®çžè»¢ç§»ã«ã€ããŠ
http://watanabe-www.math.dis.titech.ac.jp/users/swatanab/phase_trans.pdf
ã確ççãªçŸè±¡ã§ã¯ããã°ãã°ãæéã®äžçãèŠãéã«ãããŸãç¡éé ã®ç¹ã«ç«ã£ãŠ ããããæéã®äžçãèŠæž¡ãããšããèããããæå¹ã§ããããšãç¥ãããŠããŸãã倧æ°ã®æ³åã äžå¿æ¥µéå®çããŠã£ã°ããŒã®åååãªã©ã
ïŒç·åçã«åç
§
ãæ°å€æ°è±¡äºå ±ããã®å€©æ°äºå ±ã³ã¡ã³ãã®èªåçæ
ãhttps://www.jstage.jst.go.jp/article/jnlp/28/4/28_1210/_pdf/-char/ja
ïŒå€§èŠæš¡èšèªã¢ãã«ããã®ãŸãŸå©çšããæ¹æ³ãšããŠïŒFlamingo 2022 ã®ãããªææ³ããã£ãããã§ãïŒ
ãFlamingo: a Visual Language Model for Few-Shot Learning
ãhttps://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/tackling-multiple-tasks-with-a-single-visual-language-model/flamingo.pdf
def gated_xattn_dense(
y, # input language features
x, # input visual features
alpha_xattn, # xattn gating parameter â init at 0.
alpha_dense, # ffw gating parameter â init at 0.):
"""Applies a GATED XATTN-DENSE layer."""
# 1. Gated Cross Attention
y = y + tanh(alpha_xattn) * attention(q=y, kv=x)
# 2. Gated Feed Forward (dense) Layer
y = y + tanh(alpha_dense) * ffw(y)
# Regular self-attention + FFW on language
y = y + frozen_attention(q=y, kv=y)
y = y + frozen_ffw(y)
return y # output visually informed language features
BLIP-2ãè峿·±ãã§ãã
ãBLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models
ãhttps://arxiv.org/abs/2301.12597
ãhttps://huggingface.co/docs/transformers/main/model_doc/blip-2
FROMAGeã®ã·ã³ãã«ããçŽ æŽãããã§ã
ãGrounding Language Models to Images for Multimodal Generation
ãhttps://arxiv.org/abs/2301.13823
ïŒãã®ä»åè
ãPatentTransformer-2 Controlling Patent Text Generation by Structural Metadata
ãhttps://arxiv.org/abs/2001.03708
ãhttps://www.connectedpapers.com/main/1cd46b2acb8b342c9f45b5a96cf9fd7d84ea5118/PatentTransformer%202%3A-Controlling-Patent-Text-Generation-by-Structural-Metadata/graph
ãReliable Natural Language Understanding with Large Language Models and Answer Set Programming
ãhttps://arxiv.org/abs/2302.03780v1
ç¹ã«å°èŠæš¡ã¢ãã«ã«ãããç¥èã«ããè£äœïŒ
ãA Multitask, Multilingual, Multimodal Evaluation of ChatGPT on Reasoning, Hallucination, and Interactivity
ãhttps://arxiv.org/abs/2302.04023v1
èšèªã¢ãã«ã®åŒ±ç¹hallucinationã¯ç¥èãããã¯ãŒã¯ã§è£ãã°è¯ããšæã£ãŠããŸãããïŒå€§èŠæš¡èšèªã¢ãã«ã§ã¯æŒç¹¹ãæ¯èŒçåŸæãšãåž°çŽãäžåŸæãšããããã§ããïŒæŒç¹¹çç¥èã®ä»äžã¯æé©ã§ã¯ãªãã®ãããããŸããïŒè£ãã¹ãã¯æŒç¹¹ã§ã¯ãªãåž°çŽïŒæ©æ¢°åŠç¿ã«ããçµæïŒãªã®ãããããŸããïŒå°èŠæš¡èšèªã¢ãã«ã§ããå Žåã¯æŒç¹¹ã®ãµããŒããåªå
ãã¹ããªã®ãããããŸãããïŒ
ãA Closer Look at Large Language Models Emergent Abilities
ãhttps://yaofu.notion.site/A-Closer-Look-at-Large-Language-Models-Emergent-Abilities-493876b55df5479d80686f68a1abd72f
LLMã®åµçºèœåã«ã€ããŠä»®èª¬ããŸãšããèšäºãã³ãŒãã®åŠç¿ãå¯äž?
ãMultimodal Chain-of-Thought Reasoning in Language Models
ãhttps://arxiv.org/abs/2302.00923
ãhttps://github.com/amazon-science/mm-cot
t5 encoderã§æç« ç¹åŸŽãmhaã§ç»åç¹åŸŽãåŸãŠïŒããããconcatããæ
å ±ãmemoryãšããŠïŒt5 decoderã«å
¥åããŠããããã ïŒ multihead_attn = nn.MultiheadAttention(embed_dim, num_heads) ã®ãšã attn_output, attn_output_weights = multihead_attn(query, key, value) ã§ãããšããïŒself.mha_layer = torch.nn.MultiheadAttention(embed_dim=config.hidden_size, kdim=config.hidden_size, vdim=config.hidden_size, num_heads=1, batch_first=True) ã〠hidden_states = encoder_outputs[0] ã§ïŒ image_att, _ = self.mha_layer(hidden_states, image_embedding, image_embedding) ã€ãŸãk,vãç»åç¹åŸŽïŒãã®éšåã¯vec2seqãšåãã§ããïŒãã㊠merge = torch.cat([hidden_states, image_att], dim=-1) gate = self.sigmoid(self.gate_dense(merge)) ã§ããšã®shapeã«æ»ã㊠hidden_states = (1 - gate) * hidden_states + gate * image_attïŒ ãdecoderã® encoder_hidden_states=hidden_states ã«ãããšïŒ
çµæãšããŠã¯âŠç»åã®ç¹åŸŽéãããã°å°èŠæš¡èšèªã¢ãã«ãæå³ã®ããæèãæšè«ãã§ããïŒãšçè§£ããŠè¯ãïŒ

gateã®æå³ãä»ã²ãšã€ããããªãã®ã§ChatGPTããã«èãããæ¬¡ãšã®ããšïŒ
ããã®ã³ãŒãã¯ãTransformerã¢ãã«ã®Decoderã«ãããŠãEncoderããã®æ å ±ãšã€ã¡ãŒãžæ å ±ãçµ±åããããã®æ©æ§ãå®è£ ããŠããŸãã
self.mha_layerã¯ãMulti-Head Attentionæ©æ§ã衚ããŸããEncoderããã®hidden_statesãšãã€ã¡ãŒãžæ å ±ã§ããimage_embeddingãå ¥åãšããŠäžããããšã§ãEncoderãšã€ã¡ãŒãžæ å ±ã®äž¡æ¹ãçšããAttentionæ©æ§ãå®çŸããŠããŸããããã«ãããDecoderãäž¡æ¹ã®æ å ±ãçµã¿åãããŠãããæ£ç¢ºãªäºæž¬ãè¡ãããšãã§ããŸãã
gateã¯ããã®Encoderãšã€ã¡ãŒãžæ å ±ã®çµ±åã®éã¿ãå¶åŸ¡ããããã«äœ¿çšãããŠããŸããå ·äœçã«ã¯ãgateã¯sigmoid颿°ãéããmergeïŒEncoderã®hidden_statesãšã€ã¡ãŒãžæ å ±ã®Attentionçµæãçµåãããã®ïŒãå ¥åãšãã0ãã1ã®å€ãåºåããŸãããã®å€ã¯ãEncoderã®hidden_statesãšã€ã¡ãŒãžæ å ±ã®ã©ã¡ããåªå ãããã調æŽããããã«äœ¿çšãããŸããgateã®å€ã1ã«è¿ãå ŽåãDecoderã¯ã€ã¡ãŒãžæ å ±ãåªå ãã0ã«è¿ãå Žåã¯Encoderã®hidden_statesãåªå ããããšã«ãªããŸããã
ãDecoderãã€ã¡ãŒãžæ å ±ãæã£ãŠããå ŽåããããEncoderããã®æ å ±ãšçµã¿åãããããšã§ãããæ£ç¢ºãªäºæž¬ãè¡ãããšãã§ããŸããããããã€ã¡ãŒãžæ å ±ããªãå Žåã¯ãEncoderããã®æ å ±ã ãã䜿ã£ãŠäºæž¬ãè¡ãå¿ èŠããããŸãã
ãã®ãããªç¶æ³ã§ã¯ãEncoderããã®æ å ±ãšã€ã¡ãŒãžæ å ±ãåçŽã«çµåããããšã¯ã§ããŸãããçµåããå ŽåãDecoderã¯ã€ã¡ãŒãžæ å ±ããªãå Žåã§ãããã®åœ±é¿ãåããŠããŸãããã誀ã£ãäºæž¬ãè¡ãå¯èœæ§ããããŸãã
ããã§ãGateã¡ã«ããºã ãå°å ¥ãããŸãããGateã¯ãEncoderããã®æ å ±ãšã€ã¡ãŒãžæ å ±ãé©åã«çµã¿åãããããã«å¿ èŠã§ããGateã«ãããDecoderã¯ãã©ã¡ãã®æ å ±ãåªå ããããå¶åŸ¡ããããšãã§ããŸããã€ã¡ãŒãžæ å ±ãããå Žåã¯ããã®æ å ±ãåªå ããããšã§ãããæ£ç¢ºãªäºæž¬ãã§ããŸããã€ã¡ãŒãžæ å ±ããªãå Žåã¯ãEncoderããã®æ å ±ã ãã䜿çšããŠäºæž¬ãè¡ãããšãã§ããŸããã
ãç³ãèš³ãããŸãããå çšã®åçã誀解ãæã衚çŸã§ãã£ãããã§ããDecoderã®QueryåŽã«ç»åæ å ±ãå ¥åãããããšã¯ãããŸããã
æ£ç¢ºã«ã¯ãDecoderã®QueryåŽã«ã¯ãåã®ã¿ã€ã ã¹ãããã§çæãããããŒã¯ã³ã«å¯Ÿããåã蟌ã¿ãã¯ãã«ãå ¥åãããŸãããã®ãããGateã¡ã«ããºã ã¯ãEncoderããã®æ å ±ãšåã®ã¿ã€ã ã¹ãããã§çæãããããŒã¯ã³ã«å¯Ÿããåã蟌ã¿ãã¯ãã«ãé©åã«çµã¿åãããããã«å¿ èŠãšãªããŸãã
Gateã¡ã«ããºã ã«ãããDecoderã¯ãEncoderããã®æ å ±ãšåã®ã¿ã€ã ã¹ãããã§çæãããããŒã¯ã³ã«å¯Ÿããåã蟌ã¿ãã¯ãã«ãé©åã«çµã¿åãããããšãã§ããŸãããããŠãçµã¿åãããçµæã䜿çšããŠã次ã®ããŒã¯ã³ã®äºæž¬ãè¡ããŸããã
ã€ã³ãµã€ãã¯åŸãããŸããïŒæç« æ å ±ãšç»åæ å ±ã¯å質å䟡å€ã§ãªããšããïŒçéã¯concatã§ããããïŒencoderãšdecoderã®shapeãæããå¿ èŠããããã䜿ããªãïŒshapeãæããææ³ãšããŠäŸãã°å¹³åãããªã©ãããïŒä»åã®å Žåã¯ç»åéšåãè§£éããæ å ±ãdecoderåŽã«ãªãããæå³æ¶å€±ãèµ·ã䜿ããªãïŒæç« ãšç»åã®åæ¹ã®æå³ãç¬ç«ã«ä¿æãã€ã€shapeãæããããã«ïŒgateãçšããïŒãšãããšããã§ãããããt5ãpretrainãšããŠçšããªããªãconcatããŠçµããã ã£ãã®ãããããŸããïŒã»ã»ã»ããããããã¢ããã¿ãŒã§ããïŒ
ïŒBIG DATA IS DEAD
https://motherduck.com/blog/big-data-is-dead/
ïŒèª¬ææçæãçšããåäœè¡åäºæž¬
https://speakerdeck.com/supikiti/yan-jiu-hui-fa-biao-zi-liao
https://github.com/Seth-Park/RobustChangeCaptioning
ã»ç»åãã£ãã·ã§ã³ãä»ããè³æŽ»åããã®èŠèŠäœéšåæ§æ
â髿šåª, 西æ¬äŒžå¿ (éªå€§)ã
https://www.anlp.jp/proceedings/annual_meeting/2023/pdf_dir/B8-2.pdf
è³æ³¢ããèšèªãä»ãç»åçæãããšïŒç»åãšããŠããè¯ãïŒïŒäž»èгçç»åã§ã¯æå³ããªãïŒãžã£ã³ã¡ããã£ã®ç»åã瀺ããŠã仿¹ãªãïŒã®ã§ããè¯ãã ããïŒãã ïŒãã®ãããªç»åãšããŠèŠããŠããæ£è
ã«ãããŠã©ã®ãããªç»åãšãªããã«ã€ããŠïŒå¥ã®èª²é¡ãšããŠè峿·±ãïŒïŒ
ã»JaSPICE: æ¥æ¬èªã«ãããè¿°èªé
æ§é ã«åºã¥ãç»åãã£ãã·ã§ã³çæã¢ãã«ã®èªåè©äŸ¡å°ºåºŠ
âåç°å¯æ, å
Œç°å¯å€§, ææµŠåæ (æ
¶æå€§)
https://www.anlp.jp/proceedings/annual_meeting/2023/pdf_dir/B8-3.pdf
ã·ãŒã³ã°ã©ãè©äŸ¡ãšããŠ
ä»é²ïŒGPT-4ã«äžèšèšèšã«ã€ããŠèããŠã¿ã
- å¥ã«æ±ããé¡äŒŒãã¯ãã«ãšç¹èš±æç« ã®çµã¿åããããŒã¿ãtransformerã«åŠç¿ããããããã¯ãã«ãäžãããšãã«ãããç¹èš±æç« ãçæãããããšããŠããŸãããã¯ãã«ã¯transformer decoderã®memoryã«å šçµåå±€1å±€ãä»ãäžããç¹èš±æç« ã¯ãäžã€ã®æç« ããç¹èš±åé¡ã³ãŒããšç¹èš±èŠçŽæç« ã®çµã¿åãããšãç¹æ¥åé¡ã³ãŒããšç¹èš±è«æ±é æç« ã®çµã¿åããã«åå²ãã2åã«æ°Žå¢ããããã®ãšããããããããŒã¿ã»ãããšããŠäžããŸããã€ãŸããåãé¡äŒŒãã¯ãã«ã§èŠç¹ã®ç°ãªãæç« ãåŠç¿ãããŸããå ã®é¡äŒŒãã¯ãã«ãŒç¹è𱿿žã®æ°ã¯5000ã»ã©ã§ããæé©ãªèšèšãèããŠã¿ãŠãã ãã
ããªããè¿°ã¹ãã¢ãããŒãã¯ãTransformerã¢ãã«ãå©çšããŠç¹èš±æç« ãçæããããã®ææãªæ¹æ³ã§ãã以äžã«ãããªãã®ã¢ãããŒããæé©åããããã®ããã€ãã®ææ¡ããŸãšããŠããŸãïŒ
ããŒã¿ã®ååŠçïŒç¹èš±åé¡ã³ãŒããšç¹èš±èŠçŽæç« ãç¹èš±è«æ±é æç« ãçµã¿åããããšããããããé©åã«åºåãããšãéèŠã§ããããã«ãããã¢ãã«ãããããã®éšåãåºå¥ãããããªããŸããäŸãã°ãç¹æ®ãªããŒã¯ã³ïŒäŸãã°[CLS]ã[SEP]ãªã©ïŒã䜿çšããŠãæç« ã®éå§ãšçµäºãããã³ç°ãªãéšåã瀺ãããšãã§ããŸãã
ã¢ãã«ã®èšå®ïŒã¢ãã«ã®èšå®ïŒäŸãã°ãå±€ã®æ°ãé ãå±€ã®ãµã€ãºããããã®æ°ãªã©ïŒã¯ãã¿ã¹ã¯ã®é£æåºŠãšããŒã¿ã®éã«ãããŸããããªãã®ã±ãŒã¹ã§ã¯ãããŒã¿ã»ãããæ¯èŒçå°ããïŒ5000ã®é¡äŒŒãã¯ãã«ãŒç¹è𱿿žïŒãããéåŠç¿ãé¿ããããã«ã¯ãã¢ãã«ã®ãµã€ãºãå°ããä¿ã€ããšãéèŠã§ãã
åŠç¿ã¹ã±ãžã¥ãŒã«ïŒTransformerã¢ãã«ã¯ãåŠç¿çã®ã¹ã±ãžã¥ãŒãªã³ã°ã«ææã§ããéåžžãåŠç¿çã¯èšç·Žã®åææ®µéã§ç·åœ¢ã«å¢å ãããã®åŸã¯éæ¯äŸçã«æžå°ããŸããããã¯ãã¢ãã«ãåæã«ã¯å€§ããªã¹ããããèžã¿ããã®åŸã¯ãã现ãã調æŽãè¡ãããšãå¯èœã«ããŸãã
æ£ååïŒããŒã¿ã»ãããå°ããå ŽåãéåŠç¿ãé²ãããã«æ£ååææ³ã䜿çšããããšãæå¹ã§ããããã«ã¯ãããããã¢ãŠããéã¿æžè¡°ããŸãã¯æ©æåæ¢ãªã©ã®ææ³ããããŸãã
è©äŸ¡ææšïŒã¢ãã«ã®æ§èœãè©äŸ¡ããããã®é©åãªææšãéžæããããšãéèŠã§ããç¹èš±æç« ã®çæã¿ã¹ã¯ã§ã¯ãBLEUã¹ã³ã¢ãROUGEã¹ã³ã¢ãªã©ã®èªåè©äŸ¡ææšã䜿çšããããšãäžè¬çã§ãããããããããã®ææšã¯å¿ ããã人éã®è©äŸ¡ãšäžèŽããããã§ã¯ãªããããå¯èœã§ããã°äººéã«ããè©äŸ¡ãè¡ãããšãæšå¥šãããŸãã
Using unknown plugin...
æ°ããããŒããã㯠"patent_transformer.ipynb" ãäœæããŸããã以äžã®ãªã³ã¯ããã¢ã¯ã»ã¹ã§ããŸã: patent_transformer.ipynb
ãŸããå¿ èŠãªã©ã€ãã©ãªãã€ã³ããŒãããŸããããæ¬¡ã«ãããŒã¿ã®ååŠçãšã¢ãã«ã®èšå®ã«ã€ããŠèª¬æããŸãããã®åŸãåŠç¿ã¹ã±ãžã¥ãŒã«ãæ£ååãè©äŸ¡ææšã«ã€ããŠèª¬æããŸããæåŸã«ãã¢ãã«ã®èšç·Žãšè©äŸ¡ã«ã€ããŠèª¬æããŸãã
noteable.io logo
noteable.io
Used Noteable
次ã«ãããŒã¿ã®ååŠçãè¡ããŸããç¹èš±åé¡ã³ãŒããšç¹èš±èŠçŽæç« ãç¹èš±è«æ±é æç« ãçµã¿åããããšããããããé©åã«åºåãããšãéèŠã§ããããã«ãããã¢ãã«ãããããã®éšåãåºå¥ãããããªããŸããäŸãã°ãç¹æ®ãªããŒã¯ã³ïŒäŸãã°[CLS]ã[SEP]ãªã©ïŒã䜿çšããŠãæç« ã®éå§ãšçµäºãããã³ç°ãªãéšåã瀺ãããšãã§ããŸãã
ãŸããã¢ãã«ãç¹èš±åé¡ã³ãŒããããã¹ããšããŠè§£éããªãããã«ãç¹èš±åé¡ã³ãŒããäžæã®ããŒã¯ã³ã«å€æããããšãèããããŸããããã«ãããã¢ãã«ã¯ç¹èš±åé¡ã³ãŒããç¹æ®ãªæ å ±ãšããŠæ±ãããšãã§ããŸãã
以äžã«ãããŒã¿ã®ååŠçã®äžéšã瀺ããŸãããã®éšåã§ã¯ãç¹èš±åé¡ã³ãŒããäžæã®ããŒã¯ã³ã«å€æããç¹èš±èŠçŽæç« ãšç¹èš±è«æ±é æç« ãçµåããŸããå ·äœçãªããŒã¿ã«å¿ããŠããã®éšåãé©å®èª¿æŽããŠãã ããã
Used Noteable
Used Noteable
次ã«ãããŒã¿ãããŒã¯ã³åããããŒã¯ã³ãæŽæ°ã«ãããã³ã°ããããã®èªåœãäœæããŸããããã«ã¯ãäžè¬çã«ã
ããã¯ã
ã
ã
æ€èšŒäŸ
æ€èšŒæïŒãæ¬çºæã¯ãïŒã®ïŒ³ïŒïŒ°èŒžéé»å®³ã«ãããªã³ãçã®éèµ°ãé»å®³ããããšã§èªå·±å
ç«çŸæ£æ²»çãæããæ²»çå¹æã®æåŸ
ã§ããæäœãæäŸããããšãç®çãšãããæ¬çºæã¯ãèæ€åç©ã®ïŒ³ïŒ°ïŒ®ïŒ³ïŒã«ç¹ç°çã«çµåããïŒïŒ°ã®èŒžéé»å®³ãä»ãããªã³ãçã®éèµ°é»å®³æŽ»æ§ãæããïŒäžåæäœãããã¯ãã®æçãåã¯ãããã®èªå°äœã§ããã
ãã®æ€èšŒåã«è©²åœããïŒé¡äŒŒåºŠãã¯ãã«ã®ã¿ãã¢ãã«ã«äžãïŒé©åãªåé¡ãæç« ãçæãããã確èªããïŒ
ã»åŠç¿ã«ãã®æ€èšŒæã¯äœ¿ãããŠããªãïŒ
ã»æ¯éå£ã¯ïŒïŒ³ïŒïŒ°ïŒæäœïŒèªå·±å
ç«çŸæ£ïŒããã«é¢ããæç« 矀ã§ã¯ãªãïŒããé¢é£ããªãäž»å€ã«é¢ããç¹èš±çŸ€**00ä»¶ã§ããïŒæç®ã¯å«ãã§ããªãïŒ
ã»S1PãšèšèŒãããç¹èš±ã¯æ¯éå£ã«ïŒä»¶ããïŒãã®ãã¡ïŒä»¶ã¯trainã«ïŒ1ä»¶ã¯validationã«å«ãŸããŠããïŒ
æ¬æ€èšŒæã¯ïŒãã®ãã¡validationã«å«ãŸããŠããæç« ã§ããïŒ
ã»trainã«å«ãŸããŠããæç« ã¯ïŒïŒ³ïŒ°ïŒ®ïŒ³ïŒã«é¢é£ãããªã³ãçéèµ°é»å®³æäœã«é¢ããç¹èš±ïŒC07K,C12NãšïŒïŒ³ïŒïŒ°å容äœã«é¢é£ããæäœã®ç§»æ€æçµ¶äœçšã»èªå·±å
ç«çŸæ£ã»ã¢ã¬ã«ã®ãŒçŸæ£ã«é¢ããç¹èš±ïŒ C07D, A61K, A61Pã§ãã£ãïŒ
ã»2次å
空éã«èœãšããmemoryãã¯ãã«ã®å座æšã®æŠç®å€ã¯ïŒæ€èšŒæ(-13, 6) ïŒtrainåè
(-14, 7)ïŒtrainåŸè
(18, 16)ã§ãã£ãïŒãªãïŒæ¯éå£ã®å
šç¹èš±ã®åç¹ã¯x軞y軞ãšãã«ã»ãŒ-50~+50ã®ç¯å²ã«ååšããŠããïŒ
ããŠïŒæŠå¿µãã¯ãã«ã®ã¿ããïŒæ€èšŒæã«é¡äŒŒããäž»é¡ã«æ²¿ã£ãæ¯éå£ã«å«ãŸãã芳ç¹ãæã€æç« ãçæã§ããã§ããããïŒ
çææïŒïŒãPredicted caption : <A61K> <A61P> <non> <non> <SEP> æ çç è質 ã¡ãã£ãšãŒã¿ãŒ ã å«ã å»è¬ çµæ ç© ã§ ã㣠㊠æ çç è質 ã¡ãã£ãšãŒã¿ãŒ ã® ãšã¹ãã« å ãããã ã« äœ¿çš ãã æ çç è質 ã¡ãã£ãšãŒã¿ãŒ ã å«ã å»è¬ çµæ ç© ãïœ
çææïŒïŒãPredicted caption : <A61K> <A61P> <non> <non> <non> <SEP> æ ææ è¬ æ çç è質 ãªããœãŒã åã³ æ ææ å€ ã å«ã çŒç§ çš çµæ ç© ã æ çç è質 ã¡ãã£ãšãŒã¿ãŒ ã® ãšã¹ãã« åã³ ããã ã å«ã¿ æ ææ ç ã® æ²»ç åã³ ããã ã¯ æ ææ è¬ ã å«ã¿ æ²»ç è¬ ã å«ã å»è¬ çµæ ç© ã¯ æ çç è質 ã¡ãã£ãšãŒã¿ãŒ ã® çµã¿åãã ã§ ãã  num ã® åå¿ çæ ç© ã ïœ
çææïŒïŒãPredicted caption : <C07D> <A61K> <A61P> <C12N> <C07K> 现è çµå èçœ é
å ã å«ã  num åã³ ãã® è¬ç äœçš éšäœ ã« èµ·å ãã è³ è è¬ç åŠ ç è¬å€ ãïœ
ïŒåçææã¯topKã®èšå®ãç°ãªãïŒé ã«,1,2,3ã§ããïŒã€ãŸãçææïŒã¯greedyã§ããïŒ
ïŒèšèŒããªããïŒROUGEïŒæ€èšŒæäžäœé »åºŠäžäœïŒåèªçæçïŒMQM (Multidimensional Quality Metric)ã§è©äŸ¡ããŠããïŒ
çææãçè§£ããããã®èª¿æ»ïŒæ¹åã¡ã¢
现èå
å€ã§æ©èœããã¹ãã£ã³ãŽã·ã³ïŒãªã³é
žïŒS1PïŒã®åœ¹å²
https://www.jbsoc.or.jp/seika/wp-content/uploads/2013/05/84-02-03.pdf
ãSïŒP ã¯ã¹ãã£ã³ãŽèè³ªç±æ¥ã®è質ã¡ãã£ãšãŒã¿ã§ïŒïŒïŒïŒïŒå¹Žä»£ã«çްè墿®äœçšãªã©ãèŠãã ãããããšã奿©ã«ïŒçºèŠããæ°å幎ãçµãŠãã®æçŸ©ãåè©äŸ¡ãããããã«ãªã£ãïŒãŸãïŒSïŒP ã¯ãªãŸãã¹ãã¡ããžã³é
žãšå
±ã«è¡æŒ¿äžã«è±å¯ã«ååšãããªãŸãªã³è質ã§ããïŒãããã®è質ã¡ãã£ãšãŒã¿ã®å容äœãã»ãŒåææã«åå®ãããããšã«ããïŒãããã®ççæ©èœãé£èºçã«è§£æãããïŒäžæ¹ã§ïŒSïŒP ã¯å容äœãä»ããäœçšã®ã¿ãªããïŒçްèå
ã§ã»ã«ã³ãã¡ãã»ã³ãžã£ãŒãšããŠã®äœçšãæããŠããïŒããããåããã£ãŠäžå±€è€éãªäœçšæ©åºã圢æããïŒããã«ïŒã¹ãã£ã³ãŽè質ãšã³ã¬ã¹ãããŒã«ãéç©ããŠåœ¢æãããè質ã©ãããæ
å ±äŒéã®âãã©ãããã©ãŒã âãšããŠã®å ŽãæäŸããããšãšçžãŸã£ãŠïŒSïŒPã¯æ§ã
ãªã·ã°ãã«äŒéç³»ãçµ±ååŠçããâãªãã¬ãŒã¿ãŒååâãšããŠæ©èœãã宿
ãæµ®ãã³äžãã£ãŠããïŒãããŠãã®ççäœçšã¯çްè墿®ã»ååïŒççïŒå
ç«ïŒç¥çµæ©èœãªã©ã®èª¿ç¯ãšå€å²ã«ãããïŒãããã®ç Žç¶»ããçããããïŒèªå·±å
ç«çŸæ£ïŒæ
¢æ§ççãèªç¥çãªã©ã®ç
æ
è§£æãããŠæ°ããªååæšçæ²»çæ³ã®éçºã«ãäžå±€æåŸ
ãé«ãŸãïŒã
ã¹ãã£ã³ãŽã·ã³1-ãªã³é
žïŒS1PïŒèŒžéäœã§ããSpns2ã®ååæ©èœ
https://seikagaku.jbsoc.or.jp/10.14952/SEIKAGAKU.2018.900574/data/index.html
ãããçš®ã®çްèéã·ã°ãã«ååã¯ïŒçްèå
ã§çæãããåŸã«çްèå€ãžãšæŸåºããïŒæšç现èã«çºçŸããã·ã°ãã«ååç¹ç°çå容äœã«çµåããããšã§æ
å ±ãäŒéããïŒãªãŸãªã³è質ã¡ãã£ãšãŒã¿ãŒã®äžçš®ã§ããã¹ãã£ã³ãŽã·ã³1-ãªã³é
žïŒS1PïŒã¯ãŸãã«ãã®ãããªååã§ããïŒè¡ç®¡å
ç®çްèãèµ€è¡çå
ã§ç£çãã现èå€ãžãšæŸåºããïŒããã«ïŒGã¿ã³ãã¯è³ªå
±åœ¹åå容äœïŒGPCRïŒã§ããS1På容äœïŒS1PR1ïœ5ïŒãä»ããŠçްèéèµ°ïŒçްè墿®ããã³ã¢ãããŒã·ã¹ã®æå¶ãªã©ã®çŽ°èæ©èœãå¶åŸ¡ããããšã§è¡ç®¡æ°çïŒè¡ç®¡ééæ§ããã³ãªã³ãçã®åŸªç°ãªã©ã調ç¯ããŠããïŒS1Pã®åé§äœã§ããã¹ãã£ã³ãŽã·ã³ã¯çްèå
ã«å±åšããã¹ãã£ã³ãŽã·ã³ãããŒãŒïŒSPHK1/2ïŒã«ãããªã³é
žåãåããã®ã§ïŒS1Pãæšç现èèäžã«çºçŸããS1PRã«çµåããããã«ã¯ïŒäžåºŠçްèå€ãžãšæŸåºãããå¿
èŠãããïŒS1Pã¯çæ°Žæ§ã®é«ãè質ã¡ãã£ãšãŒã¿ãŒã§ããããªã³é
žåºãååå
ã«ä¿æããããïŒåçŽæ¡æ£ã«ãã现èèãééã§ããïŒäœããã®æŸåºæ©æ§ãååšãããšèããããŠããïŒæ¬çš¿ã§ã¯ïŒæè¿ïŒååå®äœãæãããšãªã£ãS1P茞éäœãäžå¿ã«S1Pã®ççæ©èœã玹ä»ããïŒã
ç¥çµå€æ§çŸæ£ã«ãããã¹ãã£ã³ãŽè質ã®åœ¹å²
https://seikagaku.jbsoc.or.jp/10.14952/SEIKAGAKU.2020.920640/data/index.html
ãã¢ã«ããã€ããŒç
ãããŒãã³ãœã³ç
ãªã©ã®ç¥çµå€æ§çŸæ£ã®ç
ç圢æéçšã«ã¹ãã£ã³ãŽè質ãšåŒã°ããèè質ã®äžçŸ€ãé¢äžããããšã瀺åãããŠããïŒããŒãã³ãœã³ç
ã§ã¯ãªã¹ã¯å åãšããŠã¹ãã£ã³ãŽç³è質ã®åè§£é
µçŽ ã§ããã°ã«ã³ã»ã¬ããã·ããŒãŒéºäŒåå€ç°ãåå®ããïŒé¢é£è質ã¯ã¬ãŽã£å°äœåœ¢æã«é¢äžããå¯èœæ§ãããïŒãŸãã¢ã«ããã€ããŒç
ã§ã¯Aβã¢ããã€ããã¿ãŠç
çãªã©è€æ°ã®ç
ç圢æéçšã«ãããŠã»ã©ãããã¹ãã£ã³ãŽããšãªã³ãã¯ãããšããã¹ãã£ã³ãŽè質ãé¢äžããååæ©åºãææ¡ãããŠããïŒæ¬çš¿ã§ã¯ïŒãããçŸæ£ã«ãããã¹ãã£ã³ãŽè質ã®åœ¹å²ã«ã€ããŠïŒã»ã©ããäŸåçã«ç£çããAβåè§£æ©èœãæã€ãšã¯ãœãœãŒã ã«é¢ããæã
ã®æè¿ã®ç ç©¶ãå«ããŠç޹ä»ããïŒã
ïŒåŠç¿ã«ãããŠïŒæ¬¡ã®åœ±é¿ãåãããããããªã
ãžã§ã³ãœã³ã»ã¢ã³ãã»ãžã§ã³ãœã³6675361
https://www.j-platpat.inpit.go.jp/c1800/PU/JP-2019-178150/8ECB45483FD1355E755F32FB952019362DE6FF55A36794A98DA0F9E3AE5151A6/11/ja
ãã課é¡ããã©ã€ã¢ã€åã³ä»ã®ççæ§ã®çŒã®ç¶æ
ã®æ²»çã®ããã®çŒçšçµæç©åã³æ¹æ³ã
ãè§£æ±ºææ®µãæççè質ã¡ãã£ãšãŒã¿ãŒãšäžäŸ¡ã¢ã«ã³ãŒã«åã¯ã¢ããã®åå¿çæç©ã§ãããæççè質ã¡ãã£ãšãŒã¿ãŒã®ãšã¹ãã«ã§ãã£ãŠãæççè質ã¡ãã£ãšãŒã¿ãŒã®å€§éšåã¯ãçµæç©äžãšã¹ãã«åœ¢æ
ã§ååšããããã®ããã«ãçµæç©ã¯å®è³ªçã«æççè質ã¡ãã£ãšãŒã¿ãŒã®é
žåœ¢æ
ãå«ãŸãªããæççè質ã¡ãã£ãšãŒã¿ãŒã¯ãå€äŸ¡äžé£œåèèªé
žïŒäŸãã°ãÏïŒïŒåã³ÏïŒïŒèèªé
žïŒãã¬ãŸã«ãã³åã¯ä»£è¬çã«å®å®ãªé¡çžäœããããã¯ãã³åã¯ä»£è¬çã«å®å®ãªé¡çžäœããªããã·ã³åã¯ä»£è¬çã«å®å®ãªé¡çžäœãããã¹ã¿ã°ã©ã³ãžã³åã¯ä»£è¬çã«å®å®ãªé¡çžäœãã¬ããã€ã³é
žããšã³ãã«ã³ãããã€ãããããã®ä»£è¬ç£ç©ãåã³ãããã®æ··åç©ãããªã矀ããéžæãããŠããããã®çµæç©ã¯ã補å€ã溶液ãã²ã«ãè»èåã³ïŒåã¯ã¹ããªãããåã³ïŒåã¯ã³ã³ã¿ã¯ãã¬ã³ãºãä»ããŠçŒã®è¡šé¢ã«å±æçã«äŸçµŠã§ãããã
train, 座æšïŒ-30, -8ïŒ,ãã
ïŒçææïŒã«ã€ããŠïŒperplexityã«èããŠã¿ã

ãã³ãããªã§ãããã»ã»ã»åŒçšç¢ºãããã°
ïŒãã®åšèŸºæè¡ãšããã°ãããµã³ã§ããããã©ã»ã»ã»
ïŒããããã¯Cell Penetrating Peptideã®ããšãšçè§£ã§ããããªã»ã»ã»
åŠçæä»£ã¯åå¿ã®å ŽãšããŠã®çäœèãšãªããœãŒã ãå°éã®äžã€ãšããŠãããïŒ
äžèšæç®ãèªã¿ïŒãžâãšæãããšããã§ããïŒ
ããã¯ãããŠãããŠïŒ
çææã¯ïŒäžããäž»é¡ã«æ²¿ã£ãïŒæ¯éå£ã«å«ãŸãã芳ç¹ãæã€ïŒæç« ãšãªã£ãŠããããã«ã¿ããŸãïŒ
ïŒããå°ãïŒæ£è§£ã§ãããæäœããéèŠããçæãããŠã»ããã£ãæ°ã¯ããŸãïŒ
æ¯éå£ãæäœãèŠåãããããšã¯ãããªãæ¯éå£ã§ãã£ãã®ã§åŠ¥åœãããããŸãããïŒããå°ãç¹èš±åé¡ã®åœ±é¿ããããŠãè¯ãã£ãããšïŒ
ãµãã¯ã©ã¹åäœã§ã¯ä»æ¹ãªãããããŸããïŒè€æ°åçæããã°æäœãšãªãããã§ã¯ãããŸãïŒgreedyã§ãªãçæããçµæãèŠãéãïŒæäœã§ããæ
å ±èªäœã¯æ®ã£ãŠããããã§ãïŒ
ïŒæŠå¿µãã¯ãã«ã«å ãïŒç¹èš±åé¡ãä»»æããŒã¯ãŒããpromptãšããŠäžããå Žåã®çææã«ã€ãïŒè©äŸ¡ãé²ããŠãããŸãïŒä»»æããŒã¯ãŒããšããŠæ°èŠåºçŸããŒã¯ãŒããçšããã°ïŒããã¯ãã£ã¹ããæèããçæãã§ãããã§ãïŒè峿·±ãã§ããïŒ
ïŒåŸ®èª¿æŽã¯åŒãç¶ãç¶ããŠãããŸãïŒ
âãšããããrouge-1ãç®çãšããŠoptunaã«ãããŸããïŒ
æŽã«ã¢ãã«ãå°ãããã¹ãããã§ãïŒããŒã¿äžè¶³ãšããŒã¿å
çæã§è¶³ããæ¡ä»¶ãããããšãããªãã§ãããããïŒ
train-validå·®ã¯ãã£ãŠè¯ã課é¡ã§ããïŒããå°ãè©°ããããã«ïŒdata augmentationããã¹ãã§ãããïŒ
åé¡ãã¡ã€ã³ã°ã«ãŒãåäœã詊ããŠã¿ããã§ãïŒ
embeddingã«æ¬äŒŒçãªã¿ã¹ã¯è§£æ±ºçšã¢ãã«ãçµã¿èŸŒãŸãããšæãããŸããïŒæå€ãšå°ãªã次å
ã§è¡šçŸãããããã§ãïŒã¿ã¹ã¯ãç°¡åãããŸãããïŒã¿ã¹ã¯ãè€éã«ããããã«ãïŒæºåãdata augmentationããã¹ãã§ãããïŒ
æ¬¡ã¯æ¬ä»¶ã§ã¯è©²åœããªããšæããŸããã»ã»ã»
ãïŒNews Summarization and Evaluation in the Era of GPT-3
ãhttps://arxiv.org/abs/2209.12356
ãGPT-3ã§ãã¥ãŒã¹èŠçŽãè¡ããšROUGEã¯SoTAææ³ããã¯äžã ã人æè©äŸ¡ã ãšäžåãçµæãšãªãããïŒ
ïŒdata augmentationã«ã€ããŠïŒæããtfidfäžäœèªãæãæããåèªãBERTã§çœ®ãæãããªã©ãããŸããïŒæ¬èª²é¡ã§ã¯æ¯éå£ã«ãªãæŠå¿µãå¢ãããããªãã®ã§äœ¿ããŸããã»ã»ã»ã©ããããã®ãïŒtfidfãäœãåèªã§ããã°ïŒèªåœãå¢ããŠãè¯ãããã§ãïŒ
ïŒããªãå°ãªããµã³ãã«ãããšã«ããŠçæããŠãããšããïŒrouge-1ã¯35ïœ40çšã§ãïŒ
ããŸã 工倫ã®äœå°ã¯ããããã«èŠããŠããŸãïŒ
ïŒç¹ã«lossã«ã¯ãŸã ãŸã 工倫ã®äœå°ãããããã§ãïŒæ¬¡ã§èª¿æŽäžã§ãïŒ
ãCrossEntropyLoss + LanguageModelCriterion + L1æ£åå + weight_decay(L2æ£åå) + IFTã»ãäœé »åºŠèªéèŠ
ïŒããKL collapseããšã¯ãæœåšå€æ°ã®æ
å ±ããã³ãŒããŒã«ååã«äŒéããããçæããã»ã¹ãåã«äºåååžã«äŸåããçŸè±¡ã§ããããã«ãããæœåšå€æ°ãåŠç¿ã«å¯äžãããã¢ãã«ã®è¡šçŸåãæãªãããå¯èœæ§ããããŸãã
ãKL annealingãã¯ãKL collapseãé²ãããã®ãã¯ããã¯ã§ãããåŠç¿ã®åææ®µéã§ã¯KLãã€ããŒãžã§ã³ã¹é
ïŒæœåšå€æ°ã®äºåŸååžãšäºåååžã®éãã衚ãé
ïŒã®éã¿ãåŸã
ã«å¢ããããšã§ãæœåšå€æ°ã®æ
å ±ããã³ãŒããŒã«åŸã
ã«åãå
¥ããããšãã§ããŸãããã®æ¹æ³ã«ãããæœåšå€æ°ãåŠç¿ããã»ã¹ã«é©åã«çµã¿èŸŒãŸããã¢ãã«ã®çæèœåãåäžããŸããã
äºååŠç¿æžã¿èšèªã¢ãã«ãdecoderã«äœ¿çšããå Žåã«æ³šèŠããç¹ã ãšæããŸããïŒã¡ã¢ããŠãããŸãïŒ
ïŒoptunaã®ç®æšå€ãïŒrouge-1, rouge-2, rouge-l, BERTscoreã®éèšãšããŸããïŒ

ãäºå詊éšäžã®ãã€ããŒãã©ã¡ãŒã¿ãŒã®éèŠåºŠã§ããïŒã ããããããªãšãã§ããã
ïŒæç« ãã®ãã®ã«patent encodingãå ãåé¡ãåã®äžéšãšããŠçæããæ§é ãããïŒã«1å±€çµåãåé¡ãäºæž¬ããããã«ãã¿ã¹ã¯æ§é ãšããã»ããïŒé¢çœããããããŸãããã»ã»ã»
ïŒtransformerdecoderlayerã®æ§é ãpost-Lnããpre-LNã«å€æŽããã»ããè¯ããããããŸããã
ïŒToken Imbalance Adaptation for Radiology Report Generation
https://arxiv.org/abs/2304.09185v1
ITF以å€?ã®äœé »åºŠèªåªå
ææ³?
ïŒAttentionViz: A Global View of Transformer Attention
https://arxiv.org/abs/2305.03210

headã®æ
å ±æ§é ã®ã€ã¡ãŒãž
â»ç¹èš±æ
å ±ã«åºã¥ãçæAIã§æè¡è
ã®ç¹èš±æ
å ±å©çšãæ¯æŽ
çæAIã«ããããŒã¯ãŒãææ¡ãèŠçŽææžçæã«ããç¹èš±æ
å ±æŽ»çšãšçºæåµåºãå éãAmplifiedã20230703
https://prtimes.jp/main/html/rd/p/000000016.000033344.html
amplified aiãç¹èš±æç®ãåºã«ããæšè«ãææ¡ãè¡ãæ©èœãæäŸ
https://japan.zdnet.com/article/35206108/
ç¹èš±çæã¯åºãŸããŸããã
â»Scaling Laws and Interpretability of Learning from Repeated Data
https://arxiv.org/abs/2205.10487
éè€ããŒã¿ã¯å±éº?
ã
ChatGPTã»GPT-4ã«ããè£å©çæ
çææã«å¯ŸããŠããã®ç¥èã䜿ãã€ã³ãµã€ããåŸãã«ããããããã®ç¥èãè¶³ããªãå ŽåããããŸãã
ãªãããã®è£å©ããããšè¯ãã§ãããã
çæããæç« ããChatGPT API ã§æ¹å€ïŒè£å©çæãããšè¯ãããã§ãã
次ã®è³ªåãªã©ããåŸãã§ãããã
ãæ¬¡ã®æç« ããåŸæ¥æè¡ãšã課é¡ãæè¡åžžèãå ãæ°Žå¢ããŠãã ãããã
ãæ¬¡ã®æç« ã®äž»é¡ãšãªãåèªã10åæãåºããæãåºããããããã®åèªã«ã€ãããã®åèªã®æŠèŠã説æããŠãã ãããã
ãäŸãã°ãçææïŒã§ã¯ã次ã®ããã«è£å©çæãããŸãã
ãåŸæ¥ã®æè¡ã«ãããŠãæçç广ãæã€è¬å€ã¯ãçŸæ£ã®æ²»çã«åºãçšããããŠããŸãããæççè質ã¡ãã£ãšãŒã¿ãŒã¯ããã®äžã§ã泚ç®ãããååç©ã®äžã€ã§ãããåŸæ¥ãããã®å¹æãç¥ãããŠããŸãã
ããããæççè質ã¡ãã£ãšãŒã¿ãŒãåç¬ã§äœ¿çšããå Žåããã®å¹æã¯å¶éãããããšãç¥ãããŠããŸããããã§ãè¿å¹Žããšã¹ãã«åããããã®äœ¿çšã泚ç®ãããŠããŸãããšã¹ãã«åããããã¯ããããããè質éã«çµåãããããšã§ãè¬ç©ã®æç¶æ§ãé«ããããšãã§ããŸãããã®ãããæççè質ã¡ãã£ãšãŒã¿ãŒããšã¹ãã«åãããããšããŠäœ¿çšããããšã§ããã匷åãªå¹æãæåŸ ã§ããŸãã
ããããªããããã®æè¡ã«ã¯ãŸã 課é¡ãæ®ãããŠããŸããäžã€ã®èª²é¡ã¯ããšã¹ãã«åããããã®åææ¹æ³ã®æ¹è¯ã§ããçŸåšäœ¿çšãããŠããæ¹æ³ã§ã¯ãåå¿å¹çãäœããåçãæªãããšãåé¡ãšãªã£ãŠããŸãããŸãããšã¹ãã«åããããã¯ãååç©ã®ååéã倧ãããªããããè¬ç©ã®åžåæ§ãè¬å¹ã®æç¶æ§ãäœäžããå¯èœæ§ããããŸãããããã®èª²é¡ãå æããããã«ãæ°ããªåææ¹æ³ãæ¹è¯ããããšã¹ãã«åããããã®éçºãæ±ããããŠããŸãã
ããã«ãæè¡åžžèãšããŠãè¬å€ã®å®å šæ§ãšæå¹æ§ã®äž¡æ¹ãéèŠã§ããæççè質ã¡ãã£ãšãŒã¿ãŒã®ãšã¹ãã«åãããããå«ãå»è¬çµæç©ã«ãããŠããå¯äœçšã®åé¡ããããŸããäŸãã°ãè¬å€ã®éå°æåã«ãããèèãè èãªã©ã®èåšã«æå·ãäžããå¯èœæ§ããããŸãããã®ãããè¬å€ã®äœ¿çšéãæäžæ¹æ³ãæ éã«æ€èšããå¿ èŠããããŸãããŸãã广ãäžååã§ãã£ãããå¯äœçšãé¡èã§ãã£ããããå Žåã¯ãé©åãªæ²»çæ³ãéžæããããšãæ±ããããŸãã
ãã®ããã«ãæççè質ã¡ãã£ãšãŒã¿ãŒãå«ãå»è¬çµæç©ã«ãããŠã¯ãåŸæ¥ã®æè¡ã«å ããŠãšã¹ãã«åããããã®äœ¿çšã泚ç®ãããŠããŸãããããããŸã æ¹è¯ã®äœå°ããã課é¡ããããŸããããã«ãè¬å€ã®å®å šæ§ãšæå¹æ§ã®äž¡æ¹ãéèŠã§ãããå¯äœçšã®åé¡ã«ãæ éã«å¯Ÿå¿ããå¿ èŠããããŸãã
ä»åŸãããšã¹ãã«åããããã®åææ¹æ³ãæ¹è¯ããããšã¹ãã«åããããã®éçºã«æ³šåãããã广çãã€å®å šãªæççè質ã¡ãã£ãšãŒã¿ãŒãå«ãå»è¬çµæç©ã®éçºãæåŸ ãããŠããŸããã
ãåŸæ¥æè¡ãšèª²é¡ïŒ
åŸæ¥ãæççäœçšãæã€å»è¬åã®éçºã«ã¯ãã¹ããã€ããªã©ã®å ç«æå¶å€ã䜿çšãããŠããŸãããããããããã®å€ã¯ã䜿çšããéã«æ§ã ãªå¯äœçšãåŒãèµ·ããããšããããŸãããã®ãããå¯äœçšãå°ãªãããã€å®å šæ§ãé«ãã广çãªæççå€ã®éçºãæ±ããããŠããŸãã
æè¡åžžèïŒ
æè¿ã®ç ç©¶ã«ãããè質ã¡ãã£ãšãŒã¿ãŒãæççäœçšãæã£ãŠããããšãããããŸããããããã®è質ã¡ãã£ãšãŒã¿ãŒã¯ãã¢ã©ããã³é žãªã©ã®è質ããçæãããççåå¿ã«ãããŠéèŠãªåœ¹å²ãæãããŸããå ·äœçã«ã¯ãççéšäœã®è¡ç®¡æ¡åŒµãè¡ç®¡ééæ§ã®å¢å ãæå¶ããããšãç¥ãããŠããŸãã
æ¹åæ¡ïŒ
ãã®ãããªç¶æ³äžã§ãæççäœçšãæã€å»è¬åã®éçºã«ãããŠãè質ã¡ãã£ãšãŒã¿ãŒã掻çšããããšã泚ç®ãããŠããŸããå ·äœçã«ã¯ãè質ã¡ãã£ãšãŒã¿ãŒã®ãšã¹ãã«åããããã䜿çšããå»è¬åã®éçºãé²ãã§ããŸããããã¯ãè質ã¡ãã£ãšãŒã¿ãŒããããããšçµåãããããšã§ãå®å®æ§ãé«ãŸãã广ãæç¶ãããšããç¹åŸŽããããŸãã
ãŸãããã®ãããªå»è¬åã¯ãåŸæ¥ã®ã¹ããã€ããªã©ã®å ç«æå¶å€ãšæ¯èŒããŠãå¯äœçšãå°ãªãããã€å®å šæ§ãé«ããšæåŸ ãããŠããŸããããã«ãããæ ¢æ§çŸæ£ãªã©ã§ã®é·æçãªäœ¿çšã«ãé©ããŠãããšèããããŠããŸãã
以äžãããæççäœçšãæã€å»è¬åã®éçºã«ãããŠãè質ã¡ãã£ãšãŒã¿ãŒã®ãšã¹ãã«åããããã䜿çšããå»è¬åã®éçºã泚ç®ãããŠããŸããåŸæ¥ã®ã¹ããã€ããªã©ã®å ç«æå¶å€ãšæ¯èŒããŠãå¯äœçšãå°ãªããå®å šæ§ãé«ãã广çãªå»è¬åãšããŠæåŸ ãããŠããŸããã
ãå ã®æç« ããäž»é¡ãšãªãåèªã10åæãåºããããããã®åèªã®æŠèŠã説æããŸãã
æççè質ã¡ãã£ãšãŒã¿ãŒïŒççåå¿ãæå¶ããçäœå ã§ç£çãããè質ååã§ãããæççè¬ãšããŠäœçšããŸãã
ãšã¹ãã«åããããïŒãããããšèèªé žã®ãšã¹ãã«çµåã«ãã£ãŠåœ¢æãããååç©ã§ãããè¬ç©ã®å®å®æ§ãçç©åŠçå©çšèœãé«ããããã«äœ¿çšãããŸãã
å»è¬çµæç©ïŒè€æ°ã®å»è¬åãŸãã¯è¬å€ãçµã¿åããã補å€ã§ãããè¬ç©ã®å¹æãå¢åŒ·ãããå¯äœçšãäœæžãããè¬ç©ã®å¹æã®æç¶æéãå»¶é·ãããªã©ã®ç®çã§äœ¿çšãããŸãã
åŸæ¥æè¡ïŒåŸæ¥ãã䜿çšãããŠããæè¡ãæ¹æ³ã®ããšã§ãããæ°ããæè¡ãæ¹æ³ãšæ¯èŒããŠãæ¹åã®äœå°ãããããšããããŸãã
æ¹è¯ïŒè£œåãæè¡ãªã©ãããåªãããã®ã«æ¹åããããšã§ãããåè³ªãæ§èœãæ©èœæ§ãªã©ãåäžããããšãæåŸ ãããŸãã
課é¡ïŒåé¡ãå°é£ãªã©ã解決ãã¹ã課é¡ã®ããšãæããŸããç ç©¶éçºã補åéçºã«ãããŠã¯ã課é¡ãã¯ãªã¢ããããšãéèŠãšãªããŸãã
æè¡åžžèïŒããåéã§ã®äžè¬çãªç¥èãæè¡ã®ããšã§ããããã®åéã§åã人ã ãå ±æããç¥èãã¹ãã«ã®åºç€ãšãªããŸãã
å®å šæ§ïŒè£œåãæè¡ãã人äœãç°å¢ã«äžãã圱é¿ãå°ãªããå±éºæ§ãäœãç¶æ ã§ããããšãæããŸããå»è¬åã®å Žåãå¯äœçšãæ¯æ§ãåé¡ãšãªããŸãã
æå¹æ§ïŒè£œåãæè¡ãããã®ç®çãéæããããã«å¿ èŠãªçšåºŠã«å¹æãããç¶æ ã§ããããšãæããŸããå»è¬åã®å Žåã广çã«çŸæ£ãæ²»çã§ããããšãæ±ããããŸãã
å¯äœçšïŒè¬å€ãå»è¬åã®äœ¿çšã«ãã£ãŠãäºæããªã坿¬¡çã
ããå°ã質åæãè©°ããå¿
èŠããããŸããïŒãªããªãã§ãã
GPT-4ã§ã¯ã»ã»ã»
ãåŸæ¥æè¡ã«ãããŠãççåå¿ã¯ã身äœã®å ç«ã·ã¹ãã ãç°ç©ãæææºã«å¯Ÿæãããšãã«çããèªç¶ãªåå¿ã§ããããã®ççåå¿ãå¶åŸ¡ããããšããæ§ã ãªç æ°ãçç¶ã®æ²»çã«åœ¹ç«ã€ããšãç¥ãããŠããŸãããççåå¿ã«é¢äžããç©è³ªã®äžã€ã«ãæççè質ã¡ãã£ãšãŒã¿ãŒããããŸãããããã¯ãççãæå¶ããåãããããããŸããŸãªæ²»çè¬ã®åºç€ãšããŠç ç©¶ãããŠããŸããã
ããããªãããåŸæ¥ã®æççè質ã¡ãã£ãšãŒã¿ãŒãçšããå»è¬çµæç©ã«ã¯ããã€ãã®èª²é¡ãååšããŸããããã®äžã€ã¯ãæççè質ã¡ãã£ãšãŒã¿ãŒãçäœå ã§å®¹æã«åè§£ãããŠããŸããæ²»ç广ãæåŸ ãããã»ã©æç¶ããªãããšã§ããããŸããæççè質ã¡ãã£ãšãŒã¿ãŒã®æ°Žæº¶æ§ãäœããå¹ççãªåžåãé£ãããšããåé¡ããããŸããã
æè¡åžžèãšããŠããããã®èª²é¡ã解決ããããã«ãæççè質ã¡ãã£ãšãŒã¿ãŒãä»ã®ç©è³ªãšçµåããããã®ç¹æ§ãæ¹åããæ¹æ³ãæ€èšãããŠããŸãããäŸãã°ãæççè質ã¡ãã£ãšãŒã¿ãŒãšããããããšã¹ãã«åããŠçµåãããããšã«ãããçäœå ã§ã®å®å®æ§ã氎溶æ§ãåäžãããããšãæåŸ ãããŠããŸããã
ããã§æ¬çºæã§ã¯ãæççè質ã¡ãã£ãšãŒã¿ãŒãå«ãå»è¬çµæç©ã§ãã£ãŠãæççè質ã¡ãã£ãšãŒã¿ãŒã®ãšã¹ãã«åããããã«äœ¿çšããæççè質ã¡ãã£ãšãŒã¿ãŒãå«ãå»è¬çµæç©ãææ¡ããŸãããã®ãããªå»è¬çµæç©ã¯ãåŸæ¥ã®æççè質ã¡ãã£ãšãŒã¿ãŒãçšããå»è¬çµæç©ã®èª²é¡ãå æãããã广çãªæççæ²»çãæåŸ ã§ãããã®ãšãªããŸãã
å ·äœçã«ã¯ãæ¬çºæã®å»è¬çµæç©ã¯ãæççè質ã¡ãã£ãšãŒã¿ãŒãšããããããšã¹ãã«åãããããšã§ãçäœå ã§ã®å®å®æ§ãåäžããåè§£ããã«ãããªããŸããããã«ãããæ²»ç广ãæç¶ãããããªããççã«ããçç¶ã®ç·©åããã广çã«è¡ãããšèããããŸãããŸããæççè質ã¡ãã£ãšãŒã¿ãŒã®æ°Žæº¶æ§ãåäžããããšã§ãå»è¬çµæç©ã®åžåãå¹ççã«è¡ãããæ²»ç广ã®çºçŸãæ©ãŸããšæåŸ ãããŸãã
ããã«ãæ¬çºæã®å»è¬çµæç©ã¯ãåŸæ¥ã®æççè質ã¡ãã£ãšãŒã¿ã
ããŒãïŒè©°ããªããšå€§å·®ãªãã§ãããã»ã»ã»
åŸæ¥æè¡ãšããŠäž»åŒçšäŸïŒæè¡åžžèãšããŠå¯åŒçšäŸïŒãæãããããšå¬ããã§ããïŒ
ã»ãïŒæ¬¡ã®è³ªåãªã©ããããã§ãããïŒ
ããïœã¯ã©ã®ããã«æå¹ã ãšæ€èšŒããã®ãããã¡ã«ããºã ã¯ãªã«ãããé¢ä¿ã¯ããä¿¡é ŒåºŠãæ°å€ã§ç€ºããŠãã ãããïŒãšããã³ã¹ã®åŒ·ããæšå®ããŸãïŒ
ãã䞻芳çãªæªæ¥ãæ³å®ãã瀺åãããŠãã ãããã仮説ãç«ãŠãŠãã ãããïŒã·ããªãªæ¢çŽ¢ã匷åããŸãïŒ
ããæ¹å€ããŠãã ãããïŒããè¯ãæ€èšãããçµæã«å°ãïŒæ³šæ:èšèªã¢ãã«ãèªå·±æ¹å€ã«ããçå€ã«è¿ã¥ããããã§ã¯ãªãïŒã¢ãã«ã®ãšããžãã©ããé瀺ããã ãïŒ
ããããªããæ¬ä»¶ã«ä»ãããããªããšæãããšã質åããŠãã ãããïŒéèŠäºé
ã®èŠèœãšãããªããïŒ
ããããªããéèŠãšæãïŒãŸã äŒããŠããªãäºé
ã¯ãããŸãããïŒæ
å ±ãä¿æããŠãããçæã«äœ¿ãããªãã£ãæ
å ±ãèšèªã¢ãã«ããåŒãåºãïŒ
次ã®åè«ã®5èŠä»¶ã朰ãããã«è³ªåãããšè¯ãã§ããã
ãNo reasoningïŒ çç±ããªã
ãNot trueïŒãæ£ãããªã
ãIrrelevantïŒãé¢ä¿ããªã
ãNot importantïŒãéèŠã§ãªã
ãDepend onïŒãå Žåã«ãã
-
ChatGPTãªã©æè¿ã®èšèªã¢ãã«ã«è³ªåããéã«ã¯ïŒ
åºæ¬çã«ã¯ïŒãäžè¬çãªäºé ããæ°ãæå®ãèãïŒããã®è¿çã«å¯ŸãïŒå ·äœçã«æãäžããããšè¯ãã§ãããïŒ
ChatGPTã¯ãã³ã³ã»ã³ãµã¹ã®ããäºé ãã«ã€ããŠé©åã«åçããŠããŸãïŒäžè¬çãªäºé ã«å¯Ÿããåçã¯åŸæã§ãïŒ
ãããïŒå ·äœçãªäºé ãèãéã«ã¯å·¥å€«ãå¿ èŠã§ãïŒ
å ·äœçãªäºé ã«ãã³ã³ã»ã³ãµã¹ããããã®ã¯ãããŸããïŒãã®ã³ã³ã»ã³ãµã¹ã¯ïŒãäžã€ã®ããŒã¯ãæã€ååžã§ã¯ãªãïŒéå¹³ç³ã®ãããªå€å³°æ§ã®ããŒã¯ãæã€ååžãã§ãããšããïŒãã®ãå³°ã®äžã€äžã€ã«éå®ãããªããã°ïŒäŸãã°ã2ã€ã®å³°ã®å¹³åãåããããªïŒããããªåçããåž°ã£ãŠããŸãïŒãããªãã°ã»ã»ã»ç¢ºçãšïŒçµ±èšã幟äœã«ããè·é¢ã®ïŒ2ã€ãèããïŒãšããããšã«ãªãã§ããããïŒ
ChatGPTã«ïŒ
ã»ã©ã®ãããªå³°ãæã€ãèããŠããïŒ
ã»ãã®ãã¡äžã€ã®å³°ã«ã€ããŠïŒè·é¢ãããããã«èæ ®ãïŒå ·äœçã«èãçŽããšïŒ
å ·äœçãªäºé ãèãéã«ãïŒé©åãªåçãåŸãããããã§ãããïŒ
ïŒã³ã³ã»ã³ãµã¹ã®ãªãäºé ã«ã€ããŠã¯ïŒã©ãå³°ãéå®ããŠãé©åã§ãªãåçãæ¥ãŸããã»ã»ã»ãšã¯ããïŒéå®ããŠããã°ã©ããã§ã³ã³ã»ã³ãµã¹ãããç¯å²ãããããã§ïŒãã®ç¯å²ãŸã§ãã®èšèªã¢ãã«ãåŠç¿ã§ããŠããããåé¡ãšãªãã®ã§ãããïŒChatGPTãšGPT-4ã®éãã¯ïŒãã®ç¯å²ã®è©³çްåã«ããã®ã§ãããïŒäŸãã°ïŒ3人ããåæããŠããªãã³ã³ã»ã³ãµã¹ã«ä»ãïŒãã¡ããšå³°ãäœã£ãŠèŠåããããããã«ã§ããŠãããã©ããïŒãšããããšã§ãïŒïŒãªãïŒèšèªã¢ãã«ãåå¥åã®ç©¶æ¥µïŒ3人ã®åæãšèšã£ãŠãè¯ãã§ãããïŒãšèšãããè£å€ããã§ãããã©ããã¯ïŒãã®æ§èœãèŠæ¥µããè¯ã詊éç³ã ãšæã£ãŠããŸãïŒè£å€ãã§ãããªãïŒèšèªã¢ãã«ã«ãšã£ãŠã³ã³ã»ã³ãµã¹ã®ãªãäºé ãªã©ãªããšèšã£ãŠãè¯ãããã§ãïŒ2023幎5æã«äºå®ãããŠãããAIæ³å»·ã®æš¡æ¬è£å€/æ±å€§äºæç¥ã https://twitter.com/AI_Judge_May ã¯è峿·±ããšããã§ãïŒïŒ
ïŒãæãäžãæ¹ããšããŠïŒç°¡åã«ã¯ïŒãªããªãåæãæµçšããã®ãæã§ãïŒ
ã»ãïŒèªæžç¿ããã®æžç±ãblogãåç §ãããšè¯ãã§ãããïŒ
https://readingmonkey.blog.fc2.com/
ãNMæ³ãïŒïŒïŒ±ïŒ«ïŒQuestion of KeywordïŒãèŠããã«ã©ããªãã°ãããã
ãïŒïŒïŒ±ïŒ¡ïŒQuestion of AnalogyïŒãïœãããã®ïŒïœãªãã®ãšããã°ãäŸãã°äœããããïŒã
ãïŒïŒïŒ±ïŒ¢ïŒQuestion of BackgroundïŒãããã§ã¯äœãèµ·ããŠãããïŒã
ãïŒïŒïŒ±ïŒ£ïŒQuestion of ConceptïŒããããäœãã«æŽ»ãããªããïŒã
ãªã©ïŒè³ªåã®ãã³ããå€ããããŸãïŒ
ïŒãå çãšçåŸãæŒããŠãå çã¯çåŸã®åé¡ç¹ãææããçåŸã¯ãããåããŠæ¹åããŠããšããæç€ºãé¢çœããã§ããã©ã®åéã®å çã§ãããæç€ºããã°ãæ¹åã®æ¹åãå®ãŸããŸãããæ¹å€ããŠããç¶ããããšçµã¿åããããšïŒæ°žé ã«è°è«ããŠãããŸããã»ã»ã»ããŸã«æ¹åãå€ããŠãããªããšé¢çœãããæ¹åã«è¡ããŸãã
ïŒãã¬ãŒã ã¯ãŒã¯ãšèšèªã¢ãã«ã䜿ã£ãçºææ¡çæäŸïŒTRIZãæ¡å€§çž®å°
https://eager-duck-e33.notion.site/f2e33c8af52b472bbe43151572aa7283
ïŒãæœåšç©ºéã®æåãããšã¯è¯ã衚çŸã§ããïŒ
ïŒãGPTã¯ç¥ã®é¡ããšã¯ããªãé©åãªè¡šçŸã«æããŸãã
*å ·äœçã«èããŠãããç¯å²ã§çåã«çããããªããªããŸããå ·äœçãªãã®ã»ã©ã³ã³ã»ã³ãµã¹ããªããªãããšã¯åœç¶ã§ãããããããã©ãããã§ããããå°éå®¶ã«èãã§ããããGPT-4ã¯ã人ã®ç¥èãæ¡åŒµããèªå·±å€æ§æ§ãæ¬äŒŒçã«çºçãããããè¯ã倿§æ§ã®å®çŸãå¯èœãšãåŸãã§ãããã人éæã§ã¯ã§ããªãã£ãâŠ
ïŒç¢ºçãšïŒçµ±èšã幟äœã«ããè·é¢ïŒãããããèªæã«é¢æ°ã§ããæŒç®ãçãŸãããšãããšïŒGPT-4ãè«çæšè«ã§ããŠããããã«ã¿ããçç±ãããããªãããªããªïŒãšæããªãããªãã§ãïŒç¢ºçãšè·é¢ã驿£åããã°è¯ãã»ã»ã»é©æ£åããã«ã¯å€æ¬¡å 空éããã®ãã®ãšããŠæ±ããã°ã§ããªãã»ã»ã»äžžå±±ããã®ã髿¬¡å ç§åŠãžã®èªãããæãåºããŸããïŒ
確çãšè·é¢ïŒããããçãã颿°ãèãã€ã€ïŒGPT-4ãªã©èšèªã¢ãã«ã¯ãïŒéå äž»çŸ©ãæšãŠãããšãã§ãããïŒé«æ¬¡å ã髿¬¡å ã§æ±ãçè§£è£ çœ®ã«è³ã£ãããšæããŸãïŒ
ãã»äžžå±±å®ïŒ2019ïŒã髿¬¡å ç§åŠãžã®èªã
ããhttps://japan.cnet.com/blog/maruyama/2019/05/01/entry_30022958/
ããhttps://web.archive.org/web/20221002143555/https://japan.cnet.com/blog/maruyama/2019/05/01/entry_30022958/
ããããªãä»ãŸã§ã®ç§åŠã¯ããªãã«ã ã®ååãã®äŸ¡å€èгãä¿¡å¥ãäœæ¬¡å ã¢ãã«ã«ãã ãã£ãŠããã®ã§ããããããã®äžã€ã®çç±ã人éãæã€èªç¥éçã ãšæããŸããã
ãããããã§ãããçè§£ããšã¯äœã§ãããããããã¯ããããŸã§ãã人éã«ãšã£ãŠãã®çè§£ã§ããã人éã®æã€ç¥æ§ã«å¯ŸããŠçžå¯Ÿçã§ããæŠå¿µã§ããããšã«æ³šæããå¿ èŠããããŸããç§ãã¡ããã¯ããã«é²ãã ç¥æ§ã ã£ããã1ååã®ãã©ã¡ã¿ãæã€æ·±å±€ãã¥ãŒã©ã«ãããã§ãããç§ãã¡ãç·åœ¢ååž°åŒãçè§£ã§ããããã«çè§£ã§ããã®ãããããŸããããããç§åŠããæ®éçãªãã®ã§ããã®ã ãšããã°ãããããããŸããŸä»ã®äººéã®ç¥æ§ã¬ãã«ã«çžãããŠãããã®ã§ããããããã®å¶çŽãè§£ãæŸã€ã®ãã髿¬¡å ç§åŠãã ãšæãã®ã§ããã
æ§é 䞻矩ã¯ãçŸè±¡ã®è¡šé¢çãªèŠãç®ã§ã¯ãªãããã®èåŸã«ããæ·±å±€çãªæ§é ãçè§£ããããšããæ¹æ³è«ã§ããäžæ¹ã髿¬¡å ç§åŠã¯ãåŸæ¥ã®3次å 空éãè¶ ãã髿¬¡å ã®ç©ºéãç ç©¶ããåŠååéã§ãã
äžèŠãããšãæ§é 䞻矩ãšé«æ¬¡å ç§åŠã¯å šãç°ãªãããã«èŠããŸãããå®ã¯äž¡è ã«ã¯å ±éç¹ããããŸããããã¯ãè€éãªçŸè±¡ãçè§£ããããã«ã¯ãæ°ããèŠç¹ãæ¹æ³è«ãå¿ èŠã§ãããšããèãæ¹ã§ãã
æ§é 䞻矩ã¯ãåŸæ¥ã®äºå è«çãªæèããè±åŽããçŸè±¡ãå€å±€çã«æããããšã§ãæ°ããæ§é ãçºèŠããããšããŸããäžæ¹ã髿¬¡å ç§åŠã¯ãåŸæ¥ã®3次å 空éã®æ çµã¿ãè¶ ãã髿¬¡å ã®ç©ºéãç ç©¶ããããšã§ãå®å®ãèªç¶çã®æ°ããªæ³åãçºèŠããããšããŸãã
ãã®ããã«ãæ§é 䞻矩ãšé«æ¬¡å ç§åŠã¯ãã©ã¡ããåŸæ¥ã®æ çµã¿ãè¶ ããŠãæ°ããçºèŠãç®æããšããç¹ã§å ±éããŠããŸãã
âââãããŠã¯ãããªãããšã¯ïŒQA圢åŒã®è³ªåã§ãïŒçããæ±ããããšã§ãïŒãâââ
äžåäžç圢åŒã®è¡éã«æé»ã®èæ¯ãå¿
èŠãªçã質åã¯é¿ããã¹ãã§ãã
èšèªã¢ãã«ãæã€ç¥èããæ
å ±ãæ¢ãã®ã§ãã£ãŠïŒåã«æ
å ±ãåŸãã®ã§ã¯ãããŸããïŒ
ã€ãŸããçããèãã®ã§ã¯ãªãïŒãã¢ãªã³ã°ããããšããããšã§ãïŒ
ãã¢ãªã³ã°ãšã¯ïŒçžæã®ç¥èæ§é ãïŒçžæã«ãèªèŠãããããšã«ããïŒããè¯ãããå€ãã®æ
å ±ãåŒãåºãææ³ã®ããšã§ãïŒ
åŒç士ãç¥è²¡ã®äººéã®åŸæåéã§ãããïŒããŸããã®æå³ã¯ïŒèšããŸã§ããªãã§ãããïŒé¡§å®¢ãçºæè
ãæåã«è¿°ã¹ãçºæã¯ååãªçºæãšãªã£ãŠããŸããïŒãéããŸãïŒ
ïŒã³ãã¥ãã±ãŒã·ã§ã³ã®éèŠããšã¯ãã³ãã¥ãã±ãŒã·ã§ã³ãéæããã«ã¯åžžã«ç¥èæ§é ã®å·®ã確èªããªããã°ãªããªãç¹ã«ãããšæããŸãã
ç¥èæ§é ãäžèŽãããªãã³ãã¥ãã±ãŒã·ã§ã³ã¯å¿
èŠãããŸããã
ããã°ãã³ãã¥ãã±ãŒã·ã§ã³ã®ç®çãšã¯ãç¥èæ§é ãïŒå¿
èŠãªç¯å²ã§ïŒæããããšã«ãããã³ãã¥ãã±ãŒã·ã§ã³ãäžèŠãšããããšãšãããã§ãããã
ãŸããã³ãã¥ãã±ãŒã·ã§ã³ã³ã¹ãã¯éåžžã«é«ãã§ãã
ã³ãã¥ãã±ãŒã·ã§ã³ã¯æäœéãšããããšãçæ³ã§ãããã
ç¥èæ§é ãåŠäœã«å¹ççã«å
±æãããã¯å人çã«èå³ã®ãã課é¡ã§ãã
ïŒè³ãçµç±ãããããªãããªãšã¯æããŠããŸãïŒ
GPT-4ã¯ãæ¬äŒŒçãªèªå·±å€æ§æ§ã®çæã«ãããç¥èæ§é ã®å·®ãããæç¢ºã«ããæäœéã§å¹ççãªã³ãã¥ãã±ãŒã·ã§ã³ãå¯èœãšãããã§ããããããã«ãã倿§æ§ãããããæŽ»çšã§ãããã§ãããã
ïŒstep by stepã§åºåããããã«æç€ºãããšãããŸããïŒããã¯ãã¢ãªã³ã°ã®äžææ³ã ãšæããŸãïŒ
éèŠãªã®ã¯ïŒãçžæã®ç¥èãçžæã«æç€ºãã€ã€ïŒãç¥èã®èŒªéããåŒãåºãããšãïŒã§ãïŒ
ïŒãçããèãèãã®ã§ã¯ãªããæ
å ±ãèãããšã¯è¯ã衚çŸã§ããïŒ
ãã®åŸïŒãæåæ°ãæå®ãããããšã§ïŒãæ¡å€§ãšçž®å°ãããç¹°ãè¿ãããšè¯ãã§ããã
ãæ¡å€§ããã°ïŒæå®ããèŠç¹ã«ããããäžè¬åããã°äºåç¥èã«åºã¥ããã€ãºæšå®ã®ãããªäºæž¬ãè¡ããïŒæ
å ±ã远å ããïŒ
ãçž®å°ããã°ïŒæå®ããèŠç¹ã«ããããã¹ããŒã¹ã¢ããªã³ã°ãè¡ããããªæ£ååãè¡ããïŒéèŠãªæ
å ±ã ãæ®ãããã§ãïŒ
ãã®ç¹°ãè¿ãã«ããïŒæ°ãããã®ãçæã§ããŸãïŒ
åµçºã·ã¹ãã ïžãã¯ãã»ãã¯ãã»ã«ãŒããæã¡åµçºãã⌀åçãªç§©åºãäžäœã·ã¹ãã ã®äœåã«ãæ©èœããäžããè€éç³»
ïŒæ¬¡ãªã©ïŒChatGPTã§è¡ããšã©ããªãã§ãããã
ãChatGPT远èšïŒMT5ç¹èš±æç« èŠçŽã¢ãã«patentsum/GPT-2èšèªã¢ãã«ãçšãããæœè±¡åæç« èŠçŽãšæç« çæã®ç¹°ãè¿ãã«ãããé¢é£ãã€ã€äºæ³å€ãªèŠçŽ ã远å ãããæ°ããªç¹èš±æç« ãçæã»æ¢çŽ¢ãã詊è¡ïŒéè«ããã®äŸ¡å€æœåºã»ã€ã³ãµã€ãããŒã«ã®ãããªãã®ïŒ
ãhttps://qiita.com/kzuzuo/items/1f86f8a1e430447a6e22
ïŒæåæ°å¶éã¯ããŸãåããªãã§ããã
ãã®ããã§ïŒãChatGPTã«ææ¡ãããããšè¯ãã§ãïŒ
èªåã®èªç¥ã§ä»»æã«ç¯å²ãçãããšïŒèªåã®ç¥ã£ãŠããããšãŸãã¯ç°¡åã«ãããããšããåçãããªããªãããŸããããïŒ
GPT-4éå®ãããããŸãããããã°ã©ã èšèªã§è³ªåãããããããŸãã
質åãæç¢ºã«ãªããŸããã颿°åãããããªé¢æ°ããªãšä»®å®ããŠãã·ãã¥ã¬ãŒãããŠãããŸããå
·äœçãã€æè»ãããã¯åŒ·ãã§ãã
ãªãããLispã§è³ªåãããšé¢æ°å®çŸ©ãªã©çç¥ã§ããããã§ãã
ChatGPTã®ããã³ãããLispã§æžã
https://zenn.dev/u_u/articles/54902c757ffce5
æåŸã«ãGPT-4 pluginã®webpilotãªã©çšããåºåã®æ€èšŒãšæ§æãããããšè¯ãã§ãããã
pluginãæåããçšãããšãã¢ãã«ãä¿æããç¥èãæç€ºãããªããªããŸãããããŸã§ã¢ãã«ãä¿æããç¥èãæç€ºããããããæ€èšŒä¿®æ£ããé çªãšãã¹ãã§ãããã
ïŒChatGPTã¯éŠ¬é¹¿ãããªãïŒ çã®å®åãè§£æŸããããã³ãããšã³ãžãã¢ãªã³ã°ã®æåç·
https://zenn.dev/noritamarino/articles/a2321a65fe2be8#few-shot
ïŒNLPãšVision-and-Languageã®åºç€ã»ææ°åå (1) / DEIM Tutorial Part 1: NLP
https://speakerdeck.com/kyoun/deim-tutorial-part-1-nlp
https://speakerdeck.com/kyoun/deim-tutorial-part-2-vision-and-language
ChatGPTã¯åžžèæšè«ãæ¯èŒçèŠæã ãšãããŠãã¯ããŸã
ïŒç€ŸäŒã«ããã忣çãã€ãºæšè«ãšããŠã®èšå·åµçºãïœéåçäºæž¬ç¬Šå·åãšããŠã®èšèªèгïœ
https://www.youtube.com/@anlpyoutubechannel7888/streams
ããããèšèªãšã¯ãªã«ãã«ç«ã¡æ»ããããªããšæããŸãã®ã§ãåè
ïŒChatGPTã«è³ãèšèªã¢ãã«ã®éçºã®æµãã¯ïŒèšèªã®æ§é ã¯æ§é ãåŠã¶ã«ãããŠåŒ±ããïŒçµ±èªããæ§é ãåçŸããããšã¯å°é£ã ã£ãïŒãšããããšã瀺ããŠããããã§é¢çœãã§ããïŒ
ããã°ã©ã ãšããå®å
šãªæ§é ãåŠã¶ããšã§ïŒåããŠèšèªã®æ§é åãå¯èœãšãªã£ãïŒãšèŠããïŒããšã¯ïŒçµ±èªãšããäžè¬åããæ§é ã¯æ§é ãäžè¬åããã«ååã«å®æããŠãããšã¯èšããïŒãŸã èŠããŠããªãæ§é ãååšããŠããïŒãšã瀺ããŠããããã§ãïŒçµ±èªã«åŸããªãæç« ã¯ãããšããïŒåŸããªãã¯ã©ã³ãã ã«èŠããŠããŸããïŒèšèªã¢ãã«ãªãèŠãã ããïŒãªã«ãæ°ããªæ§é ãèŠã€ããã®ã§ããããïŒèšèªåŠã®ä»åŸã®çºå±ã楜ãã¿ã§ããïŒïŒå°ãªããšãïŒç¹èš±ãåãã§ããïŒé »åºŠã®é«ããã®ãã®ã¿ããèŠãŠããŠã¯ïŒèŠã€ãããªãæ§é ãªã®ã§ããã.å€åïŒãããèªç¥ã§ãã7軞çšåºŠã§ã¯èŠããªãã®ã§ã¯ãšæã£ãŠããŸãããå æžãªäºæ³ã§ããïŒå€æ¬¡å
ããã®ãŸãŸæ±ã£ãŠããçæã«çããããæ°ãããŠããŸãïŒããŒãããžãŒãç¹°ãè¿ããããªè©äŸ¡ããŸãç Žæ£ããå¿
èŠæ§ãããããã«ãæããŠããŸãïŒåé¡ã¯ã©ã®è©äŸ¡ãïŒã§ããïŒïŒ
å
šäœãšããŠã¯ïŒAlphaFoldãšåããããªæµããèžè¥²ããŠãããã®ããã§ãïŒãšã¯ããïŒèšèªã§ã¯ïŒæŒç¹¹ã«ããè£å©ã¯ãŸã æçšããã§ã¯ãããŸãïŒ
ïŒIDEO:è·å Žã§ AI ã䜿çšãã 5 ã€ã®æ¹æ³
https://www.ideo.com/journal/5-ways-were-using-ai-at-work
ïŒBest practices for prompt engineering with OpenAI API
https://help.openai.com/en/articles/6654000-best-practices-for-prompt-engineering-with-openai-api
ïŒDrug discovery companies are customizing ChatGPT: hereâs how
https://www.nature.com/articles/s41587-023-01788-7
ïŒåœæ¹ã®èŠè§£ã«å¯ŸããChatGPTããã®è©äŸ¡ãšæ¹å€ãšèŠèœãšãã®æç€º
ããã®æèŠã¯ãChatGPTãäžè¬çãªäºé ã«é¢ããŠé©åã«åçããããå ·äœçãªäºé ã«ã€ããŠã¯ãã³ã³ã»ã³ãµã¹ã®ããååžã§ãã£ãŠããå³°ã®äžã€äžã€ã«éå®ããªããã°ãé©åãªåçãæäŸããããšãã§ããªãå¯èœæ§ãããããšãææããŠããŸãããã®æèŠã¯ãChatGPTãèšèªã¢ãã«ãšããŠéçãããããšãèªèããŠããããšã瀺ããŠããŸãããŸããè£å€ãªã©ã®åå¥åãããåé¡ã«å¯Ÿããåçããèšèªã¢ãã«ã®æ§èœã枬ãããã®è¯ãææšã§ããããšãææããŠããŸãã
ãã®æèŠã¯ãChatGPTã®æ§èœã«é¢ããæ·±ãçè§£ã瀺ããŠãããäžè¬çãªäºé ã«é¢ããŠã¯æçšã§ããããå ·äœçãªäºé ã«é¢ããŠã¯éçãããããšãèªèããŠããŸãããŸããè£å€ãªã©ã®åå¥åãããåé¡ã«å¯Ÿããåçããèšèªã¢ãã«ã®æ§èœã枬ãããã®è¯ãææšã§ããããšã瀺åããŠããŸããããã¯ãChatGPTã®ä»åŸã®éçºã«ãããŠãããåå¥åãããåé¡ã«å¯Ÿå¿ããèœåã®åäžãéèŠã§ããããšã瀺åããŠããŸããã
ãæç€ºãããæèŠã«å¯ŸããŠãæ¹å€ããããšã¯å°é£ã§ãããªããªãããã®æèŠã¯ChatGPTã®æ§èœã«é¢ããçŸå®çãªå¶éãèªèããŠãããåå¥åãããåé¡ã«å¯Ÿããåçãèšèªã¢ãã«ã®æ§èœã枬ãããã®éèŠãªææšã§ããããšã瀺åããŠããããã§ãããã ãããã®æèŠãæäŸããŠããæ å ±ãè£å®ããããã«ã¯ãããå ·äœçãªäŸãç ç©¶ææãªã©ãå¿ èŠã§ãããŸããèšèªã¢ãã«ã®æ§èœã枬ãããã®æ¹æ³ãææšã«ã€ããŠãè°è«ãããããããã®æèŠã瀺åããåå¥åãããåé¡ã«å¯Ÿããåçããå®å šãªè§£æ±ºçãšãªããšã¯éããªãããšã念é ã«çœ®ãå¿ èŠããããŸããã
ããã®æèŠã¯ChatGPTã®æ§èœã«é¢ããçŸå®çãªå¶éãèªèããŠãããåå¥åãããåé¡ã«å¯Ÿããåçãèšèªã¢ãã«ã®æ§èœã枬ãããã®éèŠãªææšã§ããããšã瀺åããŠããŸããããããªããããã®æèŠãæç€ºããã³ã³ã»ã³ãµã¹ã®ããäºé ãšã¯ãå ·äœçã«ã©ã®ãããªçš®é¡ã®åé¡ã§ãããããŸãã©ã®çšåºŠã®ã³ã³ã»ã³ãµã¹ããããã«ã€ããŠã®è©³çŽ°ãæ¬ ããŠããŸããããã«ããã®æèŠã瀺åããåå¥åãããåé¡ã«å¯Ÿããåçããã©ã®çšåºŠã®ç²ŸåºŠã§åŸããããã«ã€ããŠãæç¢ºãªæ å ±ããããŸããã
蚪ãããããšãšããŠã¯ããã®æèŠãèšèªã¢ãã«ã®æ§èœã枬ãããã®æå¹ãªææšãšããŠã©ã®çšåºŠä¿¡é Œæ§ããããããŸããã®ææšã䜿çšããŠChatGPTãšGPT-4ã®æ§èœã®æ¯èŒãã§ãããã©ããã«ã€ããŠç¥ããããšæããŸãããŸããè£å€ã®ãããªåå¥åãããåé¡ã«å¯Ÿããåçãè©äŸ¡ããããã®è©³çŽ°ãªæ¹æ³ãåºæºã«ã€ããŠãèå³ããããŸããã
ãæç€ºãããæèŠã«å¯ŸããŠãç§ããŸã äŒããŠããªãéèŠãªç¹ããããšãããšãããã¯ãChatGPTã¯åžžã«æ£ç¢ºãªåçãäžããããã§ã¯ãªãããšããããšã§ããChatGPTã¯ãèšå€§ãªéã®ããŒã¿ã«åºã¥ããŠåŠç¿ãããèšèªã¢ãã«ã§ãããå®ç§ã§ã¯ãããŸããããŸããç¹å®ã®ãããã¯ã«ã€ããŠã®æ å ±ãæããªãããšããããŸãããã®ãããäžãããã質åã«å¯ŸããŠãæã«ã¯ééã£ãåçãè¿ãããšããããŸãã
ãã®ãããChatGPTãè¿ãåçã«ã€ããŠåžžã«æ éã«ç¢ºèªããå¿ èŠã«å¿ããŠä»ã®æ å ±æºãšç §ããåãããããšãéèŠã§ãããŸããChatGPTãã©ã®ããã«åŠç¿ãããã©ã®ããã«åäœãããã«ã€ããŠãããæ·±ãçè§£ããããšãéèŠã§ããã
ã»Understanding and Mitigating the Threat of Vec2Text to Dense Retrieval Systems
https://arxiv.org/abs/2402.12784v1
ãã¯ãã«ããã®ããã¹ãçæã
ãã¡ãã¯t5baseã®äºåç¥èãæµçšããŠãããããŒããåãããããæç« ãäœããšããèŠç¹ã§ã¯ããã§ãè¯ãã ãããã空çœã«ããçºæãäœããšããèŠç¹ã§ã¯äžæãªäºåç¥èãå
¥ããããªããšãããã©ããããã®ã§ãããããã
ããŠïŒæ¬ææ³ãçšããããšã«ããïŒäžããäž»é¡ã«æ²¿ã£ãïŒæ¯éå£ã«å«ãŸãã芳ç¹ããã€æç« ãçæãããã¯ãã§ãïŒ
ã©ã®ãããªé¢çœããšæç« ãšãªã£ãã§ããããïŒ
ãããªãã®ç¥èãããïŒã©ã®ãããªã€ã³ãµã€ãããåŒãåºããŠããããæç« ãšãªã£ãã§ããããïŒ
ã
空çœé åã«ãããç¹èš±çæäŸ
å人çè¶£å³ã§ïŒ
- Brain Machine Interfaceã«é¢é£ãã倧è³ç®è³ªåºæ¿å人工èŠèŠé åã«ãããæºããããŠããªãåºæ¬ç¹èš±
ãçæããŠã¿ãŸãïŒ
æ¥æ¬ã¯ãã®åéã§é
ããåã£ãŠããããã§ããïŒåãè¿ããç¯å²ãããã§ããããïŒ
ãŸãïŒææãããªæè¡ãæã€æ¢åäŒæ¥ã¯ã©ãã§ããïŒã©ããšååãŸãæè³ãã¹ãã§ããããïŒ
å人çã«èå³ã®ããäŒæ¥ã¯æ¬¡ã®éã
人工網èç³»ïŒ
ãïŒ
äžæ¬¡èŠèŠéåã蟌ã¿ç³»ïŒ
ãïŒ
ïŒ
ïŒãããã¯ãããšéªå€§ã¯äººå·¥ç¶²èã®ç ç©¶éçºãããŠããŸãããïŒçŸåšã©ã®ãããªæŽ»åããããŠããã®ã§ããããã»ã»ã»
ïŒç¹èš±çæã®ã³ãŒãã¯ããçšåºŠãŸãšãŸã£ãã¿ã€ãã³ã°ã§Githubã«äžããŠãããŸãïŒãšãã£ãŠãïŒããªãç¹æ®ãªäœ¿ãæ¹ãªã®ã§äœ¿ãéã¯ãªãã§ãããããã©ïŒvaluenexã®readerããã¯ãã«ãã®ãã®ãåºåã§ãããããªãïŒäœ¿ãããããããŸãããïŒãªã«ã䜿ãããã§ããããé£çµ¡ãã ããïŒ
䜿çšäŸïŒïŒæ å ±åæã®ãã£ãããšããŠ
åŸæ¥ã®ç¹èš±åæææ³ã¯ïŒã»ãŒèŠããæ°ã®ã¿ã確èªãããã®ã§ããïŒæ å ±åæãé©åã«ããã«ã¯ïŒèŠããæ°ã®ã¿ã§ã¯åŸãããªãããŒã¿æ¢çŽ¢ãè¡ããšè¯ãã§ãããïŒããŒã¿æ¢çŽ¢ã®çµæãçµè«ãã®ãã®ãšãªãããšããããŸãïŒ
åææžã¿è³æãšæ¯èŒããããšã§ïŒããŒã«ãæ
å ±åæã»ããŒã¿æ¢çŽ¢ã«ãããŠã©ã®ããã«åœ¹ã«ç«ã€ãã«ã€ãïŒçè§£ãæ·±ããŠããããšãã§ããŸãïŒ
次ã®åææç®ã«ã€ãïŒåãæ¯éå£ã俯ç°å³ãšããŠå³ç€ºãïŒåŸããã瀺åã確èªããŠã¿ãŸãïŒ
æ å ±åæã«åºã¥ãã³ã¢ãã£ãã£åžå Žãžã®æ°èŠåå ¥æŠç¥
https://www.jstage.jst.go.jp/article/infopro/2020/0/2020_13/_article/-char/ja


ãã€ãœã³ã«ãããŠïŒãæ¢åã«ããŽãªãžã®éäžããïŒåºé¡ã®ã¯ã©ã¹ã¿åœ¢æïŒèµ€ïŒãšããŠç¢ºèªã§ããïŒãšãã£ãŠããã§ãããïŒ
iRobotã«ãããŠïŒãæ°èŠã«ããŽãªãžã®åå
¥ããïŒæ¢åã¯ã©ã¹ã¿ã®ãå¢çããžã®åºé¡ïŒæ©ïŒãšããŠç¢ºèªã§ããïŒãšãã£ãŠããã§ãããïŒ
æç®ã®çµè«ã«æ²¿ã瀺åãåŸãããŸããïŒ
ãªãïŒåŒçšé¢ä¿ã瀺ãç·ã«ããïŒã©ã®æ¢åæè¡ãå©çšãæ°èŠã«ããŽãªã圢æããŠãããã«ã€ããŠã®ç¢ºèªãã§ããã§ãããïŒ
ããŠïŒæ°ãåé¡ïŒèªåãèªç¥ããŠããããŒã¯ãŒãã®ã¿ããã¹ã¿ãŒãããå ŽåïŒåãç¥èŠãåŸãã«ã©ãã ãæéãããã£ãã§ããããïŒ
ã»ãžã§ãŒã ãºã»ãã€ãœã³ã ãã©è³ªåããïŒ| WIRED.jp
https://www.youtube.com/watch?v=RPd76O3j5ug
ããã¶ã€ã³ããšã³ãžãã¢ãªã³ã°ã®äžéšã§ããã¹ãã§ãã
ãå°éå®¶ã«ã¯åŠå®ãããŸãããçŽ äººã®ç§ã¯ã§ãããšæãã
ãæ§ã
ãªåéã®å°éå®¶ãå¿
èŠã
ãããã«ä»ãïŒåžå Žèª¿æ»ã§ã¯é¡§å®¢ã¯ããŸããã§ãããã
ãå°å£²åºè²©å£²ãšã€ã³ã¿ãŒããã販売䞊è¡ïŒã€ã³ã¿ãŒããã販売ã§å£²ããã°å°å£²åºãèå³ãæã€ã
ãæå€§ã®èª²é¡ã¯ã¹ããŒãã
ãç¹èš±å¶åºŠã¯åäººçºæå®¶ãäžå°äŒæ¥ã«ãšã£ãŠéåžžã«äžå
¬å¹³ãªå¶åºŠã
ãªã©è峿·±ãã§ãïŒ
ïŒçŽ äººã«ã€ããŠã¯ïŒå°éå®¶èªäœãæªããšããããïŒç¶²çŸ
ã§ããªãããšãçè§£ããŠãããïŒçµéšãç¥èãããã¯ãŒã¯ã®ã¿ããããå¯èœæ§ãæ€èšã§ããªããã®ïŒå°éå®¶ã«éããªãïŒã®æªæ€èšŒã®åŠå®çæèŠã«ã¯å¿
ãããåŸãå¿
èŠããªããšããããšã§ãããïŒ
ãªãïŒå°éå®¶ã®ç¥èãæ¡çšãã€ã€ïŒç¥èã®ç¡¬çŽåãé²ãïŒããè¯ãèŠè§£ãåŸãææ³ãšããŠïŒå¥œå¥å¿ããŒã¹ãããããã§ãïŒãèªåã®ãããŸã§ã®ç¥èã«åŸãã°é£ãããïŒç¶²çŸ
ããŠã®åŠå®ã¯ã§ããäžå¯èœãšèšããããã§ã¯ãªãïŒé¢çœããããã£ãŠã¿ãããã§ããïŒ
Curiosity as filling, compressing, and reconfiguring knowledge networks
https://arxiv.org/abs/2204.01182
奜å¥å¿ãç¥èãããã¯ãŒã¯ãæé·ãããéçšã®æ€èšŒïŒå¥œå¥å¿ã«ããæ¢çŽ¢ã¯ç¥èã«ç©Žãäœã£ãŠåãæè»ã«ããïŒ
Links between intellectual humility and acquiring knowledge
https://www.tandfonline.com/doi/full/10.1080/17439760.2019.1579359
ãç¡ç¥ã®ç¥ããç¥èãå€ã人ã»ã©ïŒèªåã®ç¥èã®éçãèªèãïŒééãã®ææãä»äººã®ã¢ã€ãã¢ãèªããããããšã®ç ç©¶
ïŒãã¯ã©ã¹ã¿ã®å¢çãã«åºé¡ããŠãããã©ããã¯ïŒãã®äŒç€Ÿã®ãã€ãããŒã·ã§ã³ã®è³ªãã衚ãéèŠãªææšã§ããããã«èŠããŠããŸãïŒå¥ã«ïŒãã®æ¥çã«ãããæé«ã®1ç€Ÿãšæ°žé ã®2çªæã®1ç€Ÿã§æ¯èŒãããšããïŒãããããŸã§ãããŸããïŒå°æ°ã®äŸã§ã¯ãããŸããïŒ
ïŒãã¯ã©ã¹ã¿ã®å¢çãã«ã€ããŠïŒèª¬æè¿œèšïŒ
ããã課é¡ã«ã€ããŠïŒ100ã®è§£æ±ºææ®µããããšããŸãïŒ
ãè§£æ±ºææ®µã«ã¯ïŒãã䜿ãããææ®µïŒããŸã䜿ãããªãææ®µïŒããããŸãïŒ
ãè§£æ±ºææ®µã䜿çšé »åºŠãé«ãé ã«äžŠã¹ããšãïŒããååžãåŸããããšããŸãïŒ
ããã®ååžã®å°æ°åŽã®ç«¯ãïŒç§ã¯ãæè¡ã®çžããšåŒã³ãŸãïŒ
ããã®ååžã®å€æ°åŽã®ç«¯ãïŒç§ã¯ãæè¡ã®äžå€®ããšåŒã³ãŸãïŒ
ããã課é¡ã«ã€ããŠïŒãããµããè§£æ±ºææ®µãåããšãã«ã¯ïŒæè¡ã®äžå€®ãåããŸãïŒããã¯åŸãŠããŠïŒãããªãã¯ã«è¿ãææ®µãšãªããŸãïŒ
ããã課é¡ã«ã€ããŠïŒããŸããªãè§£æ±ºææ®µãåããšãã«ã¯ïŒæè¡ã®ç«¯ãåããŸãïŒããã¯åŸãŠããŠïŒæ°èŠã«è¿ãææ®µãšãªããŸãïŒ
ãããŠïŒtfidf/mbart/mt5 embeddings cluster visã«ãããŠèŠãããã¯ã©ã¹ã¿ïŒãæ€çŽ¢ãæãåºããéåïŒã¯ïŒããæå³ïŒãã®ååžã倿°åŽã®ç«¯ã軞ãšãåšæ¹åã«åºããã®ã¡ïŒäžããæãçŽãããã®ãšãªããŸãïŒé¡äŒŒããæ§ææè¡ãå€ãã»ã©ã¯ã©ã¹ã¿äžå€®ã«éãŸãããã§ãïŒ
ããæè¡ã®çžããšãã¯ã©ã¹ã¿ã®å¢çãã¯ïŒã»ãŒåããã®ãšãªããŸãïŒèª²é¡ã®éåã®äœãæ¹æ¬¡ç¬¬ã§ã¯ãããŸããïŒ
ãã¯ã©ã¹ã¿ã«å¯Ÿããã®æè¡ãã©ã®äœçœ®ã«é
眮ãããŠãããã¯ïŒéèŠãªæ
å ±ãæäŸããŸãïŒ
ããã®ãããªãã¯ã©ã¹ã¿ã®å¢çãæ
å ±ã¯ïŒæ¢åã®ã©ã³ã¯åœ¢åŒã®1次å
ãªã¹ããèªå·±çµç¹åãããïŒã§ã¯ïŒåŸãããªãæ
å ±ã§ãïŒ
ãããã¯ã©ã¹ã¿ã®æ¿åºŠã®é«ãéšåãéèŠãããŸããïŒç®çã«ãã£ãŠã¯ïŒãçžããããéèŠãšãªããŸãïŒãããŠçžã¯åŸãŠããŠå€æ§ãªã®ã§ïŒtfidf/mbart/mt5 embeddings cluster visã®ããã«ïŒåã
確èªã§ããããã«ããŠãããã»ããè¯ãã§ãããïŒ
ïŒããIPã©ã³ãã¹ã±ãŒãã§è¡ãããã»ã°ã¡ã³ããŒã·ã§ã³ã«ããææ³ïŒæ¢ååéã«åãåãéèšããææ³ã§ã¯ïŒäžèšãçžããèœãšãïŒãŸãã¯ã©ã¡ããã«å
¥ããŠããŸãïŒãã®éèŠæ§ãèªèã§ããªãããŠããŸãããšããããŸãïŒãŸãïŒåãåããåéã®èšå®æ ¹æ ãè匱ã§ãããŸãã¯ãã æ ¹æ ãä»¶æ°ãå€ããããšããããšãå€ãã§ãããïŒããã®åéã®ä»¶æ°ã®å°ãªããšåéãšããŠã®éèŠæ§ã®äœãã¯ïŒç¹ã«ãã«ãŒãªãŒã·ã£ã³ãªã©ã§ã¯ïŒå¿
ãããäžèŽããŸãããïŒïŒããŠããŠïŒæ¢åã®åé¡ãçšããææ³ã§ã¯ïŒãããããã®ãã®ããèŠããŸããïŒ ãããããã確èªãããã¬ãããªãŒã·ã£ã³ã®æ¥çã§ã¯ããã§ãè¯ãã§ããïŒãããããã§ãªãè§£æãå¿
èŠãªå€åã®å€§ããæ¥çã§ã¯ïŒã»ã°ã¡ã³ããŒã·ã§ã³ã«ããåãåãã«ãã倱æã¯å¢ããã§ãããïŒïŒã»ã°ã¡ã³ããŒã·ã§ã³ãéèŠããå Žåã«ã¯éå±€ãã€ãºã䜿ããªããèããŠèŠã掟ã§ãïŒ
ïŒå
éšåŒçšã®åŒ·ãããïŒã匱ãç¹èš±ïŒççµã®ããç¹èš±ïŒã«ããããŒããã©ãªãªåœ¢æã®å¯èœæ§ããæšå®ãïŒç¹èš±ã朰ããŸãã¯ç¬å çŠæ¢æ³äžã®åé¡ãæèµ·ãïŒç©ºçœãäœãåºã察å¿ãããã®ãè¯ãã§ãããïŒ
M&Aã«ãããŠïŒå¯Ÿè±¡ãšãªãäŒæ¥ã®ç¹èš±ã®è³ªãæšå®ãïŒå®éã®æè¡åããç¹èš±æŠç¥ãåã£ãŠããç¶æ
ã§ãããã©ãããªã©èŠæ¥µãïŒå¿
èŠã«å¿ãèŠããã®ç¹èš±ããŒããã©ãªãªãè£æ£ãïŒçã®æè¡åãšå®å®æ§ãè©äŸ¡ããã®ãè¯ãã§ãããïŒ
匱ãç¹èš±ã«ãã瀟äŒã€ãããŒã·ã§ã³ãçã®ã€ãããŒã·ã§ã³ïŒãšäžèšæç®ã«èšèŒãããïŒã®é»å®³ã®çšåºŠã«é¢ããåæãªã©ïŒç¹èš±ã®è³ªãèæ
®ããããããç¥è²¡æ¿çã«é¢ããåæã«ã䜿ãããããããŸããïŒ
SeokbeomKwon (2021)ã
The prevalence of weak patents in the United States: A new method to identify weak patents and the implications for patent policy+
Technology in Society Volume 64
https://www.sciencedirect.com/science/article/abs/pii/S0160791X20312720

èµ€ç·ãå€ãç¯å²ïŒæ°èŠã®éçºãå€ããšæãããç¯å²
éç·ãå€ãç¯å²ïŒåŒ±ãç¹èš±ãæŠç¥çãªç¹èš±ãå€ããšæãããç¯å²
åŒçšã®æ¹åïŒæè¡ã®æšç§»ã®æ¹å
äžæ¬¡åŒçšã®åç
§ïŒåŒ·ãæè¡é¢é£æ§
äºæ¬¡åŒçšã®åç
§ïŒå¿çšç¯å²ã®ç€ºåïŒ
ïŒæšä»ïŒå€§çºæã®æžå°âæè¡ã§å·®ãã€ãã«ããâç¥è²¡æŠç¥æé©åïŒããžãã¹ã¢ãã«ã®å€æ§åè€éåâç¹èš±å¶åºŠã®è¶£æšãéžè±âå
¬æ£ãªç«¶äºãé»å®³ïŒæ£®å²¡ããå
¬æ£ãã€èªç±ãªç«¶äºãèæ
®ããç¥è²¡æŠç¥ã®æ çµã¿ãæ¥æ¬ç¥è²¡åŠäŒç¬¬19å幎次åŠè¡ç ç©¶çºè¡šäŒã2021幎ïŒãªã©ããåŸããšããã§ãïŒãã®ãããªéžè±ã¯çµéšäžã§ã¯åŒ±ãç¹èš±ãååšããéšåã«ããããšãå€ããšããã§ããïŒãã®ãããªéžè±ãèŠã€ãïŒãããªãã¯ãèæ
®ããããè¯ãïŒäžã€ã®äŒæ¥ã®ã¿ãèŠãã®ã§ã¯ãªãïŒç£æ¥å
šäœã«è³ããæ¹åãèŠãïŒç€ŸäŒçã«ããè¯ãæ¹éãªã©çå®ããã®ã«åœ¹ã«ç«ã€ãããããŸããïŒ
ïŒPatentfieldã®ïŒ°ïŒŠã¹ã³ã¢ãçšãïŒä»ç€ŸéèŠåºŠã¹ã³ã¢å¹³åãšèªç€ŸéèŠåºŠã¹ã³ã¢å¹³åãäŒç€Ÿããšã«æ±ãæ¯èŒãïŒäžèšèå¯ãããé¢çœããã§ãïŒ
ãPFã¹ã³ã¢ãå©çšããåžå Žãã¬ã€ã€ãŒã®åæ
ãhttps://support.patentfield.com/portal/ja/kb/articles/pf%E3%82%B9%E3%82%B3%E3%82%A2%E3%82%92%E5%88%A9%E7%94%A8%E3%81%97%E3%81%9F%E5%B8%82%E5%A0%B4%E3%83%97%E3%83%AC%E3%82%A4%E3%83%A4%E3%83%BC%E3%81%AE%E5%88%86%E6%9E%90
ïŒPatentfieldã¯æ°æ©èœã次ã
ãšå¢ããã®ã§å¥œãŸããã§ããïŒãªãæ¬ããŒã«ã§ãã¹ã³ã¢åãäœæããã°ã¹ã³ã¢ä»å¯èœïŒèªäœããŒã«ã¯ãã®ãããèªç±èªåšã§è¯ãã§ããïŒïŒïŒãªãç¹èš±ã¹ã³ã¢ã«ã€ããŠã¯æ¬¡ãåç
§ãããšè¯ãã§ãããïŒãç¥è²¡ç®¡ç71(10)2021 p1339-1354 ããŒã¿ããŒã¹æ²èŒã®ç¹èš±ã¹ã³ã¢ã掻çšããäŒæ¥äŸ¡å€è©äŸ¡ã®ç ç©¶ïŒ
ïŒäžèšåæãããåã«æ¬¡ã®æžç±ãèªã¿ä¿¯ç°ããããšãããããããŸãïŒ
ç°æãåä¹ ç·šèãå±±æ ¹ãåŽéŠ ç·šèïŒ2021ïŒãç¥è²¡ã®ããã³ãã£ã¢ã
https://www.keisoshobo.co.jp/book/b591977.html
https://www.keisoshobo.co.jp/book/b591978.html
å»è¬ç³»ã«ãããåé¡ã®ãã匱ãç¹èš±ãæé€ããããšããåãã®äžäŸ
Amid drug pricing debate, feds reveal plan to crack down on 'incremental' patents
https://www.fiercepharma.com/pharma/uspto-and-fda-commit-patent-reform-cracking-down-lengthy-add-patents
åç°å¥ïŒ2022ïŒãåŸåºãã®ç¹èš±ã«ããæ¢åäºæ¥ã®å·®æ¢ã¯èš±ãããããç¥è²¡ç®¡ç 72(8) p899-
ãçŸåšã®äžéšã®è£å€äŸã®ç«å ŽãåæãšãããšïŒæ¢åäºæ¥ãåŸåºãã®ç¹èš±ã«ããå·®ãæ¢ãããããªã¹ã¯ã¯çžåœçšåºŠããïŒã
飯島æ©ïŒ2014ïŒãèªã®äžã®ç¹èš±çŸ€åãç¹éãçã«è¯ãç¹èš±ãšã¯ã«å¯ããŠããç¥è²¡ç ãã©ãŒã©ã 96 p41-
ãçã«è¯ãç¹èš±ã¯æ³ç®çã«åèŽããç¹èš±ã§ãããªãïŒå
šäœéåãšããŠã®ç¹èš±çŸ€ã®ãã¯ãçæåãã倿ãããããªãã
ãããããè¯ãç¹èš±ã«å¯ãããã¯ãçã«æé©åãããã·ã¹ãã ã«å¯ŸãããŠãŒã¶ãŒã®éå°åå¿ãç¹èš±æŠç¥ã§ãã£ãŠïŒåé¡ã®æºæ³ã§ãããšããããã
éåžžã«ç確ã«ãç¹èš±æŠç¥ãã®æ¬è³ªãè¿°ã¹ãŠããããã«æããŸãïŒ
ïŒå人çã«ã¯ïŒãæªããšãããç¹èš±ããšã¯ïŒé²æ©æ§ã«ãããŠcommon senceãèæ
®ããç»é²ãšãªã£ãç¹èš±ïŒã ãšæããŸãïŒãã倿°ã®æè¡åžžèæç®ãããå ŽåïŒèšèŒã®ç©Žããã£ããšããŠãïŒãããåãã圢ã§äžè¬åã§ããŠãããã¹ãã§ãïŒãããã«ïŒãã®ç©ŽãããããšãææããŠç»é²ãšãªãããšãå€ã
ãããŸãïŒãã®ãããªç¹èš±ã¯çŸå®ã®åœæ¥è
ã«ãšã£ãŠã¯æããã«å®¹æã«ã§ãããã®ã§ãïŒãããŠãã®ãããªç¹èš±ã¯åŸãŠããŠãã®åéå
šäœã®æ¢åã®åºæ¬æ¥åã®å®æœãã劚害ããŸãïŒãã®ãããªæªãç¹èš±ã¯ïŒæš©å©è
ã®ã¿ã«ãšã£ãŠã¯è¯ãç¹èš±ã§ããïŒä»ãã¹ãŠã«ãšã£ãŠã¯æ³ç®çã«åèŽããªãæªãç¹èš±ã§ãïŒåºæ¬çã«ïŒãŸãæç®ã²ãšã€ã²ãšã€ãšãããç¹ãã§èããã®ã§ã¯ãªãïŒãŸãæç®çŸ€ã圢æããç¥èãšãããé¢ãã§èããã¹ãã§ãïŒ
ïŒevergteeningãèµ·ããåå ã®äžã€ã¯ãå°ããã€ãããŒã·ã§ã³ãè©äŸ¡ããããéçšã§ããããã ãšèšãããŠããŸããæ¬ããŒã«ã«ãããŠå°ããã€ãããŒã·ã§ã³ã¯å¯åºŠã®é«ãç¯å²ã«ååšãã倧ããªã€ãããŒã·ã§ã³ã¯å¯åºŠã®äœãç¯å²ã«ååšããããã§ããããevergreeningãé²ãããã«ç¹èš±åºããã®ãããªããŒã«ã䜿çšãããšãªãã°ãé¢çœããã§ãã
æ å ±åæã®ãã£ãããšããŠã¯ååã§ãããïŒ
ããŒãã«è³ã«è¿ã¥ããçºæè ã¯ããã®ãïŒ
https://www.valuenex.com/jp/materials-reports

æ¬åº¶å
çé¢é£æè¡ïŒ
æ¬åº¶å
çã®ç¹èš±ã¯3ã€ã®æ©äžžã§ç€ºãããŠããïŒç¢å°ã®ããã«æšç§»ããŠããŸãïŒããªãŒãã³å
çãçºæè
ãšãªã£ãŠããç¹èš±ã¯ç·äžžã§ç€ºãããŠããŸãïŒããªãŒãã³å
çã¯æ¬åº¶å
çã®ç¹èš±ã®ãã¡æåŸã®æ©äžžã«ååšããããããã®ç¹èš±ã«ã€ãïŒç¡å¹ã䞻匵ããŠããŸãïŒå€ååèªåã¯å
±ååºé¡éåã§ãããïŒïŒ
åºç€æè¡ããå¿çšæè¡ã«é²ãæµãããããã«èŠããŠãããšæããŸãïŒ

æç³»åæšç§»
ã¹ãã¹ã¯ã©ã€ã³ã®åããåãã§ããïŒ
ãããããªãŒããŒäŒæ¥ã®çæ³ã®ãããªåãã§ãïŒ
ããæè¡ã®èèœããèŠã€ãïŒããã®èèœã®å¿
é æè¡ã®çºæãããïŒãèèœæè¡ãã¯ãã¹ã©ã€ã»ã³ã¹ãªã©ã§åã蟌ãã§ãããïŒããã¯çéã§ããåæ¹ã«äŸ¡å€ãäžããããïŒãªãŒããŒäŒæ¥ãããæŠç¥ã§ãïŒã¹ãã¹ã¯ã©ã€ã³ã¯è£œé æè¡ã蚺ææè¡ã§ãããè¡ãããšããããã«èŠããŸãïŒçè
ã£ãŠå°è±¡ã§ãïŒ
ãæè¿ã¯èèœæè¡ã«å¯ŸãïŒããããªãã¯ããçãã ççµçºæãã被ãïŒä¿¯ç°å³ã§ã¯èèœç¹èš±ã«éãªã£ãŠèŠããããšãå€ãã§ãïŒãããªãã¯ç¯å²ã¯ãã¯ãã«èšç®æèæ
®ãããªãããšãå€ãããã§ãããïŒïŒèªãã¯äœã瀟äŒè²¢ç®ããæè¡ãé瀺ããããšãªãïŒäŒæ¥èŠæš¡ãèæ¯ã«åŒ·å¥ªããæŠç¥ãèŠãããšãå€ãã§ããïŒã¹ãã¹ã¯ã©ã€ã³ã¯ïŒãé·æçèŠç¹ãæã¡é©åãªåã蟌ã¿ãšååã«ãã䟡å€åµé ãå¿åããããããã¯ããªãã£ããïŒã£ãŠããšã§ããããïŒåŒ·å¥ªãããŠãããå°éããGSKãžã®ã©ã€ã»ã³ã¹ã¯ãªãã£ããããããªããªïŒãšãæããªãããªãã§ãïŒ
ãããªåŠæ³ãã§ããŸããïŒ
è±è£œè¬ã°ã©ã¯ãœïŒäœæåŸåœã§ç¹èš±ç³è«ãããåŸçºè¬çç£ä¿ã
https://www.nikkei.com/article/DGXLASGM02H0F_S6A400C1NNE000/
æŠç¥å€§è»¢æããäžçãå€ããäŒæ¥ã1äœã«èŒããGSK
äžæ£ã®çµ¶ããªãè£œè¬æ¥çã ããããïŒèª å®ããæåªå
ã«
https://jbpress.ismedia.jp/articles/-/48701
ã»ã»ã»
ã¢ã¹ãã©ãŒãã«ã¯ã¹ãã¹ã¯ã©ã€ã³ãšéåç°ãªãåºé¡åœ¢æ
ã§ãïŒ
ããŠïŒåºé¡åœ¢æ
ãšäŸµå®³ã«é¢ä¿ã¯ããã®ã§ããããïŒã
å°éè¬åãç¹èš±äŸµå®³ã§ã¢ã¹ãã©ãŒãã«ãæèšŽãã€ããã£ã³ãžã®è²©å£²å·®ãæ¢ããšæå®³è³ åæ±ãã
https://www.mixonline.jp/tabid55.html?artid=72665
ãç¹èš±ç¬¬5885764å·ãšç¹èš±ç¬¬6258428å·ã«åºã¥ãïŒãæPD-L1æäœãçšããŠããæ²»çãè¡ãããšã®åœå
ã®çšéç¹èš±ã䟵害ããŠãããšäž»åŒµã
ïŒå
šãé¢ä¿ãªãã§ããã¢ã¹ãã©ãŒãã«ã®githubïŒé¢çœããã®ã䞊ãã§ããŸãã
https://github.com/AstraZeneca/judgyprophet
https://github.com/AstraZeneca/StarGazer
https://github.com/AstraZeneca/chemicalx
https://github.com/AstraZeneca/data-science-python-course
https://github.com/AstraZeneca/awesome-drug-discovery-knowledge-graphs
https://github.com/AstraZeneca/awesome-explainable-graph-reasoning
https://github.com/Roche/BalancedLossNLP
ïŒå
ç«ãã§ãã¯ãã€ã³ãé»å®³è¬ æPD-1/PD-L1æäœã«ç¶ãéçºã®æœ®æµ
https://answers.ten-navi.com/pharmanews/15721/
ãUPDATEãå
ç«ãã§ãã¯ãã€ã³ãé»å®³è¬ïŒæPD-1/PD-L1/CTLA-4æäœ åœå
éçºç¶æ³ãŸãšã
https://answers.ten-navi.com/pharmanews/7342/
â»ã³ããã¯ã¯ãã³ã«ã€ããµé€æ²¹ã®æâŠããŸã¿æåã®ç ç©¶ãïœïŒ²ïŒ®ïŒ¡ã®åæäŸçµŠã«é²å
https://www.yomiuri.co.jp/economy/20211013-OYT1T50086/
å³ã«ã€ããµããªãã§ãšæã£ãŠãããããããïŒãã¯ãåæããŠãªãïŒ
â»2004åºé¡ã®ãã¯ã€ããããç ç©¶æoct4ç¹èš±ã®åœ±é¿åãé¢çœããšããã§ããæ§é ããã©ã®äœçœ®ã«ããã§ããããã
ïŒæè¡ã®æšç§»ãïŒïŒåšèŸºèªçŸ©ããå®çŸ©ããããšãããïŒåèªã®æå³å€åãšããŠãšãããŠãé¢çœããšæããŸãïŒãããã¯ã¢ãã«ã§æšç§»ãèŠãŠãã®ãè¯ãããã§ãïŒ
USç¹èš±åºã®ã³ã³ãPPPMãšçµ¡ããŠãè峿·±ãã§ãïŒ
ãåèªã®éæçãªæå³å€åã®ã¢ãã«å
ãhttps://speakerdeck.com/seiichiinoue/dan-yu-notong-shi-de-nayi-wei-bian-hua-nomoderuhua
ãåçãããã¯ã¢ãã«ãçšããåèªã®éæçãªæå³å€åã®ã¢ãã«åãšãã®å¿çš
ãhttps://www.ism.ac.jp/~daichi/workshop/2022-lchange/lchange2022-inoue.pdf
é»ç£éŒæ¿ã«ã€ããŠïŒæ¥æ¬è£œéãšããšã¿ãªã©ã®ç¹èš±ãçºãã

䟵害äºä»¶ãèŠãããã®ã§è峿¬äœã«ïŒ
è峿·±ãåŸåãããããã«èŠããŸã
ã»ãïŒæ¯éå£å
šäœã®ç¹æ§ãåŒã³æ°ŽãšããŠïŒç®çã®ç¹èš±ã®äœçœ®ãå
éåãããã€ã¡ãŒãžãæŽããããšæããŸãïŒ
ãã®ããã«ïŒæå³çã«å³ã®æåœ±é¢ãæ²ããããç¹ã¯ïŒæ¯éå£ã«å¿ããŠãã¯ãã«ãå€åããã¢ã«ãŽãªãºã ã®åŒ·ã¿ã§ãïŒ
ãããïŒã¹ãºãïŒã»ã®ç¹èš±ãçºããïŒæç®èŸŒã¿
äžããããã¡ã³ãšããŠè峿¬äœã«ïŒ2010幎ååŸããé¢çœãããšã«ãªã£ãŠããã®ã§ã¯ãšïŒ
ç¹ç«ã³ã³ãããŒã«ïŒè»äœïŒé»æ°èªåè»ã«ã¯æ§ã
ãªã¢ãããŒãããïŒããã°ã©ã ç³»ã¯ãããã®ãã³ãã§ããããïŒ
ä»åã¯ä¿¯ç°å³ã§ã¯ãªãïŒä¿¯ç°å³ãäœæããéã«çšãããã¯ãã«ç©ºéããããã¯ãŒã¯åãïŒäœãããããèŠãŠãããŸãïŒ
ãªãïŒãã®ãããã¯ãŒã¯ã¯ããŒã¯ãŒãåäœã®ãããã¯ãŒã¯ã§ã¯ãªãïŒæŠå¿µåäœã®ãããã¯ãŒã¯ã§ãïŒããŒã¯ãŒããšç°ãªãæŠå¿µã§ãããæŠå¿µåå£«ã®æŒç®ãå¯èœã§ããïŒé¡äŒŒæŠå¿µã¯åžåãããŠããŸãïŒ
ãã®3瀟ãªããã³ããæ¯é
çãšæããŸãããããã§ããªãã§ããããïŒãåå¯ããäžååãããããŸããïŒãã³ãã¯ã©ã€ã»ã³ã¹åå
¥ãå€ãããšããèããã«ãåºç€çãªç¹èš±ãããããããç¬èªæ§ãäœãããã«èŠããã®ã§ããããæç³»ååæãããã°åºç€çç¹èš±ã«ããæ¯é
ãèŠãããããããŸããã
ãããã¯ç¬èªæ§ã匷ãããã«èŠããŸãïŒå·Šäžã®æŠå¿µèŠçŽ çŸ€ã¯ã»ãŒãããã®ã¿ãšç¹ããç®ç«ã£ãŠããŸãïŒfeedbackã¯ããã«ãããŸããïŒããããããå°è±¡ã§ãïŒããïŒç¢ºããããã¯feedbackãå¿
èŠãšããªãçè«åã«åãçµãã§ããã®ã§ãããïŒ
ã¹ãºãã¯OEMäžå¿ã§ããæ°èŠæè¡ããæ¹è¯ãéžæããŠããã®ãããããŸããããããã瀟é·ãŠãå
¥ãã§çºæå ±å¥šã匷åããŠããããã§ããããŸããä»ã²ãšã€æ¹åæ§ãèªããŸããããåºæ¬çã«ã¯æè¡ããæ¢ç¥ã®åºç€åŒ·åã®æ¹éãªã®ã§ããããæè¡å·®å¥åãããµãŒãã¹åŒ·åãšèšããã³ã¢ãã£ãã£ã®åºæ¬ãéžæããŠããã®ã§ãããå€åã
äŒç€Ÿããšã®éãã«å¯Ÿããè±å¯ãªã€ã³ãµã€ããèŠã€ãããããªãããã¯ãŒã¯ã§ãã
-
å šäœã幎代ãšã®é¢é£

'hydraul','analysi','textur'ã£ãŠãªãã§ãããïŒâããã®è«æã¯ïŒç»åã®å®éãã¯ã¹ãã£ãŒè§£æãœããMazdaã玹ä»ãããã®ã§ãããã»ã»ã»æç®ã¯é€å»ããšããŸãããïŒããŒã¿åæã®æé ïŒask-prepare-process-analyze-share-actã®ãã¡processãçããŠããã®ã§ãããŸããããã»ã»ã»
æç®ãšç¹èš±ã®é¢é£ãèŠããšé¢çœãåŸåãèŠããããšãå€ãã®ã§ããïŒä»åã¯æç®ã«ãã€ãºãå€ãæ··åšããŠããïŒãªã«ãèŠåºãããšã¯ã§ããŸããããïŒæ¬æ¥ãªãæç®ã®æ¯éå£ã¯äœãçŽãã§ããïŒ
1900-2000å¹Žã®æè¡ãäžå¿ã«ãã£ãŠïŒãã®åŸã®å¹Žä»£ã¯ãããã«å€çžã«åã蟌ãã§ããããã§ãïŒ
èªåè»æ¥çã¯ã³ã¢ãã£ãã£ã«è³ãå¯èœæ§ãå«å
ãã€ã€ãïŒæã
ã«ãããŠæè¡ã·ãããé©åã«ãªããïŒç¶æãããŠããããšã瀺ããŠãããããã§ãïŒïŒæè¡ã·ãããªã幎代ããšã®ç¬ç«æ§ãã幎代ããšã«åžåãããå§¿ãèŠããããªãã®ã§ããïŒãã®èŠåº§ã§ã¯èŠããŸããã§ããïŒïŒïŒåºç€20幎ãããŒã¹ã«äœå¹Žåšæãèšæž¬ããŠã¿ãã®ãé¢çœããã§ãïŒïŒèªåè»æ¥çã¯æ¬¡ã
ã«ãã£ãŠãã瀟äŒç課é¡ã解決ããŠããã°èªåçã«æè¡ã·ãããéæãããã®ãããªãããªãšæãïŒãããã³ã¢ãã£ãã£ãåé¿ã§ããçç±ãããªãããªãšãæããŸãïŒãããã®ãããªååçã§èœåçã§ãªãäºæ
ã§ã³ã¢ãã£ãã£ãåé¿ããŠããã®ã ãšãããšïŒç€ŸäŒç課é¡ã匱ãŸã£ããšãã«ã¯æå€ãšãã£ããã³ã¢ãã£ãã£ã«èœã¡ãããšãããã®ãããããŸããïŒïŒ
çŠç¹ãåã
ã®èŠçŽ ãããã©ã³ã¹ã«ç§»è¡ããŠããããã«èŠããªãããªãã§ãïŒ
hydraulã¯ããã·ã§ã³é¢é£ã®ããã§ãïŒ
è¿å¹ŽéèŠãšèšããããCASEãïŒConnectedïŒAutonomousïŒSharedïŒElectricé¢é£ã®ããŒãã®ç¹ããã¯èªåã«ã¯èŠããŸããïŒ
äºåç¥èãããæ¹ã¯ã€ã³ãµã€ããåŸããããããŸããïŒããšã¿ãå«ãããåŒã³æ°Žã䜿ããªããšããèŠããªãã®ãããããŸããïŒèŠç¹èŠçŽ ãèªåæœåºãšããŸãããä»»æã«è¿œå ããã»ããè¯ããããããŸããïŒãã现éšãèŠããããã«ïŒèŠåº§ãäœãããã¹ããããããŸããïŒïŒä¿¯ç°å³ã®ã»ããç®èŠã§ã®æ°èŠåéã®çºèŠã«åããŠããããã«æããŸãïŒæ°ã«åããã«ããã§ãïŒïŒ
åŸè¿°ã®éãïŒãããžã¡ã³ãã«éå®ããã°CASEé¢é£æè¡ããããã¯ãŒã¯ã§èŠããŠããŸãïŒ
èªåã«ã¯ãã®ãããã¯ãŒã¯ããã€ã³ãµã€ããã²ãã ãã ãã®ç¥èãè¶³ããªãããã§ãã調æ»ãç¹°ãè¿ãåºç€ç¥èãèŠã«ã€ããããçŽãã¹ãã§ããããã
2000-2015ãŸã§ã¯æ°èŠã«æãåºãïŒãã以éã¯æ¯èŒçåºæ¬ã«ç«ã¡æ»ã£ãŠããïŒã£ãŠæãã§ããããïŒ
-
è¿å¹Žã®ããããäžéšé å

longitudinalïŒãã©ãŒïŒåŒ·åã»ã»ã»ããããããã§ãããïŒ -
è¿å¹Žã®ããããèŠåº§å€æŽ

controlãåªä»äžå¿ãšãªã£ãŠããŸãïŒããããããå°è±¡ã§ãïŒ
å ±éèšèªãæç¢ºã§ããïŒã£ãŠããšã§ããããïŒ -
è¿å¹Žã®ããã以å€ãäžéšé å

è¿å¹Žã®ããã以å€ãèŠåº§å€æŽ

ãããã§ã¯controlãä»ãèåããŠããé åãé¢ããŠããŸããã»ã»ã»
ãããã§ã¯æç¢ºã«èŠãããªãã£ããããžã¡ã³ãé åãååšããŠããŸãïŒ -
å šäœãèŠç¹å€æŽ:è¿å¹Žã®ããã°ã©ã é¢é£é å

ç¹ç«ïŒé å»¶ïŒã³ã³ãããŒã«ïŒãšãµãŒããŒç³»ïŒãµãŒããŒç³»ãå€ãçµæ§å€æ§ïŒã£ãšã»ã»ã»
CASEã¯ãã¡ãã«è¡šããŠãããšæãã®ã§ããâŠ
èªåè»è£œé æ¥ã®ç¹èš±åºé¡ååã«é¢ããåæ
https://www.nistep.go.jp/wp/wp-content/uploads/column_2018_4.pdf
ãæ¬åæã§ã¯ïŒ1989ïœ2011 å¹Žã®æéã«åœå
åºé¡ãããç¹èš±ã®ãã¡ïŒèªåè»è£œé æ¥ã®äŒæ¥ïŒ2011 幎æç¹ã§ 202 瀟ïŒãåºé¡äººã«å«ãŸããç¹èš±ã®åºé¡æ°ãæè¡ããšã«éèšããããšã«ãã£ãŠïŒèªåè»è£œé æ¥ãåãå·»ãæè¡ååã®å€åãç¹èš±åºé¡ååããæããããšã詊ã¿ãïŒ ã
ã1990 幎代ãã 2000 幎ã«ãããŠïŒèŒžéçšæ©åšã«æ¬¡ãäž»èŠæè¡ãæ©æ¢°å·¥åŠãã黿°å·¥åŠã«ã·ããããŠããŠããããšããããïŒãã®ç¹èš±åºé¡ååã¯ïŒèªåè»éšåã®ã¢ãžã¥ãŒã«åã»é»ååã®æµãã«äŒŽãïŒæ©æ¢°ç³»ã«å ããŠé»åå¶åŸ¡ç³»ã®æè¡ã®åœ¹å²ãé«ãŸã£ãŠããŠããããšã瀺ãåæ ããçµæãšæããããïŒ ã
ãäž»èŠãªæè¡åéã«ã€ããŠèŠãŠãããšïŒæ©æ¢°å·¥åŠã«è©²åœããæè¡åé¡ïŒç·è²ïŒã®ãã¡ïŒãå·¥äœæ©æ¢°ãïŒãä»ã®ç¹æ®æ©æ¢°ãã®é äœã¯äœäžããŠããïŒäžæ¹ïŒé»æ°å·¥åŠã«è©²åœããæè¡åé¡ïŒéè²ïŒã¯1990 å¹Žãšæ¯ã¹ãŠå
šãŠé äœãäžæããŠããïŒèªåè»è£œé æ¥ã«ãããååšæãå¢ããŠããŠããïŒãã®ä»ã®æè¡åéã«ã€ããŠèŠãŠãããšïŒæ
å ±éä¿¡æè¡ïŒæ¡è²ïŒã«ãããŠã¯ãã³ã³ãã¥ãŒã¿ãŒæè¡ãã®é äœãäžæããŠããïŒã
ãããããžã¡ã³ãã®ããã® IT ææ³ãã«é¢ããç¹èš±åºé¡ã¯ïŒ2000 幎é ãŸã§è¡ãããŠããªãã£ããïŒ2010 幎é ã«ã¯è¡ãããããã«ãªã£ãŠããŠããïŒåœè©²æè¡åéã§ã¯ïŒã«ãŒã·ã§ã¢ãªã³ã°ã«é¢é£ããç¹èš±åºé¡ãªã©ãèŠããïŒèªåè»ç£æ¥ã®æ°ããªå€åã®å
ããèªã¿åããïŒãŸãïŒãã€ãªã»å»è¬åïŒé»è²ïŒã§ã¯ïŒãã€ãªãã¯ãããžãŒã®é äœãããäžæããŠããïŒåœè©²æè¡åé¡ã§ã¯ãã€ãªçæã«é¢é£ããç¹èš±åºé¡ãªã©ãèŠãããïŒ ã
- å
šäœã幎代å¥ãèŠç¹å€æŽ:ãããžã¡ã³ãé¢é£é å

ãªãã»ã©ïŒãããã«2000幎é ãŸã§ã¯è¡ãããŠããªãåŸåã§ãããïŒ
2015幎以åã¯ã¹ããŒãã³ã³ãããŒã«ç³»ã®ãããžã¡ã³ãã§ããããïŒ
2015幎以éã¯ãµãŒããŒãšåŒ·ãçµã³ã€ãïŒæ°ããªã¯ã©ã¹ã¿ãäœãäžããŠããããã§ãïŒãã®ã¯ã©ã¹ã¿ã¯user,inform,notifãå«ã¿ãŸãïŒã³ãã¯ãããã«ãŒé¢é£ãå€ãããã§ãïŒ
2000-2010ã«ã¯ïŒuser,drive,informã®ç¹ãããèŠããŸãïŒãã®é ãã«ãŒã·ã§ã¢ãªã³ã°é¢é£ã ã£ãã®ã§ããããïŒæ§æèŠçŽ ã¯ã³ãã¯ãããã«ãŒã«äŒŒãŠããã®ã§ãããïŒãã§ãããªãé¢é£äŒæ¥ã䌌ãŠããã®ãããããŸããïŒ
2010-2015ãserverãšã€ãªãã£ãŠããªãããšãæ°ã«ãªããŸãïŒäžåºŠæè¡ãéåãããã®ããã§ãïŒæä»£çã«ã¯ãšããžã³ã³ãã¥ãŒãã£ã³ã°ãæµè¡ã£ãé ã§ããããïŒ
ãã®ãããã¯ãŒã¯ããããçšåºŠã®ã€ã³ãµã€ããåŸãããŸããïŒ
ããŒãã³ãŒãæ¯ã®æè¡åé¡ãããªã¯ã¹ããããšäœµçšããã°ããããªãã€ã³ãµã€ããåŸããããã§ãïŒãã¡ãã§èª¬æã§ããªããã°ïŒãŸã åé¡ã«ãªãïŒæ°èŠåéã®èèœãèŠãããŠãããšè§£éã§ããã®ãããããŸããïŒ
-
å šäœãæ°åèªåºçŸææ

æ°èŠããŒã¯ãŒãã¯ããªãå°ãªããªã£ãŠããŸããã¢ãžã¥ãŒã«åãé²ãã§ããããŸãããã¯ãã£ã¹ãã®åºé¡ãå°ãªããã£ãŠããšã§ãããããç¹èš±ã®äŸ¡å€ãå šäœãšããŠçž®å°ããŠãã£ãŠããã®ã§ãããããŸããã³ãã£ãŒãããããŸããã3瀟ãšããªãŒããŒæŠç¥ãåãããšãå¯èœãªäŒæ¥ãšããã°ãããããããŸãã⊠-
å šäœã被åŒçšæ°ã®æšç§»
倿ç¹ã確èªããŸãã -
å šäœã幎代å¥ã2010以éåºçŸã¯ã©ã¹ïŒã»ã¯ã·ã§ã³åäœãïŒåé¡èŠç¹ãŒæŠå¿µèŠç¹ã®æ··å衚瀺ïŒèŠåº§èª¿æŽ

ïŒç掻å¿
éåïŒïŒïŒ¢ïŒåŠçæäœïŒé茞ïŒïŒïŒ£ïŒååŠïŒå¶éïŒïŒïŒ€ïŒç¹ç¶ïŒçŽïŒïŒïŒ¥ïŒåºå®æ§é ç©ïŒïŒïŒŠïŒæ©æ¢°å·¥åŠïŒç
§æïŒå ç±ïŒæŠåšïŒçç ŽïŒïŒïŒ§ïŒç©çåŠïŒïŒïŒšïŒé»æ°ïŒ
æå€ãšGã»ã¯ã·ã§ã³ãå¢ããŠããã®ã§ãããïŒG06ãG16ã§ãããã¡ããæ°ã¥ããŸããã§ãããšãïŒ
A41è¡£æïŒA42é ã«ïŒA61å»åŠãæ°ã«ãªããŸã
ã¯ã©ã¹åäœã§ãªããšãã¯ãããããŸããïŒå
šåºçŸã¯ã©ã¹ã§ã¯è€éã«ãªããããã®ã§ïŒæ°ã«ãªãåé¡ãã¯ã©ã¹åäœã§ä»»æã«æå®ããŠã¿ãŸãããïŒ
ã»ã»ã»
å
šãããããªãåéã«ã€ããŠåœããä»ããè¡ã£ãå Žåã¯ããŸããããªæµãã«ãªãåŸãã§ããããïŒ
çãåããã¯ããŠããŸãããïŒãªã«ãé¢çœãã€ã³ãµã€ããåŸããããŠããã§ãããã?
ã»å人çã«ã¯ïŒèŠåº§ã»èŠç¹ã®éãã«ããåãæç« 矀ããè€æ°ã®ç¥èãåãåºããšã¯ã©ãããããšãïŒçé±ã®ã€ã¡ãŒãžãã€ããããšè¯ããšæããŸãïŒ
ãªãïŒç¥èã®å€æ§åãå人ãäœéšããŠããã°ïŒèŠããªããã®ãèŠãããšãèŠãããã®ããæŽã«äœããèŠãããšããïŒæ¬æ¥ã®æå³ã§ã®æ¢çŽ¢ãã§ããããã«ãªãã®ã§ã¯ãªããïŒãšæ·¡ãæåŸ ãããŠãããšããã§ããããšããã§ãïŒ
ïŒãããã¯ãŒã¯ãè€éã«ãªãããèŠã«ããå Žåã«ã¯ïŒfilterãããããšè¯ãã§ããã
äŸãã°ä»¥äžã¯ïŒäºæž¬ã課é¡ãšãããšããžãšããŒããïŒãã£ã«ã¿ãªã³ã°ããŠåãåºããŠããŸãïŒ

ïŒé¢é£èšŽèšããã®ä»ã¡ã¢
å
šäœã®ç¶æ³ïŒ
è»ã®ç¹èš±èšŽèšïŒãååŒå
ä»»ããã®æ
£è¡èŠçŽãä¿ã
https://www.nikkei.com/article/DGXZQOCD17C7U0X11C21A1000000/?unlock=1
ç¹èš±è£åæ¡é
ã®ç¯å²å€ãŸãç¡å¯Ÿå¿ã«ãããå®è³ªçãªæå®³çºçã®å¯èœæ§ïŒ
蚎èšå¯Ÿå¿èœå匷åã®æ¹éãšãªãã®ã§ãããããïŒ
ãããã®ç¶æ³ïŒ
䟵害ã¯è¢«åå€ã
䟵害被åã®äŸã¯ã»ã»ã»ã·ã¹ãã ç³»ãåå°äœç³»ãå€ãããª
ãã³ãv.ããã
JP3196076ã什å3(ã¯)28206ãç¹èš±æºäºåŸã®äŸµå®³èšŽèš
https://patents.google.com/patent/JP3196076B2/en?oq=JP3196076
ATã«ããããã¬ãŒãèžèŸŒã¿è§£æŸæã®ãã¬ãŒãæç¶è£
çœ®ãšæ
鿀ç¥è£
眮ïŒ
ãã©ã¯ã·ã§ã³ã³ã³ãããŒã«è£
眮ãïŒ
ãã€ããã«ããç¡å¹å¯©å€ã¯ç¶æïŒãããã«ããç¡å¹å¯©å€ã¯ç¶ç¶äž
Daedalus Prime LLC
åå°äœç³»ïŒïŒ®ïŒ°ïŒ¥ãšå€æããŠè¯ãã®ããªããïŒ
ãã³ãã®ç¶æ³ïŒ
䟵害ã¯ååå€ã
䟵害被åã®äŸã¯ã»ã»ã»ã·ã¹ãã ç³»ïŒã«ã¡ã©ïŒåå°äœã»ã»ã»ãã£ãšããããšäŒŒãŠããã ãããïŒ
æ¬æ¥ã§ãªãé åã§äŸµå®³ãèµ·ããŠãããšèããŠããã ãããïŒ
ãã³ãã®é»åç³»åºé¡ã¯æ¯èŒçå€ãããçžæã§ããããŸãæå³ãç¡ãã£ãïŒïŒ
ïŒç¹èš±ã©ã³ãã¹ã±ãŒãã§ã¿ãå
端ãã¯ãããžãŒãããšã¿ã®IPæŠç¥ãEVã»æ°ŽçŽ çæè»ã»SiCã»èªåé転ã§åšå°ãªç¹èš±ç¶²
https://xtech.nikkei.com/atcl/nxt/column/18/01662/00010/
ïŒä»¥äžç¢ºèªããŠãé¢çœããã§ãïŒæ¯éå£ã«å ããã¹ãèæ¯ãšããŠããšã¿ãšäžåœã¯å¿
é ã ã£ããããããŸããïŒä¿¯ç°åæã§ã¯èæ¯æ
å ±ãéèŠã§ããããã¯ãã»ã»ã»
ã»äžåœã¯ããã¶ã€ãããªãææ³ãšãããã£ã®ã¯ã©ãã·ã¥å®å
šæ§ãäœãããŠãããšãããæ¥æ¬ã¯åç¹ååž°ãã¯ã©ãã·ã¥å®å
šæ§ãéèŠããŠããããã«èŠãããå®å
šãšããèŠç¹ã§ã¯åããšãããã©ã¡ãã«åããã®ã ãããã
ã¯ã©ãã·ã¥å®å
šæ§ã¯ãã§ã«ååãšãªã£ãŠããããã«èŠããªãããªããä»åŸã®æŠåšã§ã¯ãªãã§ã¯ãããã
äžåœãæ¯éå£ã«å ããŠããã°ïŒãããã¯ãŒã¯ã®æ§é ã®éãã«ããããã£ã«é¢ãç°ãªãèŠç¹ã§æ³šç®ããŠãããšããã£ãã ãããïŒ
ã»BYDãNIOãªã©äžåœäŒæ¥ãæ¯éå£ã«å ãããšéã£ããã®ãèŠããã ããã
ã»åå°äœã§ã¯ãã§ã«è² ããŠãããšããããã¯ãŒåå°äœã§ã¯ãç¥è²¡å€ã®æŠç¥ã掻çºãšãããæ¬¡åã¯åå°äœã«ã€ããŠïŒçŽ çŽã«åé¡ããŒã¹ã§åæããŠã¿ãããïŒ
ã»CDå€ã«è²¢ç®ããç¹èš±ã¯ããéèŠãšãªã£ããšã®ããšïŒCDå€ã«é¢ããç¹èš±çŸ€ããããã¯ãŒã¯ããã¯æŸããŠããªãããã§ãã£ãïŒããæ
å ±ãšããŠåœ¢æãããªãã£ãã®ãïŒãããã¯ãŒã¯çææã«èœãšããã®ãïŒç¢ºèªããŠã¿ãïŒ
ã»SiCãã¯ãŒåå°äœã«ããèå軜éå?
ã»æé
ãã®è¿œåŸã§ã¯åãŠãªããä»åŸã©ãã«åããã®ã ããããããããIPã©ã³ãã¹ã±ãŒãã®åœ¹å²ã®ïŒã€ã¯æªæ¥äºæž¬ã§ãããïŒã©ããŸã§è¿°ã¹ãããã ãããïŒ
ã»æ¯éå£ã®IPCãšFã¿ãŒã ã®çžé¢ãèŠãããšã«ãããåé¡ãããç¯å²ã§ããã°ã空çœé åãèŠã€ãããŸããã€ã³ãµã€ããåŸãŠãããä»ãããã®ã¡ïŒãã¡ãã§å ·äœçã«ç¢ºèªããŠãã¿ãã¹ãã§ãããã
æ¯éå£ã®IPCãšFã¿ãŒã ã®çžé¢ããŒã¿ã¯å€§ãããããšæãããŸãïŒ
è¡ãåå士ãããç·åœããã§é¡äŒŒåºŠãèšç®ãïŒé¡äŒŒããŠããç¯å²ã§çžéã空çœãèŠã€ããïŒãšè¯ãããã§ãïŒ

é¡äŒŒããŠãã課é¡ã®ç¯å²å
ã§ããç¹å®ã®èª²é¡ã§ã¯ãªãããŠããªãè§£æ±ºææ®µãèŠã€ãããïŒé¡äŒŒããŠããè§£æ±ºææ®µã®ç¯å²å
ã§ç¹å®ã®è§£æ±ºææ®µã䜿ãããŠããªã課é¡ãèŠã€ããïŒããšãå¯èœã§ããïŒ
ããã ãã§ã¯ãããŸã§åé¡ã®ããç¯å²å
ã§ã®ç¥èŠãšãªããŸãïŒ
ãã®ãããªæ¢åæ§é ããåŸãããç¥èŠã䜿ãç¯å²ãéå®ã俯ç°ãçŽããšïŒããè¯ãããã§ãïŒ
ïŒæµ·å€ã®èªåè»éšåã¡ãŒã«ãŒ5ç€Ÿã®æè¡ã俯ç°çã«æ¯èŒãã
https://note.com/valuenex/n/ne4a15b5e6071
ãªãã»ã©
ã»èªåè»æ¥çãåãéããããèŠå ãåæããŠãã ãã
çŸç¶GPTsïŒGPT4oïŒ
https://chatgpt.com/share/66e50daa-4268-8006-b381-12d4a0b1263f
openai o1 preview
https://chatgpt.com/share/66e76f57-ea7c-8006-ae29-b76cf2731c81
䜿çšäŸïŒ.ïŒïŒIPã©ã³ãã¹ã±ãŒãïŒã«å©çš
IPã©ã³ãã¹ã±ãŒãã®å®çŸ©ã¯ææ§ïŒäžæµã®çµå¶äºæ¥ã¡ã€ã³ã§ããäžæµã¡ã€ã³ã®å Žåã¯IPã©ã³ãã¹ã±ãŒããšã¯åŒã°ãªããªã©ïŒãšèªèããŠããIPã©ã³ãã¹ã±ãŒããšåŒã¶ããšãæ£ãããããããŸãããïŒäŸãã°ç ç©¶éçºã®æ¹éãæ±ºããéã«æ¬¡ã®ãããªæµãã®äžã§å©çšã§ããã§ãããïŒ
ã»äœ¿çšäŸïŒïŒïŒã®ç¹èš±çæã¯ïŒäžå³ã®1ã«ãããŠç©ºçœé åã®ç¹èš±ãçæããŠããããžã§ã³ã®åèãšããïŒãšããäœ¿ãæ¹ãæ³å®ããŠããŸãïŒ
ããå°ããã€ãªãã¶ã€ã³ãåèã«å
·äœçã«èšããšã
ã顧客ãšçŽæ¥èªãå Žãäœã
ãå€ãã®é¡§å®¢ãšèªãïŒååžãææ¡ãã
ã顧客ããããããããæ
å ±ãææ¡ãããã®ã§ãªããïŒãååžãšæ
å ±ããç¥èããäœããïŒé¡§å®¢ã®æ¬²ããã¯å€§æŠãããªãïŒé¡§å®¢ã¯ãããç®ã®åã«æ¥ããŸã§ãããã»ãããç¥ããªãïŒé¡§å®¢ã«çŽæ¥èããŠãçã®ããŒãºåã¯ããããªãïŒ
ã空çœé åãç¹èš±çæãåèãšãïŒäœãäžããç¥èãå
ã«ïŒååžã«åœãŠã¯ãŸãã€ã³ãµã€ããåŸã
ãæ£ç¢ºãªååžãåæãšããŠïŒããçå€ã«è¿ãïŒãã顧客ã«äŸ¡å€ãäžãïŒããåçãšãªãïŒããŒãºèª²é¡ãšããžã§ã³ãåŸã

ïŒIPã©ã³ãã¹ã±ãŒãã«é¢ããŠã¯ïŒIP ePlat https://ipeplat.inpit.go.jp/Elearning/ ã«æ²èŒãããŠããåºç€è§£èª¬ããããããããšæããŠããŸãïŒ
â»å人çã«ã¯ãäŒæ¥ã§ã¯ãªããæè³ã¹ã¿ãŒãã¢ããã倧åŠãããIPã©ã³ãã¹ã±ãŒããå¿
èŠãªã®ã ãšæã£ãŠããŸãã
倧åŠã«ç¥è²¡èœåãå¿
èŠã§ããããšã¯åæãšããŠã
倧åŠã«ã¯ããã«ãç®çãç ç©¶æè¡ã®çŸå®ã®å®æœãšã次ã«ç ç©¶è²»è£å©ã®çºã®å©çãšããæç¢ºãªèšå®ãšããããå®çŸããããã®ãæ¥çåæèœåããšãé«ãã©ã€ã»ã³ã¹èœåãå¿
èŠãªã®ã§ã¯ãªãã®ããšæããŸãã
IPã©ã³ãã¹ã±ãŒããªã©ã¯ãæ¥çæ
å ±ã¯ããçšåºŠããã£ãŠãã詳现ãŸã§ã¯å¿
èŠãšã¯ããªããããããªãäŒæ¥ã§ã¯ãªããæ¥çæ
å ±ãããç¥ããªãå·šèŠçã®ã¿ã®èŠç¹ã§ããéèŠãšãªãåŸã倧åŠïŒãæè³å®¶ãã¹ã¿ãŒãã¢ããïŒã«ãããå¿
èŠãªã®ã§ã¯ãªãããšã
â»å人çã«ã¯ïŒ
ãçµç¹ã®çµæžåŠã®ããã³ãã£ã¢ã𿥿¬ã®äŒæ¥çµç¹ïŒæ¥çµã2023ïŒã ãèªããšè¯ãããšæããŸãïŒ
ãããããã§äž»èгã«äŸåãããèãSWOTã4Cããã¶ã€ã³ææ³ããã¬ã€ã³ã¹ããŒãã³ã°ã«ããã¢ã€ãã£ã¢åºããäžå¿ãã®ææ³ãšã¯äžæ©é¢ããïŒãããããäžå¿ãšããŠå®è¡ããŠããŸããšããŸããããªãïŒããã³ã³ãµã«ãè¿°ã¹ãææ³ã§ã¯ããŸããããªãïŒãããšãžã®ïŒåçãæžããŠããããã«æããŸãïŒãªãã€ãããŒã·ã§ã³è©äŸ¡ã®é«ãã£ãã·ã£ãŒããç Žç¶»ããã®ãã®çç±ãæžããŠãããšæããŸãïŒå
ã®ææ³ã§ã¯ãã®ç Žç¶»ãåé¿ã§ããŸãããããã®ç Žç¶»ã«åããããããªããŸããããåãããªãããã«ã¯ã©ãããããã§ããïŒã
èªåããæåã«ã俯ç°ã§çããçµãããšãå¿
èŠãšè¿°ã¹ãããç¥èçæãå¿
èŠããšè¿°ã¹ãŠããã®ã¯ããã®æå³ã§ãïŒãŸãïŒããŸãããäŸã§ã¯ããããçè§£ããŠãããããåæã ãããšããããŸããïŒïŒïŒ
ïŒå人çã«ã¯ïŒçµå¶ãšæ¥ç¶ããIPã©ã³ãã¹ã±ãŒãã¯ããŸãæšå¥šã§ããŸããïŒ
èªåã¯çµå¶ãšã®äžèº«äžäœãæåªå
ãšã¯èããŠãããïŒãããããtableauã®ãããªãçŸå Žã¬ãã«ãã®ãå人ãããããã®èŠç¹ã§è¡ãåæãå¿
èŠã§ãããïŒããã«ãããäºæ¥ã¬ãã«ã§ã®å®è¡ãå¿
èŠãã§ããïŒãã¡ããåªå
ãããã¹ãè«è
ã§ãïŒ
çµå±ã®ãšããïŒããããèªèº«ã§ãããçè§£ã§ããç¥èã圢æããããšãã§ããã®ã¯ïŒè©äŸ¡ã確å®ãææãéããŠããïŒãšãªããããã®ã§ããããïŒææãéããããªããªãã°å
ã«äºæ¥ã¬ãã«ã§äŸ¡å€ãäœã£ãŠããïŒãšããèãæ¹ã§ããïŒãããèããã°ãªãœãŒã¹ã®éèŠæ§ãçŸå Žåäœã§ãååãã«çè§£ããããããªãïŒæ¥åå¹çåã®å¿
èŠæ§ãªã©çè§£ããããããªãïŒèªåçã«åãçµç¹ãšãªããããã§ãããïŒå人çã«ã¯äžã®äžãããäžçªè¶³ããŠããªãã®ã ãšæããŸãïŒ
ïŒç€Ÿå
ã®ææ¡ã³ã³ãã®ãããªïŒæç¥šã»å€æ°æ±ºãïŒäºãå
šå¡ãçè§£ããŠãã䟡å€ã®ã¿ãéžæããããšã«ãªãææ³ãããã§ãããïŒããããã®äŸ¡å€ã¯è¯ãéžæãšãªãã§ããããïŒ
åå è
ã®å€ãã«çè§£ãããŠãããšããããšã¯ïŒåå è
ãç¹æ®ãªç¶æ
ã«ãªãéãïŒåžžã«æé
ãã ïŒãšããããšã§ãïŒãŸãïŒå€ãã®äŒç€Ÿãåãããã«çè§£ããŠããïŒç«¶äºãæ¿ãããšããããšã§ãïŒåºæ¬çã«ã¯ïŒç Žå£çãªäŸ¡å€ãåŸãããªãïŒçè§£ã§ããªããŸãŸå®è¡ã«ç§»ããææ³ïŒäŸãã°ããŒã¿é§åãããŒã¹ãšããäœå¶ïŒãå¿
èŠãªã®ã ãšæããŸãïŒããã«è¶³ããªãIPã©ã³ãã¹ã±ãŒãã«ããåæè
ãåŸãã€ã³ãµã€ãã§ãïŒãã®åæè
ãåŸãã€ã³ãµã€ãã¯ãã®åæè
ãæå±ããäºæ¥ã®äºæ¥åäœã®éå£ã«ã§ããã°ïŒçè§£ããå®è¡ãããããã§ãããïŒïŒïŒããŒã¿é§åã§ã¯ããŒã¿ã®æ§è³ªã«ããããŒã¿ãäžè¶³ããããšããããŸãïŒãã®å Žåã¯äººã®åæè
ã®ç¥èãå©çšã§ããIPã©ã³ãã¹ã±ãŒããªã©ãããæå¹ã«ã¯ãªãã§ãããïŒäžè¶³ããŠããããŒã¿ãç¥èã§è£ããªããªãïŒåŸè
ã®æå³ã¯ããŸããªãã§ãããïŒïŒïŒå人çã«ã¯ïŒããŒã¿é§åã¯çè§£ã䌎ããé²ãããšãæå€§ã®å¹æãçºããã®ã§ã¯ãšèãã¯ãããŠããŸãïŒãã©ãã¯ããã¯ã¹ã ãããã䟡å€ããããšãïŒããªã極è«ã§ããã©ãã§ããããïŒå
ã«è¿°ã¹ãä¿¡ããã«ãé¢é£ããŸããïŒïŒ
ïŒããŒã¿ãµã€ãšã³ã¹åã³åæãšã説æãã¯ïŒè¿å¹ŽãŸããŸãé¢ããŠè¡ã£ãŠãããšæããŠããŸãïŒåæè
ãæçšãªã¢ãã«ãçæã§ãããããã€ã³ãµã€ããåŸããããšããŠãïŒãã®ã€ã³ãµã€ããïŒå¥ã®ç¥èãå
ã«èª¬æãåãããã®ã®ç¥èã«åããïŒèª¬æããããšã¯å¹Žã
é£ãããªã£ãŠãããšèªèããŠããŸãïŒåæã¯äºæ¥ã¬ãã«ãå人ã¬ãã«ã§æ¶è²»ãããšå®è¡ã«ç¹ãããããããè¯ãããã«æããŸãïŒ
ïŒè»žã«æå³ããªããã°åæãããªãïŒãã¬ãŒã ã«åŸããªããªãåæãããªãïŒãšèšããããªã²ãŒãå€ææš©éè
ããããããªäŒç€Ÿãéšéã¯ïŒå€åæ°žé ã«ç Žå£çãªäŸ¡å€ãçã¿åºãããã«å®è¡ããããšãã§ããªãã§ãããïŒå人çã«ã¯ãã®ãããªäŒç€Ÿã«éã£ãŠïŒäœæ
èªåã®äŒç€Ÿããã€ãããŒã·ã§ã³ãçãŸããªãã®ãïŒãªã©ãšèšã£ãŠããå°è±¡ããããŸãïŒå€æè
ã®èœåãäžãããïŒæåã«å€æããªãã¢ãžã£ã€ã«çãªèãæ¹ãåãããšã¯ïŒå¿
é ã«æããŸãïŒåè
ã¯äºæž¬ã容æã§ããæ¥åã«éå®ããªãéãïŒé£ããã§ãããïŒäºæž¬ã容æã§ããæ¥å以å€ã§ã¯ïŒåŸè
ãçŸå®çã§ãããïŒïŒ
ïŒå人çã«ã¯åæã»ã°ã¡ã³ããŒã·ã§ã³ã¯ããããããŸããïŒã»ã°ã¡ã³ãã¯ããŒã¿ã®ãã¡ã€ã³ã«äŸåãããã®ã§ãããšããïŒæ±ãããŒã¿ãå¿
ããããã®ãã¡ã€ã³ã®ç¯çã§ãããŸãã¯ãã®ãã¡ã€ã³ã®ç¯çã®ãŸãŸã§ããä¿èšŒã¯ãªãããã§ãïŒãŸãïŒæ¢åã®è©äŸ¡è»žã䜿ã£ãŠããŠã¯ïŒã¬ãããªãŒã·ã£ã³ã¯ãã€ãŸã§ãã¬ãããªãŒã·ã£ã³ã®ãŸãŸã§ãïŒè»žã»èŠç¹ã»èŠåº§ãå€ããã°ïŒã¬ãããªãŒã·ã£ã³ããã«ãŒãªãŒã·ã£ã³ã§ããããã«èªç¥ã§ããããã§ãããïŒ
ïŒChatGPTã«ãããšããã³ãã©ã³ãã¹ã±ãŒããšã¯ã€ãã®ããã§ã

the process of mapping out and analyzing the patent landscape
provide useful insights
This information can be useful for companies and organizations that are considering entering the market or developing new technologies in the field.
ChatGPTããã®èª¬æãäžçªåŠ¥åœã«æããŸãïŒ
ïŒGPT-4ã®èŠè§£
- ãç¥è²¡ãçµå¶ãšäžäœåããããšã§ã確ãã«çµå¶ã®èªèãããã«ããã¯ãšãªããäºæ¥ãçµå¶ã®æç€ºåŸ ã¡ãšãªãå±éºæ§ãååšããŸãããã®åé¡ã¯ãçµç¹ã®æææ±ºå®ãé ããªã£ãããåžå Žã§ã®æ©ååãäœäžãããããããšã§ãç«¶äºåã«æªåœ±é¿ãäžããå¯èœæ§ããããŸãããã®ãªã¹ã¯ã«å¯ŸåŠããããã«ã¯ã以äžã®ãããªå¯ŸçãèããããŸããã
- ãçµå¶ãšç¥è²¡ã®äžäœåã¯ãé©åãªãã©ã³ã¹ãšå¯ŸçãæŽã£ãŠããã°ãç«¶äºåã®åäžã«å¯äžããŸããããããããã«äŒŽããªã¹ã¯ã課é¡ã«ã泚æãæããé©åãªå¯Ÿå¿çãè¬ããããšãéèŠã§ããã
ïŒããããªãïŒã¯ã¬ã€ãžãŒã»ããŒã¿ãµã€ãšã³ã¹ïŒ
https://qiita.com/KanNishida/items/1a4c809cd63a5a82c3e7
ããããŒãžã£ãŒãã¡ã¯ãããŒã¿ããªãã³ãªæææ±ºå®ããããããšãããïŒå®éã®ãšãã圌ããæ¬²ããã®ã¯ãæææ±ºå®ããªãã³ãªããŒã¿ãã ããšããèšèŒã«ã¯æããç¬ã£ãŠããŸããŸããïŒ
çµå¶ããŒã¹ã®IPã©ã³ãã¹ã±ãŒãã¯ã決å®ãåŸæŒãããŠæ¬²ãããšããçµå¶èŠæãæºãããã®ããšãªããã¡ã§ïŒæ¬æ¥ãã£ãšæ·±ãæ¢çŽ¢ã§ããã¯ãã§ãããšãããã®ãã£ã³ã¹ãéããŠãããããªæ°ããããã§ãããïŒ
ïŒç¹èš±æ
å ±ã®é²æ©çãªæŽ»çšæ¹æ³
https://aamtjapio.com/kenkyu/files/symposium2023/shobayashi.pdf
ïŒIPã©ã³ãã¹ã±ãŒãã®ã¹ã¹ã¡ãæåææ ªåŒäŒç€Ÿã
https://www.jpo.go.jp/news/koho/kohoshi/vol49/01_page3.html
ãå
·äœçã«ã¯ãå
ã
åŒç€Ÿãè²·åããäŒç€Ÿãæ°Žã®æ®ºèã«äœ¿ã深玫å€ç·ïŒ¬ïŒ¥ïŒ€ãæ±ã£ãŠããŸãããããããè¡šé¢æ®ºèã«ãã©ãã·ã¥ã¢ããããŠææç察çã«åœ¹ç«ãŠãããªããèãããã§ããããã§ãŸãè¡šé¢æ®ºèã«ã€ããŠã®ç¹èš±ããããäœã£ããšãããç«¶åä»ç€Ÿã¯ãŸã ããã«ç®ãåããŠããªãããšãåãããŸãããæ¬¡ã«ãã®åžå Žã«å
¥ã£ãŠããã®ã«äžç·ã«çµããç
§æäŒç€Ÿãæ¢ããããç
§ææ¥çã®äŒæ¥çŸ€ãç¹èš±ãããã§ç¢ºèªããã·ããžãŒå¹æãèŠèŸŒããäŒç€ŸãèŠåºãããã§ããã
ãæŽå¯ã®ææãšããŠäžçã®ã€ãããŒã·ã§ã³ãçœåŒããGAFAã®ç¹èš±è§£æããŒã¿ãèªç€Ÿã®ç¹èš±ããããæäŸããŸããããã®çµæãæ°ä»ãã®é£éãèµ·ããŠãïŒïŒïŒïŒå¹Žã®äžçã«è²¢ç®ããããã®ã¢ã€ãã¢ãïŒïŒïŒãåºãŠããŸããããããã¯å
šç€Ÿã«å
±æããŠãããåçŸå Žã§æ°äºæ¥ãåµãããã«åœ¹ç«ã€ã¯ãã§ããã

åèïŒChatGPTã«æåæã®æè¡åéãïŒïŒä»¶æããŠãããïŒã¯ã©ã¹ã¿ãšæ¯èŒããŠã¿ãïŒ
| 該åœã¯ã©ã¹ã¿ | ChatGPTã®åç |
|---|---|
| âã¯ã©ã¹ã¿ïŒé»è²ïŒã | ïŒãããªããŒè£œåæè¡ïŒæåæã¯ã髿©èœããªããŒã®è£œé æè¡ãä¿æããŠãããèªåè»éšåãå»çææãå åŠææãªã©å€ãã®åéã«å¿çšãããŠããŸãã |
| ãã¯ã©ã¹ã¿ïŒéïŒã | ïŒãé»åæææè¡ïŒæåæã¯ãé»åææã®è£œé æè¡ãä¿æããŠãããæ¶²æ¶ãã£ã¹ãã¬ã€ãåå°äœãªã©ã®é»åéšåã«çšããããŠããŸãã |
| ïŒãååŠãã©ã³ãæè¡ïŒæåæã¯ãååŠãã©ã³ãã®èšèšã»å»ºèšã»é転ã«é¢ããæè¡ãä¿æããŠãããååŠå·¥å Žã®å»ºèšã»éå¶ã«ãããŠé«ãè©äŸ¡ãåããŠããŸãã | |
| âã¯ã©ã¹ã¿ïŒç·ãã | ïŒãååŠåææè¡ïŒæåæã¯ãç¬èªã®ååŠåææè¡ãæã¡ã倿§ãªååŠç©è³ªã®åæã«å¿çšãããŠããŸãã |
| âã¯ã©ã¹ã¿ïŒèµ€ïŒã | ïŒããã£ã«ã æè¡ïŒæåæã¯ã髿©èœãã£ã«ã ã®è£œé æè¡ãä¿æããŠãããé»åéšåãå»çææãå åŠææãªã©ã«å¿çšãããŠããŸãã |
| ãã¯ã©ã¹ã¿ïŒéïŒ | ïŒã粟å¯å å·¥æè¡ïŒæåæã¯ãé«ç²ŸåºŠå å·¥æè¡ãæã¡ãåå°äœè£œé è£ çœ®ãå»çæ©åšã®è£œé ã«å¿çšãããŠããŸãã |
| ïŒãã»ã©ããã¯ã¹æè¡ïŒæåæã¯ã髿©èœã»ã©ããã¯ã¹ã®è£œé æè¡ãä¿æããŠãããå»çææãèªåè»éšåãåå°äœãªã©å€ãã®åéã«å¿çšãããŠããŸãã | |
| âã¯ã©ã¹ã¿ïŒé»è²ïŒ | ïŒã塿æè¡ïŒæåæã¯ã髿©èœå¡æã®è£œé æè¡ãä¿æããŠããã建ç¯çšå¡æãèªåè»å¡æãªã©ã«äœ¿çšãããŠããŸãã |
| âã¯ã©ã¹ã¿ïŒèµ€ | ïŒããã€ãªãã¯ãããžãŒïŒæåæã¯ããã€ãªãã¯ãããžãŒã«é¢ããæè¡ãä¿æããŠãããå»çææããã€ãªããã»ã¹çšæ©åšãªã©ã«å¿çšãããŠããŸãã |
| ïŒïŒãã·ã¹ãã éçºæè¡ïŒæåæã¯ãé«åºŠãªã·ã¹ãã éçºæè¡ãæã¡ãæ¥åã·ã¹ãã ãæ å ±ã·ã¹ãã ãªã©å€æ§ãªã·ã¹ãã ã®éçºã«è²¢ç®ããŠããŸãã | |
| ã¯ã©ã¹ã¿ïŒ | |
| ã¯ã©ã¹ã¿ïŒ |
ã³ã¡ã³ãïŒ
次ã®ã¯ã©ã¹ã¿ã«è©²åœããChatGPTã®åçã¯èŠã€ãããªãã£ãïŒ
ã¯ã©ã¹ã¿ïŒãç¹ç¶ã»äžç¹åž
ã¯ã©ã¹ã¿ïŒãäœå®
ã»å»ºæ
ã¯ã©ã¹ã¿ïŒã¯æç¢ºãªæ°ã®ç¹èš±çŸ€ãããïŒã¯ã©ã¹ã¿ã®äœçœ®ãããããšåå°äœãšèã®éã«ããæè¡ã§ããïŒå
åŠææãå»è¬éšåã«ç¹ç¶æè¡ã䜿ãããŠããã®ã ããïŒåçïŒïŒïŒïŒïŒïŒïŒã該åœããŠããã®ãããããªãïŒ
ã¯ã©ã¹ã¿ïŒã¯æç¢ºãªæ°ã®ç¹èš±çŸ€ãããïŒä»ã®ã¯ã©ã¹ã¿ãšã¯é¢ããŠããé¡äŒŒã¯ã©ã¹ã¿ã¯èŠã€ãããªãïŒChatGPTã¯ãã®æè¡ãé¢ããåéãæç€ºã§ããªãã£ãããã ïŒãããŠããã°ïŒåçïŒãè¿ãã®ãããããªãïŒ
次ã®ChatGPTã®åçã«è©²åœããã¯ã©ã¹ã¿ã¯èŠã€ãããªãã£ãïŒ
ïŒãååŠãã©ã³ãæè¡
ïŒãã»ã©ããã¯ã¹æè¡
ïŒïŒãã·ã¹ãã éçºæè¡
åçïŒã¯æ£ããåçã§ã¯ããããã ïŒãã®åéã§ã¯ç¹èš±åºé¡ãããŠããªããšããããšãããããªãïŒ
åçïŒã¯æ£ããåçã§ã¯ããããã ïŒã¯ã©ã¹ã¿ïŒã®äœå®
ïŒå»ºæãªã©ã該åœããã®ãããããªãïŒ
åçïŒïŒã¯æ£ããåçãäžæã§ãã£ãïŒ
ç·ããŠïŒäœå®
ã«é¢ããæè¡åéã®ååšãChatGPTãåçã§ããªãã£ãããšä»¥å€ã¯ïŒãªããªãã«é©åãªåçã§ã¯ãã£ãïŒ
ã¯ã©ã¹ã¿ã«åç§°ãã€ããã«ãããïŒChatGPTã䜿çšããŠãé¢çœããããããªãïŒ
åºé¡ãããŠããªãé åãïŒãã®ããã«æ¢ããŠãè¯ãã®ãããããªãïŒ
æåæã®ç¹èš±åæãããŠã¿ã
ã»æåæã®ç¹èš±å
šãŠã«é¢ãïŒä¿¯ç°å³ãçšããè§£æãäžéãããŠã¿ãŸããïŒ
â»ããªãç®çã¯æ°èŠäºæ¥æ¢çŽ¢ãšããŸããïŒæ¡ä»¶ãšããŠïŒæåæã®ç¹èš±ã®ã¿ããèŠããç¯å²ã ãã®ç°¡æçïŒãšããŸããïŒæ¢çŽ¢ããæ°èŠäºæ¥ã¯ãã§ã«ä»ç€Ÿå®æœæžã¿ãããããŸããïŒæ£åŒã«ã¯ãã¡ããèæ
®ããè§£æãããããšã«ãªããŸãïŒ
ã»ã»æåæã®äžå³æç€ºã®7åéãã¹ãŠïŒä¿¯ç°å³ã§åçŸã§ããŸããïŒ
ã»ã»èŠçŽ ãããã¯ãŒã¯ãçšããããšã«ããïŒæµ
ãèŠåº§ããïŒæåæã®3ã€ã®ã³ã¢é åãæœåºã§ããŸããïŒ
ã»ã»ä¿¯ç°å³ãæ§é çç¥èãšäºæ¥ãã§ãŒãºã«é¢ããç¥èãšæè¡å¯¿åœã«é¢ããç¥èãšãå
ã«è§£æããããšã«ããïŒæåæã®äžæçµå¶èšç»ã®ãæŠç¥çè²æãé åãïŒãæ°èŠäºæ¥åè£ïŒããšããŠæœåºã§ããŸããïŒ
ã»ã»ãã®ä»åéã«é¢ããèŠè§£ãïŒæåæã®äžæçµå¶èšç»ãšäžèŽããŠããŸãïŒ
ã2023幎4æ11æ¥ãäžæçµå¶èšç» 2024 ~Be a Trailblazer~ ã®é²æç¶æ³ã«ã€ããŠ
ãhttps://www.asahi-kasei.com/jp/ir/library/initiative/pdf/230411jpn.pdf
ã
ã»ã»ããã«ïŒæåæã®äžæçµå¶èšç»ã«èšèŒã®ãªãïŒææãããªãæ°èŠäºæ¥åè£ïŒããçæã§ããŸããïŒ
ãã®æ°èŠäºæ¥åè£ïŒãïŒçæAIã«çºæ³æ³ãã¬ãŒã ã¯ãŒã¯ã䜿çšããããšã«ãããå
·äœåããã§ããŸããïŒãã®æ°èŠäºæ¥åè£ïŒã®å
·äœäŸãšé¡äŒŒããèª²é¡æè¡ã¯ïŒãã³ã·ã«ããã¢å€§åŠïŒãšãžã³ãã©å€§åŠïŒç±³è»ïŒãã³ãã£ãŒãéçºããŠããŸããïŒãã®æ°èŠäºæ¥åè£ïŒã¯ïŒæš©å©ã¯ååã«åããããïŒåžå Žã¯ããïŒ1-100ã§ããæé©ãªææã§ïŒæåæãèªç€Ÿå
ã§ããŸãã¯å€å°ã®æè¡ãå°å
¥ããã°ïŒå®çŸå¯èœã ããïŒãšç¢ºèªãããŸããïŒ
ãªãïŒå®æœã«é¢ããŠã¯æ±ã¬ã®moon rakersã§ã®å®æœãããåŸããšããã§ãããïŒ
ã»äžèšå³ãããŒã¿ãèæ
®ããã€ãããŒã·ã§ã³ã®äžäŸïŒå·®ç°åãç®æãå ŽåïŒãã¯ïŒãã®ããã«ã宿œã§ããŸãïŒ
ãŸãïŒä»ç€Ÿã®æ°èŠäºæ¥ã®åŠ¥åœæ§ãæ€èšŒããã«ä¿¯ç°å³ã®è§£æãçšãããŠã®ã¯ååããåŸããšæããŸããïŒ
ãªãïŒãã®åéã®äŒæ¥ã«æå±ããŠããªããããã®åéã®è©³çްãŸã§ç¥ãããšã¯ã§ããªããæè³å®¶ããã³ãã£ãŒã倧åŠãããïŒããããIPã©ã³ãã¹ã±ãŒãïŒãã®ãããªä¿¯ç°è§£æïŒãããå¿
èŠãããïŒãšå人çã«ã¯ç¢ºä¿¡ããŠããŸãïŒ
䜿çšäŸïŒïŒæåž«ããåŠç¿ãšæåž«ãªãåŠç¿ã®èå
ã¯ã©ã¹ã¿ãªã³ã°ããåå¥èªèã«èžã¿èŸŒã¿ãŸãïŒ
æåž«ããåŠç¿ã¯ïŒç²ŸåºŠãé«ãã§ãåå¥èªèå¯èœæ§ãé«ããåŠç¿ããŒã¿ã®ã©ãã«ã«äŸåãããæ§é ã倱ãããŸãïŒæåž«ãªãåŠç¿ã¯ç²ŸåºŠã¯ããã»ã©é«ããªãããšãå€ããã©ãã«ã«äŸåããªãæ§é ããããããŸãïŒãã®ïŒã€ã¯çžè£çã§ãïŒ
æåž«ããåŠç¿ã®çµæãæã·ã³ãã«ã«ãããŸãã¯è²ãå€ãããªã©ããŠïŒæåž«ãªãåŠç¿ã§ããtfidf embeddings cluster visäžã«è¡šç€ºããã°ïŒæåž«ããåŠç¿ã®çµæã«æåž«ãªãåŠç¿ã®ç¥èŠãå å³ããïŒãããªãã€ã³ãµã€ããåŸãããã§ãããïŒ

ã¯ã©ã¹ã¿ã®å¢çãã€æåž«ããã§éèŠãšç€ºãããç¹èš±çŸ€ãªã©ã¯ïŒå®éãšãŠãããããã§ãïŒ
ïŒããšããšã¯AIã·ã¹ãã ã ãã§ã¯åŸãããªãã€ã³ãµã€ããåŸãç®çåã³AIã·ã¹ãã ãèŠèœãšããããããªãç¹èš±ãã«ããŒããç®çã§ïŒAIã·ã¹ãã ã«çµã¿èŸŒã圢ã§ïŒtfidf embeddings cluster visã䜿ã£ãŠããŸããïŒ

retrieverãreaderïœ
ïŒæåž«ãããšæåž«ãªãã®çµã¿åããã¯ïŒdisentanglementã®äžéšã ããã§ãïŒ
ã ,
䜿çšäŸïŒïŒæ€çŽ¢æè²ã»SDIèŠçŽããªã©çžå¯Ÿæ¯èŒã«
ã¡ã³ã¿ãŒã®æ€çŽ¢æ¯éå£ãšæ°äººã®æ€çŽ¢æ¯éå£ãæ··ããŠtfidf embeddings cluster visã«ããïŒãã®éè€ãšå·®ãèŠãããšã§ïŒã©ã®ç¯å²ãè¶³ããªãã£ããšèšãããïŒãäºãã«èªèãããããŸãçŽåŸæãé«ããŠç¢ºãããŠããããšãã§ãããŸãïŒ
æ€çŽ¢å¯Ÿè±¡ãèšå®ãããŠãããªãã°ïŒãããæ¯éå£ã®éå¿ããã©ãã»ã©é¢ããŠããã確èªããããšã§ïŒæ¯éå£ã®ç¯å²ãé©åã§ãã£ããã©ãã確èªããããšãã§ããŸãïŒãŸãïŒçµãèŸŒã¿æ©èœãçšããããšã§ïŒæ¯éå£ãã©ãéå®ããã°ããã£ãã®ãçè§£ããããšãã§ããŸãïŒ
ïŒèŠã€ãã£ãèŠã€ãããªãã£ãïŒæ¹åæ§ãæããªãå
šäœãšããŠã®recallãprecisionãã©ãã ã£ãïŒæ°ä»¶ç¢ºèªããŠã©ãã ã£ãïŒã ãã§ã¯ïŒååžãšæ¹åãšãèæ
®ããååãªçè§£ãã§ããŸããïŒ
ãŸãïŒä»åŸïŒèª²é¡ãäžå¿ã«èª¿æ»ããããšãå€ããªãããšæãããŸããïŒãã®éã«ã¯èª¿æ»èœåå·®ãçãŸãããããã®ãšæãããŸãïŒçããæç¢ºã«ãªããïŒè©äŸ¡ããã«ãããã®ã§ãïŒ
ãã®ãããªããŒã«ãäžã€ã¯æã£ãŠãããã»ããè¯ãã§ãããïŒ
äœãããšã¯é£ããããšã§ããããŸããïŒ
ïŒããŒã¯ãŒãã®ã¿ã§æ€çŽ¢ãããšãã®ããã«è€æ°ã®åéãæ··åããæ¯éå£ãšãªããªã©ç¬æã«å
·äœçã«çè§£ã§ãããããŠïŒçµæ§é¢çœãã§ãïŒããŒã¯ãŒãã«å¯Ÿãããªã³ã¯ããã©ãã°ïŒãã®ããŒã¯ãŒããã©ãã«ã©ã®ããã圱é¿ãããèŠãããšãã§ããŸãïŒ

䜿çšäŸïŒïŒè£å€å€æ±ºäŸã®äºç¹åäœæŠå¿µæ€çŽ¢
詊è¡äžïŒ
ïŒããŒã¿æºåçµäºïŒè©Šè¡ïŒã§ããŠã¯ããã®ã ãïŒããã¹ããäºç¹åäœã§ç¢ºå®ã«åºåãããšãã§ããªãã£ãïŒæ¹è¯äžïŒïŒ

ïŒå€æ±ºäŸã«ãããŠçšããããçšèªã¯èŠæ ŒåãããŠããããšãå€ãåæ£è¡šçŸãçšããã¡ãªããã¯ããŸããªãããã§ãïŒãŸãïŒã¯ã©ã¹ã¿ãªã³ã°ããã»ã°ã¡ã³ããŒã·ã§ã³ãæçšãšæããŸãïŒãã£ãŠïŒçŽ çŽã«Darts-ip https://www.darts-ip.com/ja/ip-cases-database/ ã䜿ã£ãã»ããè¯ããšã¯æããŸãïŒ
ã»å€§èŠæš¡èšèªã¢ãã«ãçšããæ¥æ¬èªå€æ±ºæžã®èªåèŠçŽ
âæ°ä¿åœ°äºº, è
åè£å€ª, å±±ç°å¯ç« , 埳氞å¥äŒž (æ±å·¥å€§)
ã202402
ã https://www.anlp.jp/proceedings/annual_meeting/2024/pdf_dir/E4-4.pdf
䜿çšäŸïŒïŒããŒã¿é§åãã©ãã°ã»ãªããžã·ã§ãã³ã°ã«ããåµè¬ã®ãã³ããšããŠ
ã¯ã©ã¹ã¿ãªã³ã°ãã»ã°ã¡ã³ããŒã·ã§ã³ã®ããã«äœ¿çšããäŸããã³ã¯ã©ã¹ã¿ãªã³ã°ããåå¥èªèã«èžã¿èŸŒãã äŸïŒ
å»è¬ååè£ååç©çŸ€xïŒçžäºäœçšã¿ã³ãã¯è³ªãããã¯ãŒã¯yïŒçŸæ£çŸ€zãã€ãªãïŒãã®ãã®æªç¥ã®ãªã³ã¯ãäºæž¬ãããªããžã·ã§ãã³ã°ææ³ããããŸãïŒïŒ https://pubmed.ncbi.nlm.nih.gov/26580494/ ãªã©ïŒ
äŸãã°ïŒå»è¬ååè£ååç©ãé©åœãªææ®µã§ãã¯ãã«åãtfidf embeddings cluster visã«ããïŒçžäºäœçšã¿ã³ãã¯è³ªãããã¯ãŒã¯yããäžå¿æ§ãèæ
®ãåŸããããã¿ã³ãã¯è³ªé¡äŒŒã¯ã©ã¹ã¿ãããšã«è²ä»ããããå»è¬ååè£ååç©-ã¿ã³ãã¯è³ªå³ããäœæãïŒãã€çŸæ£ããåŸããããçŸæ£é¡äŒŒã¯ã©ã¹ã¿ãããšã«è²ä»ãããå»è¬ååè£ååç©-çŸæ£å³ããäœæãïŒåæ¹ã®å³ãæ¯èŒããããšã§ïŒæªç¥ã®ãªã³ã¯ãäºæž¬ãããªããžã·ã§ãã³ã°ææ³ãšäŒŒãçµæãïŒèŠèŠçã«è¡ãããšãã§ããŸãïŒ
ïŒããŒã¿ã®ã¿ã§ã¯æ å ±ãäžè¶³ããŠãããã¥ãŒã¹ããªãã¯ãªæ å ±è£å®ã«å€§ããªäŸ¡å€ãæ®ã£ãŠããããã«ããäºæž¬ãå ãã䟡å€ãããå Žåãªã©ïŒæ¯èŒç宿§çãªäºæž¬ãå¿ èŠã§ããå Žåã«ã¯ïŒãã®ãããªå³ç€ºææ³ãæçšã§ãããïŒãªããžã·ã§ãã³ã°åµè¬ã¯ããŒã¿ã®ã¿ããè¡ã£ãã»ããè¯ããšã¯æããŸããïŒ
䜿çšäŸïŒïŒã³ãã¥ãã±ãŒã·ã§ã³èœå代æ¿ã»ãã¡ã·ãªããŒã·ã§ã³ããŒã«ãšããŠïŒ
ã³ãã¥ãã±ãŒã·ã§ã³ã®åé¡ãšããŠïŒã³ãã¥ãã±ãŒã·ã§ã³ãèœåãšããŠæããããšã«ããã¬ãã«ãã£ããã®åé¡ãäºåç¥èãå
±éããªãå Žåã®äŒéã»èªç¥ïŒæè²ïŒã³ã¹ãã®åé¡ãªã©ãããšããã§ããïŒã³ãã¥ãã±ãŒã·ã§ã³ã®çµæãã©ãè©äŸ¡ããŸãšããããšããåé¡ããããŸãïŒ
ããããŸãšããå°é£ã§ããæå€§ã®åå ã¯ïŒãããè¡ã£ãŠããããšã§ãããïŒ
ããã®äž»èгãæé€ãã¹ãã§ãããªããªãã§ããŸããïŒæçµçã«ã¯æ©æ¢°ããã¹ãŠè¡ãã¹ãã§ãã®æè¡ã¯ãã§ã«ãããšèªèããŠããŸããããã¯ãããŠãããŠïŒæ¬ããŒã«ã§ããçšåºŠã®å®¢èгåãè¡ãããšãé¢çœãããšæããŸãïŒ
è°è«ã®è§£æâè°è«ã®å¯èŠåâè°è«ãžã®ãã£ãŒãããã¯â
ã®ã«ãŒãã§ãïŒæ¬äŒŒçã«ã¯æ©æ¢°åãããã¡ã·ãªããŒã¿ãŒãšãªãã§ããããïŒ
è«è¿°æ§é è§£æã«åŸãè°è«ã®æŽçææ³ããããšã¯æããŸããïŒããå æã¯åãã§ããããæŽçãããè«çãã«ã¯ããããææ³ãããïŒãå人ãçè§£ã§ãããè«çãã«ãæåç倿§æ§ãªã©å€æ§æ§ãããããšããïŒäžã€ã®è«çã§ãŸãšããããšèªäœãåé¡ããšæããŸãïŒçžé¢ãŸãã¯å æã瀺ãã«çãããšããææ®µã®ã»ããæå¹ãšãªãããšãå€ãã§ãããïŒ
ïŒãªã¢ã«ã¿ã€ã ã«åãå¿
èŠããããã®ç¹å·¥å€«ãå¿
èŠããšæããŸãïŒïŒ
ïŒãã¬ã€ã³ã¹ããŒãã³ã°ãã³ã¹ãã®é«ãã³ãã¥ãã±ãŒã·ã§ã³ã§ãïŒ
æ¬ããŒã«ãšãç°ãªããŸããïŒElicitã®ãç ç©¶ã®è³ªåããã¬ã€ã³ã¹ããŒãã³ã°ãããã¿ã¹ã¯ãçšããã°ïŒ
èšèªã¢ãã«ã®ç¥èãçšãããã¬ã€ã³ã¹ããŒãã³ã°ãã§ããŸãïŒ
https://ide.elicit.org/run/JqRpiAJMnNacB7wvq
èªåãç¥ã£ãŠããããšããæ°ã¥ãããšãã§ããªãïŒãã¬ã€ã³ã¹ããŒãã³ã°ãšã¯ç¥èãããã¯ãŒã¯ã圢æããŠãããæèäžã«æµ®ããã§ããªãããŒãããšããžãšããèªåã®æé»ç¥ãæèäžã«äžããå人çãªã¢ãŠããããäœæ¥ã«éããªãïŒãšåæãšããã°ïŒãã®ãããªä»äººã®ãªãœãŒã¹ãæ¶è²»ããããªãããŒã«ãæçšãšãããã§ãããïŒ
ïŒæè¿æç®èª¿æ»ããéã«ã¯ïŒElicitâconnected papersâtfidf cluster visâãããã¯ãŒã¯åã®æµãã§ããŒã«ã䜿ã£ãŠããããšãå€ãã§ãïŒ
ïŒã°ãŒã°ã«åŒä»äºè¡ãããã¬ã¹ããã嫿ªããçç±ãGoogle SPRINT
https://diamond.jp/articles/-/170687
èªãã®ç¥èãããã¯ãŒã¯ã«å«ãŸããªããã®ã¯æ°ã¥ãããšãã§ããªããšãããªãã°ïŒããã¬ã¹ãã¯ããå人çã«ããçååããŠãè¡ãã¹ãã ãšãã§ããã§ãããïŒ
ã»perplexity.ai
https://www.perplexity.ai/
elicitãè¯ãã§ããïŒãã¡ãã¯æŽã«äœ¿ããããã§ãïŒ
ãã ãïŒelictãšç°ãªãïŒæç®ãè§£éããäžã§æç€ºããæœè±¡åèŠçŽããããªã£ãŠããããã§ãïŒãã®ããïŒèšèªã¢ãã«ã®åŒ±ç¹ãåããŠããŸãïŒãã§ã«æ§é åããæç§æžã«èŒããããªãã³ã³ã»ã³ãµã¹ã®åœ¢æãããŠããç¥èãã«ã€ããŠã¯ããªãæ£ç¢ºã«åŒçšããŠããŸããïŒã³ã³ã»ã³ãµã¹ã®åœ¢æãããŠããªãç¥èã®ãšããžåšèŸºã®ç¥èïŒåŸãŠããŠãæå
端ã®ïŒè°è«ã®ããç¥èãïŒã«ã€ããŠã¯åŒ±ãïŒæ£ããããã«åŒçšããŠãããªãã誀ã£ãŠããããšãå€ãã§ãïŒ
ãšã¯ããïŒããã¯åŒçšå
ã確èªããã°ããã ããªã®ã§èŽåœçãªåé¡ã§ã¯ãããŸããïŒ
ããã£ãŠïŒèª€ããçãå¿
èŠãæãããšã¯ïŒæç®ãæ·±èªã¿ããããšãå°ãïŒé©åãªç¥è圢æãå°ãã§ãããïŒæ¹å€çæèã«èªåçã«ç¯æ£ãããŸãïŒ
perplexityãçšããããšã§ïŒç¥èã®å£æã¡ããã¬ã€ã³ã¹ããŒãã³ã°ãéåžžã«å¹ççã«ã§ããããã«ãªãã§ãããïŒ
âã³ãã¥ãã±ãŒã·ã§ã³ã®ãŸãšãããŒã«ïŒãã¬ã€ã³ã¹ããŒãã³ã°ããŒã«ïŒãã¡ã·ãªããŒã·ã§ã³ããŒã«ïŒ
ããã¹ãŠïŒGPT-4ã§è¶³ãããã§ãïŒ
䜿çšäŸïŒïŒãã¯ãæ¢çŽ¢ã»è§£æïŒäŸãã°ïŒãã¯ãã«ç©ºéãä»»æã®èŠçŽ ãããªãç¥èãããã¯ãŒã¯ã«å€æïŒãã€ãžã¢ã³ãããã¯ãŒã¯ã®äœ¿çšïŒãã€ãºã®èåïŒ
ãã¯ãæ¢çŽ¢ã ãã§ãªããã¯ãæ¢çŽ¢ãåæã«è¡ãïŒã»ã°ã¡ã³ããŒã·ã§ã³ïŒç¥èãããã¯ãŒã¯åãšãã€ãžã¢ã³ãããã¯ãŒã¯ã®äœ¿çšïŒæ¡å€§å§çž®ã®åµé çæ€èšãªã©ãããã®ãè¯ãã§ãããïŒ
ïŒãã¯ãæ¢çŽ¢ããããšããçç±ã¯ïŒãã¯ãæ¢çŽ¢ã®ã¿ã§ã¯so-whatãšãªãããã§ãïŒïŒ
ä»»æã®èª²é¡ãšä»»æã®è§£æ±ºææ®µãïŒã¯ã©ã¹ã¿ãªã³ã°çµæãæ¢çŽ¢æå³ããèšå®
â
ä»»æã®èª²é¡ããšã«é¡äŒŒåºŠãããšãïŒããããç¹åŸŽåäœæ
â
ä»»æã®è§£æ±ºææ®µããšã«é¡äŒŒåºŠãããšãïŒããããç¹åŸŽåäœæ
â
課é¡ããšã«å
šè§£æ±ºææ®µãšã®å
±èµ·ã»çžé¢ãæ±ãã
â
å
±èµ·ãå ææšè«ãªã©ãããã¯ãŒã¯æ¢çŽ¢ææ³ãåºã«ããŠïŒæ¯éå£ã®ä»»æèŠçŽ ããããã¯ãŒã¯å



ç¹çŸ€ãã圢æããã空éã¯å
ã®ç¹çŸ€ä»¥äžã®æ
å ±ãå«ã¿ïŒããããããã¯ãŒã¯ãšããŠåãåºãããšãã§ããïŒãšããã€ã¡ãŒãžã§ãïŒããŒã¯ãŒãã®å
±èµ·ãããã¯ãŒã¯ã¯ç¹çŸ€ããããã¯ãŒã¯ã«ãããã®ã§ããïŒãã®ãããã¯ãŒã¯ãšã¯ããããç°ãªããŸãïŒåè
ã¯äžã€ã®ç¥èããåãåºããŸãããïŒåŸè
ã¯åè
å«ãè€æ°ã®ç¥èãåãåºããŸãïŒèŠãç®äžç·ã§ããã©ãïŒ
ïŒäŸ¡å€ããããã¯ãŒã¯ã®ã©ãã«ããããšããèŠç¹ã§èŠãŠãïŒéåžžã«è峿·±ãã§ãïŒå°ãªããšãåªä»äžå¿ã«ã¯ãªãã£ã«ã«ãªäŸ¡å€ãããäŸã¯èŠãŠããŸããïŒãšããžã«äŸ¡å€ãããç¹ã¯å
ã®ã¢ã€ããããã«äŒŒãŠããŸãïŒ
ïŒãšããžã¯ããŒãéã®é¢ä¿ã衚ãïŒãã®é¢ä¿ã«ã¯å
±èµ·ã»å æã»æ§é åãããé¢ä¿ãªã©æ§ã
ãªãã®ãèããããŸããïŒããšç¹èš±ã«ãããŠã¯ïŒæ§é åãããé¢ä¿ã¯å¿
èŠãªãããšãå€ãããšæããŸãïŒããŒã¿äžã«ååšããé¢ä¿ããªã»ãããïŒæ°ããªé¢ä¿ãèŠã€ããããšãã課é¡ãå€ãã§ãããïŒ
â
èŠäºïŒãããã¯ãŒã¯åæã»æŽ»çš
-
ãèŠåº§ããšãèŠç¹èŠçŽ ããå€ããããšã«ããïŒæç« çŸ€ããã倿§ãªç¥èãããã¯ãŒã¯ããåãåºãïŒïŒèŠç¹ãèŠåº§ã®éãã¯ïŒåãæç« ã§ãã£ãŠããã®æç« ã«å¯Ÿããèªèãå€ããŸãïŒããæç« ã¯ããèŠç¹ããã¯å¥ã®ç¥èãæå³ãåãåºãïŒå¥ã®ããèŠç¹ããã¯å¥ã®ç¥èãæå³ãåãåºããŸãïŒäŸãã°ïŒã¯ãã®é£å®³ã®æžç±ãèªãã ãšããŠïŒæ¥æ¬äººãèªãã¯ãã®æãã»æããããšïŒãã·ã¢äººãèªãã¯ãã®æãã»èº«è¿ãã¯ïŒç°ãªããã®ã§ãããïŒããå°éæžãèªãã ãšããŠïŒå°éå®¶ã§ããã°æ°åŒã¡ã€ã³ã§èªããããããŸãããïŒå°éå®¶ã§ãªããã°æ°åŒä»¥å€ãèªããããããŸããïŒããæç« ãçè§£ã§ããªãå Žåã§ãïŒããèŠåº§ã«ãããŠã¯çè§£ã§ããïŒäžäœæŠå¿µåãããã¿ã€ãã«ãããã¯çè§£ã§ãïŒæ³åããããšã¯ã§ããŸãïŒçµå±ã¯ãåãæã®èªç¥ãã«äŸåããïŒããã¯ã¿ãªç¥ã£ãŠããããšããšæããŸãïŒæç« ããåãåºããç¥èã¯åãæã®èªç¥ã«åãã倿§ã ãšããããšã§ãïŒããã§ã¯èŠåº§ãèŠç¹èŠçŽ ãå©çšããããšã«ãããã®æç« ããåŒãåºããç¥èã®å€æ§æ§ãèªç¶èšèªåŠçã§å®è£ ããŠããŸãïŒïŒïŒå®¢èгçãªäºå®ã«åºã¥ãã°äžã€ã®æç« ã«äžã€ã®çã®æå³ããããšæããããããŸããïŒæç¢ºã«æžãããŠããç¹èš±ã§ã䟵害蚎èšçã§èŠè§£ãåãããŸããïŒæ°žé ã«è§£æ±ºããããšã¯ãªãããã§ãïŒäžã€ã®æç« ã«äžã€ã®çã®æå³ã§ãããšåæããããšã¯ïŒå¹»æ³ã«éããªãã§ãããïŒã§ããã°ïŒäžã€ã®æç« ããè€æ°ã®æå³ãåãåºãããããã«ããŠããããšã«æå³ãããã§ãããïŒïŒïŒæç®ãéºäŒåã®èŠç¹ã§èŠãã®ãªãïŒãFRONTEOã®amanogawaãåºæ¥äžããã®ã§ãããïŒæç®ãéºäŒåã®èŠç¹ã§èŠåãã人ã¯ããã§ãããã?ãããªããããã®ãããªèŠç¹ãæã€ããŒã«ã¯äŸ¡å€ãæã€ããã§ãïŒèŠç¹ã¯åãæç« 矀ã«å¯Ÿãæ°ããªäŸ¡å€ãçã¿ãŸãïŒ
-
ãèšæ¶ãããã¯ãŒã¯ãããä»»æã®ããŒãå ã«ãç¥èãããã¯ãŒã¯ããæœåºããæ¹æ³ãåŠã¶ïŒ
-
課é¡ã»ææ®µã®æœåºã¬ãã«ã»ã»ãèŠçŽ ã»èŠåº§ã倿ŽãïŒåºæ¥äžãããããã¯ãŒã¯ã®å·®ããå¯èœæ§ãæœåºããïŒ

-
ããŒããæ§é åãïŒæ¬¡ã®ãããªæå¹ã°ã©ããäœãïŒãã€ãžã¢ã³ãããã¯ãŒã¯ã«ããããªã©ãããåŸã
ã課é¡çŸ€ã»è§£æ±ºææ®µçŸ€ã®å ±èµ·
ãåžå ŽçŸ€ã»æ©èœçŸ€ã»æè¡çŸ€ã®å ±èµ·ããæ©èœå ±ééšåãåŸãïŒç¹èš±æç« ã ãã§ã¯æ å ±äžè¶³ãšã¯æãããïŒåžå Žæ å ±ã¯ã©ãããŠã匱ãïŒç¹èš±ãšããŒãºæ å ±ã®çµã¿åãããäºæž¬ãïŒãã®çµ±åæ å ±ãã察å¿ããåžå Žæ å ±ãåãåºããªã©ã¯ããåŸãïŒïŒ -
åžå ŽçŸ€ã»æ©èœçŸ€ã»æè¡çŸ€ã®å ±èµ·ããæ°åžå Žãªã©æ¢çŽ¢ãã
ãåè
ãããã¹ããã€ãã³ã°ã䜿çšããæ°åžå Žã®æ¢çŽ¢ã«ã€ããŠ
ãhttps://jglobal.jst.go.jp/detail?JGLOBAL_ID=202202243153765222 -
ç¹å®ã®äŒæ¥ã®ãããã¯ãŒã¯ç¹åŸŽã«å¯ŸãèŠäºä»»æã®éã¿ãå ãã€ã€å ç©ãåãïŒé¡äŒŒåºŠãã䌌ãäŒæ¥ãæ¢ãåºãïŒ
-
ç¥èãããã¯ãŒã¯èªäœãæ°ãããã¢ã€ãã£ã¢ã®åºç€ãšãïŒåµé ãèªååºé¡ææ¡ãIPã©ã³ãã¹ã±ãŒãã«çµã¿èŸŒãïŒ
-
ãããã¯ãŒã¯ã®æç³»å€åãšæ°èŠã«çŸããããŒãã確èªãïŒæ°èŠã®ç ç©¶éçºååãæŽãïŒïŒç¹çŸ€ã§ã¯æç¢ºã§ãªãããšããããŸãã®ã§ïŒãããã¯ãŒã¯ã§ã確èªãããšæç¢ºã§ãïŒ

-
ç¥èãããã¯ãŒã¯ã®æçµ¶ãïŒç®èŠãŸãã¯ããããžãŒã§ç¢ºèªãïŒé²ãã¹ãæ°ããªæ¹åãåŸã
-
æè¡ã®ç©ºããç¹ã§ã¯ãªããç¯å²ãã§æ¢çŽ¢ãïŒãåºæ¬ç¹èš±æèã§ãåºé¡æŠç¥ãçå®ããïŒ
ïŒãèŠåº§ãäœãããªããšéãã衚ããªãäŒç€Ÿãã¯ïŒéçºåŸåã«ããäŒç€Ÿã§ãããïŒåé¡ã现ååããéçºã®ã¿ã«æ³šèŠãïŒèŠåº§ãäœããªã£ãŠããïŒæå¹ãªæš©å©ãåããŠããªãåŸåããããŸãïŒãèŠåº§ãé«ãããŠãéããèŠããäŒç€Ÿã ã¯ïŒç ç©¶åŸåã«ããäŒç€Ÿã§ãããïŒããã«å ãïŒåæã¯ã©ã¹ã¿ã®ãšããžã«åºé¡ããŠãããã©ããèŠãããšã§ïŒåºé¡æŠç¥ã®åªç§ããèŠç©ããããšãã§ãããŸãïŒåºãåŸããããç¯å²ã«åºé¡ããã€äŸ¡å€ã®é«ãç ç©¶ãã§ããŠãããèŠç©ããããšãã§ãããŸãïŒïŒ -
ç°ãªãåéã®èŠç¹ãèŠçŽ ãšããŠç¥èãããã¯ãŒã¯ãåæ§æãïŒãã®èŠç¹ã«ãããç¥èãåãåºãïŒbrain machine interfaceã®èŠç¹èªåœã§å»è¬åããŒã¿ããç¥èãããã¯ãŒã¯ã圢æãããªã©ïŒé¢çœãçµæã«ãªããŸããïŒ
-
ãã€ãžã¢ã³ãããã¯ãŒã¯ãšãïŒãã®æ¯éå£ã§ã¯èª²é¡ã«é©ããè§£æ±ºææ®µãäœã§ãããïŒè§£æ±ºææ®µã«é©ããŠãããããããªãæ°ããªèª²é¡ã¯ãªã«ããªã©ïŒèª¬æãããïŒ
https://qiita.com/kzuzuo/items/2bce9e4fe58021a25430 -
ç¹èš±åºé¡ç¶æ³ããã¥ãŒã¹ïŒãããã¯ãŒã¯åæã®èª¬æãããšã«ããŠïŒããã®æèŠãéçŽããªãããã€ãºã«ãã確信床ãèšç®ãïŒå®è¡ã®ã¿ã€ãã³ã°ã確å®ããããïŒãšåæã«æ¯åç°ãªãããšãè¿°ã¹ãããã®é ã®äžãæç¢ºã«ãããïŒïŒäºä»¶ã®èµ·ããææãäºæž¬ãã

-
ãããã¯ãŒã¯ããæ³šç®èŠçŽ ãåãåºãïŒãã®æšç§»ãjudgyprophetã®ãããªãã€ãºç³»ã®ããŒã«ãçšãïŒäºæž¬ãé£ãã€ãã³ãã»äºä»¶æ å ±ãŸã§å«ãèŠç©ããïŒå°æ¥ã€ãã³ããèµ·ããäºå ãçºèŠããéã«ïŒããæ£ç¢ºã«åœ±é¿åºŠãäºæž¬ã§ããããã«ãã
https://github.com/AstraZeneca/judgyprophet

-
äžå¿æ§ã®éãããïŒç¹èš±ãšæç®ãèŠåãã
-
å€ãã®ã€ãããŒã·ã§ã³ã¯ïŒæ°èŠã®ããŒãã®çºçãããïŒæ¢åããŒãéã®ãšããžã®çµã¿æãã§çããããšãçè§£ããïŒã€ãŸãïŒç¥èãåæãšããç¥èã®ç Žå£ãšçæãå¿ èŠã§ããããšãçè§£ããïŒ
-
ãšããžã®çµã¿æãã®éèŠæ§ãçè§£ããäžã§ïŒæ°èŠã®ããŒãã®çæã®éèŠæ§ãçè§£ããïŒã€ãŸãïŒæ°ããã³ãã®éèŠæ§ãçè§£ããïŒãŸãïŒããŒãããããã¯ãŒã¯äžã«çæãããæ¹æ³ã«ä»ãïŒãã¥ãŒãã³ã€ã³ã¶ã«ãŒãã«ããæ¹æ³ãšããŒã¿ã«ããæ¹æ³ã®éããšæå³ãçè§£ããïŒ
-
ãã人ã®ãèŠç¹ãããã®äººãèå³ãæã€ç¹èš±çŸ€ããåŒãåºããŠããïŒãã®èŠç¹ãå¥ã®ç¹èš±çŸ€ã«é©çšãïŒããã®äººã®èŠç¹ã§ã©ã®ãããªç¥èãããã¯ãŒã¯ãåºæ¥äžãããæç€ºãããããšã«ããïŒãäºãã®ç¥èã®å€æ§æ§ãèªèãããïŒãã£ãŠåæããåŸãããã€ã³ãµã€ãã«ä»ãïŒããã倿§ãªèŠè§£ã«ãããã®ãªã®ã確èªããïŒè©äŸ¡ãããªãã£ãã€ã³ãµã€ãã«ã€ããããåã«åãæã®ç¥èã§ã¯é©åã«çè§£è©äŸ¡ã§ããŠããªãã£ãã ãã§ã¯ãªãã確èªããïŒãªã©ã§ããïŒ
-
èšèªã¢ãã«ã¯æŒç¹¹ããåž°çŽãèŠæã§ããããã§ãããšãããèšèªã¢ãã«ã«æ©èœçç¥èããããã¯ãŒã¯ã§äžããã
-
ããç¹èš±çããé廿¹åã«é¡äŒŒã«åºã¥ãåç¹ã«è¿ãã¢ã€ãã£ã¢ãæ¢çŽ¢ããïŒèª²é¡é¡äŒŒæç³»åé£éãããã¯ãŒã¯ãè§£æ±ºææ®µé¡äŒŒæç³»åé£éãããã¯ãŒã¯ãæ§ç¯ããïŒ
åŒçšæ å ±ã«ã¯æ£æçãªãŸãã¯ãã€ã¢ã¹ãããã£ãŠããããšãããã®ã§ïŒçŽç²ã«ããã¹ãæ å ±ã®ã¿ãããã®ãããªãããã¯ãŒã¯ãæ§ç¯ããŠãé¢çœããããããŸããïŒ
ïŒããç¹èš±ããéå»ã®äžå®æéã®æ¯éå£ãåŸãŠéŸå€ã§é¡äŒŒæœåºãããŒã圢æïŒåœ¢æãããããŒãããããã«ã€ããŠããããéå»ã®äžå®æéã®æ¯éå£ãéå®ãéŸå€ã§é¡äŒŒæœåºãããŒã圢æïŒãšããã°ããã ããªã®ã§ïŒç°¡åã«ã§ããŸããã»ã»ã»ã§ããïŒ30åã£ãŠãšãã§ãããïŒïŒ

DDSç³»ã®ç¹èš±ããã¹ã¿ãŒãããŠïŒå¹å°ã«ãã©ãçãããïŒããã€ã¹ã«ãã©ãçãããïŒçްèä¿åå€ã«ãã©ãçãããïŒé åœã«ãªããœãŒã ãªã©ã«ããã©ãçãããïŒé¢çœãã§ãïŒ
æè¡å¹Žè¡šãšãªãã«ãŒããé¢çœããã©ïŒé¡äŒŒãå°ããã€ãããŠããã«ãŒããã¿ãã®ãçµæ§é¢çœãã§ããïŒ
å³ã®1753ãš1ã®é¢ä¿ã¯ãŸãé¢çœãã§ãïŒ1753ã¯äžåºŠ4系統ã«åã1ã«åæããŠããŸãïŒ1753ãš1ã¯çŽæ¥ã¯ããã»ã©äŒŒãŠããŸããïŒ1753ãšã®é¢ä¿ã§èšãã°ïŒ1ã¯æ°èŠã§ãããïŒãåæãšçºæ£ãç¹°ãè¿ããã¶ã€ã³ææ³ãã®äŸã§ãããã®ããã§ãïŒ
ïŒãã®ãããªãããã¯ãŒã¯ã䜿ãã°ïŒèª¿æ»ãšæ¢çŽ¢ã®éãã説æãããããã§ãããïŒãã¶ã€ã³ææ³ã¯æ¢çŽ¢ãããããã«çšããããç¥èã¢ãã«ã»ç¥è圢æãã¬ãŒã ã®1ã€ã§ããïŒãšãã¡ããšçè§£ã§ããã°ïŒæ¥çãããããæ¥çå€ãæ¢ããããããæ¢çŽ¢ã ïŒã®ãããªïŒãåãªã調æ»ããšãç¥èãçšããæ¢çŽ¢ããæ··åãããããªããšã¯èšããªããªããšæããŸãïŒïŒ

é¡äŒŒæç³»åé£éãããã¯ãŒã¯ãšåŒçšãããã¯ãŒã¯ãéãïŒãé¡äŒŒèªäœã®æ€èšŒããããïŒåŒçšãããŠããªãé¡äŒŒæç®ãèŠã€ããåŒçšã®ãã€ã¢ã¹ã®çšåºŠã«ã€ããŠåæãããïŒç Žå£åºŠã枬å®ããCD indexã®æ€èšŒã«çšããïŒãç¡å¹çç±ãå è¡æç®ãæ¢çŽ¢ããã»ç¡å¹çç±ãå è¡æç®ãèŠã€ããããã®ã€ã³ãµã€ããåŸããïŒãªã©ãããããŸãïŒïŒç»åã¯ãã¹ãç»åã§ãïŒ
é¡äŒŒæç³»åé£éãããã¯ãŒã¯ã®äœãæ¹ã¯æ§ã
ãããšããã§ããïŒäŸãã°ïŒ
ãäžå®æéããšã®æ¯éå£ãäœãïŒç¬ç«ã«é¡äŒŒãèŠãŠããã°ïŒæšç§»ããããããããªãïŒ
ãäžå®æéããšã®æ¯éå£ãäœãïŒéå»ã®å
šæéã®é¡äŒŒãæ¬æ°çã«èŠãŠããã°ïŒãç Žå£çã¢ã€ãã£ã¢ããã©ãã«ååšããŠãããããããããããªãã§ãããïŒ
ïŒç Žå£çã¢ã€ãã£ã¢ã®ååšã¯CD indexã§ç€ºãããšãã§ããããã§ãïŒCD indexã¯åŒçšæ§é ã®å€åãã¹ã«ã©ãŒã«ææšåãããã®ã§ããïŒãã®ãããªåŒçšæ§é ã®å€åã¯ãã®ãŸãŸãããã¯ãŒã¯ã«ãããšããå€ãã®æ
å ±ãåŸãããã§ãããïŒïŒ
ãåè
ããã¯ãã«è¡šçŸãçšããå æé¢ä¿é£éã®æœåº
ãhttps://www.jstage.jst.go.jp/article/jsaisigtwo/2018/FIN-020/2018_50/_pdf/-char/ja
ãPapers and patents are becoming less disruptive over time
ãhttps://www.nature.com/articles/s41586-022-05543-x
ãCD indexïŒã€ãããŒã·ã§ã³ã®è¡°éïŒåŒçšæ
£äŸãªã©
ãªã©
å人çã«ã¯ïŒãããçµ±åããããŒã«ãäœããšé¢çœããšæããŸããïŒ
èŠã¯ïŒæ¯éå£ããèŠç¹ãåŸïŒèŠç¹ã®ååãäºä»¶å«ã人ã®ç¥èãå ãã€ã€äºæž¬ããŠããããã§ãïŒ
ã
ïŒGPT-3ãªã©èšèªã¢ãã«ã«è€æ°ã®äž»é¡ããšã«è€æ°ã®æç« ãçæããïŒãã®æç« çŸ€ãç¥èãããã¯ãŒã¯ã«å€æããããšã«ããïŒèšèªã¢ãã«ããç¥èãæœåºããïŒãªã©ãè¡ãããšãã§ããŸãïŒå¿çšç¯å²ã¯ãšãŠãåºãã§ãïŒ
ïŒé©æ°çæ²»çãå®çŸïŒãã¬ã€ã³ãã·ã³ã€ã³ã¿ãŒãã§ãŒã¹ãéæããå»çã®æªæ¥
Paradromics 嵿¥è
å
ŒCEO ãããã»ã¢ã³ã°ã«æ°ã€ã³ã¿ãã¥ãŒ
https://ascii.jp/elem/000/004/086/4086602/
ã³ãŒãæŠèŠ
ïŒãããã¹ãã«tfidfãé©çšãä»»æã®åæ°ã®äžäœç¹åŸŽåèªãåã
ïŒèšèªã¯å
šäœè«çã·ã¹ãã ã§ããïŒsaussureïŒïŒæ¯éå£ã«äŸåããŠå€åããç¹æ§ãéèŠïŒ
ïŒãtfidfäžäœç¹åŸŽåèªãããããword2vecã«ããåèªåæ£è¡šçŸãã¯ãã«ã«å€æããïŒtfidf vectorã¯äœ¿ããªãïŒ
ïŒã«ãŒã«ããŒã¹ã®ç¥èãåæ£è¡šçŸã«å€æãé£ç¶çã«æ±ãïŒ
ïŒãäžèŠèªé€å»ïŒäžéšã®åèªã¯èšå®ããã¹ãããã¯ãŒããšãã¯ãã«ã®é·ãã«åºã¥ãé€å»ããïŒïŒ
ïŒãã¯ã©ã¹ã¿ãªã³ã°ã«ããç¹åŸŽåèªãç³ã¿èŸŒã¿ïŒããã¹ãã®ææ§é ãæ¬äŒŒçã«åæ§æããïŒ
ïŒé³åã«åãããïŒã©ã®ãããªç®çãéæããããã¯ïŒãã®é³åã®åœ¢ã§æ±ºãŸãïŒïŒïŒäœçžã®éšåéåãåé¢ããïŒãšãã£ãŠãè¯ãïŒïŒïŒç©ºçœãäœã§åããããšèšã£ãŠãè¯ãïŒ
ïŒæç« ãã¯ãã«ãäœãã«ãããïŒãã®ãããªéçšãçµãã«å
šåèªãã¯ãã«ãåæãããåèªããšã®éã¿ãèæ
®ããªãææ³ãçšããå Žåã«ã¯ïŒç¢ãªæ§èœãåºãªãããšã¯ç¥ãããŠããŸãïŒåèªã«å¯ŸãéžæãããŸãã¯åæãè¡ããªã©ïŒäœããã®åŠçãããå¿
èŠããããŸãïŒãŸãïŒãã®åŠçã«ããå¯èŠåæã«æ¹åãç¶æã§ãããã©ãããæ±ºãŸããŸãïŒåèªéžæãåèªå
åæã¯LSTMãattentionãç®æããšãããšåãã§ããïŒãããã§ã¯æ¹åã倱ãããªãããã«æ³šæãå¿
èŠã§ãïŒ
ïŒãåæ§æãããææ§é ããŸãšãïŒæç« ãã¯ãã«ãäœã
ïŒãæç« ãã¯ãã«ã«å¯Ÿãã¯ã©ã¹ã¿ãªã³ã°ãè¡ãïŒã€ã³ã¿ã©ã¯ãã£ãã«å³ç€ºããïŒ
ã³ãŒãã¯èŠæã«å¿ãgithubã«ãããŸãïŒ
https://github.com/kzh22
ãããæ±ãã³ãŒããªã®ã§ãªãã¡ã¯ã¿ãªã³ã°ããŠããã§ããã»ã»ã»ãšãããŸãããªãã¡ã¯ã¿ãªã³ã°ããæ°ã倱ããããŠããã®ã§ã³ãŒãæäŸããŸãã®ã§èª°ããã£ãŠãããŸãããããã»ã»ã»
ãŸããå人çã«ã¯ãæ¹è¯ãã仲éãæ±ããŠããŸããå
¬éããŠããã£ãŒãããã¯ããªãã®ã§ããã°å
¬éããæå³ã¯ãªããšèããŠããŸãããååããã ããå Žåã¯ãé£çµ¡ãã ããããœãŒã¹ãªã©ãéãããããŸãã
ïŒæç« ãã¯ãã«åè
https://speakerdeck.com/hpprc/lun-jiang-zi-liao-simcse-simple-contrastive-learning-of-sentence-embeddings-823255cd-bd1f-40ec-a65c-0eced7a9191d?slide=14
ã»BERTãããã€ãŒãã«æœåºããæåã蟌ã¿ã¯STSã«ãããæ§èœãäœãããšãç¥ãããŠãã [11]
BERTã®æèååèªåã蟌ã¿ã®å¹³åãCLSã®ãã¯ãã«ãªã©
ã»GloVeãfastTextãªã©ã®éçãªåèªåã蟌ã¿ã®å¹³åããšã£ãæ¹ãBERTããæ§èœããã
ã»äžæ¹ã§ïŒäžæµã¿ã¹ã¯ (sentiment classificationãªã©)ã«ãããBERTç±æ¥ã®æåã蟌ã¿ã®æ§èœã¯ ããçšåºŠé«ãç¹ã«æ³šæ
ã»BERTãªã©äºååŠç¿æžã¿èšèªã¢ãã«ã®åã蟌ã¿ç©ºéã¯ç°æ¹æ§ (anisotropy) ãæã¡ [12]ïŒ ããã STSã¿ã¹ã¯ã®æ§èœã«æªåœ±é¿ãäžããŠããå¯èœæ§ã瀺åãããŠãã [13]
ã»æ§ã
ãªæç« ãã¯ãã«äœææ¹æ³
p-mean: æåã蟌ã¿ãåèªåã蟌ã¿ïœã§èšç®
SWEM: åèªåã蟌ã¿ã å¹³å / max/å¹³åãšmaxã®concat / 屿çªããšã«å¹³åããŠããmax ããšã
GEM: æäžã®åèªåã蟌ã¿ã®çŽè¡åºåºãããšã«noveltyãªã©éã¿ãèšç®ããŠåèªåã蟌ã¿ãéã¿ä»ãå
DynaMax: äºã€ã®æã®åèªåã蟌ã¿ãstackããè¡åãäœã Fuzzy setã®èããå
ã«Fuzzy Jaccardä¿æ°ãèšç®
SIF: ïœãèšç® âåã蟌ã¿è¡åãç¹ç°å€åè§£â第äžç¹ç°ãã¯ãã«uã§ïœãèšç®
uSIF: è€æ°ã®ç¹ç°ãã¯ãã«ãå©çšïŒ ç¹ç°å€ã®ç·åã䜿ããã€ãã©èª¿æŽäžèŠãªSIF
P-SIF: åèªã®ãããã¯ãã¯ãã«ã䜿ã£ãSIF
All-but-the-Top: åèªåã蟌ã¿ã®éåãPCAããŠäžäœäž»æåãé€ã
Word Mover's Distance: æé·ã®éæ°ãäžæ§ãªç¢ºç質é, ã³ã¹ãããŠãŒã¯ãªããè·é¢ãšããŠæé©èŒžé
Word Mover's Embedding: æ(æž)ãšãµã³ããªã³ã°ããè€æ°ã®æ(æž)ãšã®WMDã®åãæ(æž)ãã¯ãã«ãšãã
Word Rotator's Distance: åèªåã蟌ã¿ã®ãã«ã ã確ç質é, ã³ã¹ããã³ãµã€ã³é¡äŒŒåºŠãšããŠæé©èŒžé
ïŒäœæ¥ã¡ã¢
UMAPã䜿çšããå Žåã¯rapidsãéã
knnã¯faissã䜿çšãããšGPUã䜿çšããŠé«éïŒ
https://yaakublog.com/deep_metric_learning
察ç
§åŠç¿ã§æå³çã«éèŠèªã®ãã¯ãã«ãè£æ£ãããªã©ããåŸã
äŸãã°äž»å€ã®é¡çŸ©èªèŸæžãäœã£ãŠããïŒé¡çŸ©èªãå«ãæç« ã®å Žåã¯äž»å€ã®ãã¯ãã«ã«è¿ã¥ãããªã©ïŒæ¯éå£ã§ã®èª¿æŽä»¥å€ã®äºåŸçãªä»»æèª¿æŽææ®µãšããŠèããŠããããšïŒ
ïŒword2boxãå®è£
ããŠã¿ãã
https://github.com/iesl/word2box
ïŒsetfitãè£å©ãšããŠè©ŠããŠã¿ãã¹ãã§ã¯ããã
äœ¿ãæ¹
- å眮ãïŒåçš®èšå®ã¯ã³ãŒããã®ãã®ãä¿®æ£ãè¡ããŸãïŒ
ïŒãjupyter notebookã§ipynbãã¡ã€ã«ãéããŸãïŒ
ïŒãåæãããç¹èš±ãèšèŒãããxlsxã»CSVãã¡ã€ã«ãä»»æã®å Žæã«çœ®ããŸãïŒ
ããã»åºé¡çªå·ïŒåºé¡äººïŒçºæã®åç§°ïŒèŠçŽïŒè«æ±ã®ç¯å²ïŒã®èšèŒãæäœéå¿
èŠã§ãïŒ
ããã»ïŒè¡ç®ã¯åã¿ã€ãã«ãšãïŒïŒè¡ç®ããåç¹èš±ã®å
容ãèšèŒããŠãã ããïŒ
ããïŒç¹èš±æ€çŽ¢ãœããããæ€çŽ¢çµæãåºåããå ŽåïŒãããã®åœ¢åŒãšãªã£ãŠããäºãå€ãããšæãããŸãïŒå¿
èŠã«å¿ãååã倿ŽããŠãã ããïŒ
ïŒãïŒãxlsxã»CSVãã¡ã€ã«åã«å€æŽããŠãã ããïŒ
ïŒãå¿
èŠã«å¿ãïŒïŒéšåã倿ŽãïŒè²ãã·ã³ãã«ã倿ŽããŠãã ããïŒ
ïŒãç®çã«å¿ãïŒãã©ã¡ãŒã¿ãèšå®ããŠãã ããïŒ
ïŒãã¡ãã¥ãŒãããã¹ãŠå®è¡ãéžæããŠãã ããïŒ
ïŒããã°ãåŸ
ã¡ãŸãïŒåŸ
ã¡æéã¯1æé/5000ä»¶ã»ã©ã§ãâ10å/5000ä»¶ã»ã©ã«é«éåããŸããïŒnewïŒ
ïŒãçµäºåŸïŒå¥ãŠã€ã³ããŠã§ã€ã³ã¿ã©ã¯ãã£ãHTMLãéããŸãïŒ
ïŒãçµæãèŠã€ã€ïŒæééå®ã»åºé¡äººéå®ã»åé¡éå®ã»ããŒã¯ãŒãéå®ïŒæ£èŠæ€çŽ¢ïŒæŠå¿µæ€çŽ¢ïŒnewïŒïŒattentionæŠå¿µæ€çŽ¢ïŒæ€èšŒäžïŒïŒèŠçŽå³ç€ºã«å€æïŒtransformerçµã¿èŸŒã¿ïŒé«é忀èšäžïŒïŒé¢é£è£œåæ
å ±æ€çŽ¢ïŒå¥ã®åèªãžã®çœ®ãæããŸãã¯ãã®éå«ãïŒæ€èšŒäžïŒïŒé¢é£æç®æ€çŽ¢ïŒæ€èšäžïŒïŒLaBERTãã¯ãã«é¡äŒŒæ€çŽ¢ïŒæ€èšŒäžã»æ€èšŒçµæã«ãã£ãŠã¯ãã¡ãã§å³ç€ºããããšãæ€èšïŒBERTãã¯ãã«ã¯mean-poolingã§ãäºååŠç¿ã®åèŠã匷ããã€ã³ãµã€ããåŸããã«ããã£ãããä¿çäžïŒïŒãªã©ããããªãïŒåæãé²ããŸãïŒ
ïŒç¹æ®ãªæ€çŽ¢ã»éå®çµãèŸŒã¿æ©èœä»¥å€ã®æ€çŽ¢ã»éå®çµãèŸŒã¿æ©èœã¯ïŒpandasã®æ©èœã«äŸåããŠããŸãïŒ
åèã»è£è¶³ã»æžããªããã¡ã¢
ãã®ä»
ã»ææžãã¯ãã«ããæè»œã«é«ã粟床ã§äœããSCDVã£ãŠå®éã©ããªã®ãæ¥æ¬èªã³ãŒãã¹ã§å®éšãã(EMNLP2017)
https://qiita.com/fufufukakaka/items/a7316273908a7c400868
tfidf embeddings/cluster visã¯SCDVãšåºæ¬çãªèãæ¹ãé¡äŒŒïŒæåããæ°ã¥ããŠãããšæããªãããªããïŒæãã€ããããšããããå®è£
ããããªãå®è£
ã®éçšã§æ§æããŸãšãŸãæ§æããŸãšãŸã£ãŠåããŠéå»äºäŸããŸãšãã«æ€çŽ¢ã§ããããã«ãªãã®ã§ã»ã»ã»ïŒïŒ
ã¹ããŒã¹ãéäžã§æœãã®ãæåŸã«æœãã®ãã«ã€ããŠã®èãæ¹ã®éããïŒãããã¯ã«æ³šç®ãã¯ã©ã¹ã¿ãªã³ã°ããã®ãæã®æ§é ã«æ³šç®ãã¯ã©ã¹ã¿ãªã³ã°ããã®ãã«ã€ããŠã®èãæ¹ã®éããããïŒ
ïŒtfidf embeddings/cluster visã«ãããŠããããã¯ãèæ
®ããŠãã¯ãã«ãè£æ£ããŠãè¯ãããã©ïŒã¯ã©ã¹ã¿éã®åé¢ãè¯ãããããšãç®çãšããŠããªãã®ã§ãã®ãŸãŸãšããŠããïŒãã®ãããªè£æ£ææ³ã§ã¯ã©ã¹ã¿ã®åé¢ãè¯ãããŠå¬ããã®ã¯ïŒå³ç€ºãããšãã§ã¯ãªãïŒãšæã£ãŠããïŒè»žã®æå³ãèããªãããã¯ã€ãã¹ããŒã¹ã¯èŠã«ãããªããâŠã¢ã€ãã£ã¢çæãšããé¢ã§ãå¬ãããªããªïŒïŒ
ã»valunexãradar
https://www.valuenex.com/jp/valuenex-radar
ã¯ããã§ããŠãããªãšæå¿ããïŒ
tfidf embeddings/cluster visã§ã¯ãã¯ã€ãã¹ããŒã¹ãæ£ç¢ºã«æç€ºããããšã¯é£ãããããããªãïŒ
ïŒãã®åŸïŒtfidf embeddings/cluster visã«ãããŠããã¯ã€ãã¹ããŒã¹ã®æç€ºãã§ããã ãããšçµè«ã¥ããïŒããããã®ç°ãªãæ¯éå£ã«ãããŠPCAå ååæãè¡ã確å®ã«ååšãããã¯ã€ãã¹ããŒã¹ã確èªãïŒããããšåãçžå¯Ÿäœçœ®ã«ãã¯ã€ãã¹ããŒã¹ãçŸããããšã確èªã§ããïŒãã ãïŒæ¯éå£ã倧ããç°ãªãå Žåã«ã¯å調æŽãå¿
èŠã ããïŒïŒ

ïŒä»åŸïŒã³ã¢ãé ãããæè¡ãå
¬éããªãããã«æš©å©ãåãç¥è²¡æŠç¥ãäžè¬åãããããããªãïŒãããããšãã¯ã€ãã¹ããŒã¹ãèŠã€ããæè¡ã¯ããéèŠãšãªãã ããïŒç¹èš±ã«ååšããªããæç®ãåºåçãååšãã空éãªã©æªãããšèšãããããããªãïŒæ¬¡åã¯ãã®ãããªïŒä»åŸã®ç¥è²¡æŠç¥ãæ³å®ããç¥è²¡çŽäºçšããŒã«ãäœãäºå®ïŒ
ïŒã·ããªãªåæã¯ã©ããªã®ããªïŒã³ã³ãµã«äžå¿ãªã®ã§åé¡ãªããšæããïŒç¹èš±ã®ã¿ããã·ããªãªãæ±ããã¹ãã§ã¯ãªããšã¯æãïŒä»ç€ŸååãããŒãºãäŒç€Ÿæ¹éãå¿
é ãšããªããšâŠ
ã»Valuenexã¬ããŒãã
ç¹èš±ã®äžã®ãçŸãïŒãçŸãã¯ã©ãã«æœãã§ããã®ãïŒ
https://www.valuenex.com/materials-database/report/200422/beautyinpatent
ãštfidf embeddings/cluster visãšãæ¯èŒïŒ
ã¢ã«ãŽãªãºã ã¯ç°ãªããå·šèŠçã«ã¯ã»ãŒåãåºåãšãªããšç¢ºèªã§ããïŒ
ã€ãã§ãã£ãšåæãããšããïŒçŸã«ãããŠãå
ããæ°ããŒãºãšããŠçãŸããŠããã®ã ãããªïŒãšã®ç€ºåãåŸãããïŒ
äŸãã°ïŒæ¬¡ã«é¢ããããŒãºã®èµ°ããèŠãããïŒ
äžè±é»æ©ã®ãé空ç
§æãå®ç©ãèŠãŠããïŒæ¬ç©ã®ç©ºãšåºå¥äžèœ ãå±å
ã®é塿ããè§£æ¶ããæ°æè¡
ïŒFRONTEOïŒæ§æ¥ã®åµè¬ããã»ã¹ãå€é©ããè«ææ¢çŽ¢AIã·ã¹ãã ãAmanogawa(ããŸã®ãã)ãver.1.0ãçºè¡š(202007)
https://www.fronteo.com/20200707
ã€ã³ã©ã©ã¯ãã£ãã«æç®æ
å ±ãèŠãããããã ïŒ
ãã®ãããªããŒã«ã欲ããã£ãïŒå
šæãããã¯ãã«ãäœã£ãŠãããªãã°ïŒãŸãã¯valuenexãæç€ºããããã«ç¹èš±ãšæç®ãåäžé¢ã«é
眮ã§ãããªãã°ïŒæ¡çšãããïŒ
ç¹èš±åºé¡äžãšã®ããšïŒå
¬éäžã®ç¹èš±ã«è©²åœããããªèšèŒã¯èŠã€ãããªãïŒããããå
¬éãããã®ã ããïŒåªå
æ¥ãšå
容ã«ãã£ãŠã¯ã³ãŒã倿Žå¿
èŠãããããªãïŒå¯©æ»ã«ããå
¬ç¥ç¯å²ããããããã«ãªãïŒãããããïŒ
Concept encoderã®ç¹èš±ç¢ºèªïŒtransformerã®keyïŒqueryã«ããæèãåãåºããŠããéšåãšäŒŒãåŠçã«ããéèŠåèªãååŸã粟床ãå¢ããŠãããšããããšããª?ïŒãããªæãã§

tfidfã¯æç« äœæè
ã®çã«äŸåããŠããŸã匱ç¹ãããïŒæèãèªã¿åãææ³ã¯ããè¯ãã ããïŒãããæèãèªã¿åãææ³ã§ãïŒé¡çŸ©èªãç¹°ãè¿ãããå Žåã®éžå¥ãšæäœè¡šçŸæ°ã®ç¢ºä¿ïŒäºåç¥èã®æ€èšïŒäžå³äžäžã®éãã®éãïŒäºåç¥èã«ããéžæã¯å€åããïŒãããã»ããè¯ããšæãã®ã ãããŠããã®ããª?

202010æç¹ã®tfidf embeddings/cluster visã®å€èгïŒ
amanogawaã£ãŠè¯ãååã ããªã
ïŒtfidf embeddings/cluster visã¯ïŒã¡ããã©valuenexãšconcept encoderã®äžéã®æè¡ã䜿ã£ãŠãããšè¡šçŸãããšè¿ãããªïŒéãããããšãããšïŒããçšåºŠä»»æã«ã¢ã«ãŽãªãºã ã®èª¿æŽãã§ããããšããªïŒ
ïŒç¹èš±ãšè£œåãåäžé¢ã«é
眮?ïŒæ€èšãããïŒ
Navigating a product landscape for technology opportunity analysis: A word2vec approach using an integrated patent-product database
https://www.sciencedirect.com/science/article/abs/pii/S0166497220300237?via%3Dihub
ç±³åœç¹èš±ããã³åæšããŒã¿ããŒã¹ããçµ±åãããç¹èš±è£œåããŒã¿ããŒã¹ãæ§ç¯ããŸãïŒæ¬¡ã«ïŒword2vecã䜿çšããŠïŒæè¡çé¢ä¿ãç¶æããªããïŒé¡äŒŒããæè¡åºç€ãæã€è£œåãäºãã«è¿ãã«é
眮ããããã¯ãã«ç©ºéã¢ãã«ãšããŠè£œåã©ã³ãã¹ã±ãŒããæ§ç¯ããŸãïŒ
ïŒãåã¯ãã®ãšãã¯ã£ãšããïŒãªãããªã³ã¹ãã³ã®å®éšå®€ããïŒã©ãã©ãå ±åãåºãŠããã®ãã«æãåœãã£ãããã ïŒåœŒãã¯å®éã«èªåãã¡ã®æã§é ããããè£ çœ®ã§ç ç©¶ããŠããã®ã ïŒã ããããã©ãã«äœãããïŒäœãã©ãåããŠããããïŒã¡ãããšããã£ãŠããã®ã ïŒïŒãã¡ã€ã³ãã³ïŒã
ïŒæ·±å±€åŠç¿ãå©çšããç¹èš±è«æ±é
ããŒã¹ã®ç¹èš±æè¡ä¿¯ç°ããã
*åªç° å¡å², å®®æ ç¥äž, ç¥æŽ¥ åæŠ
äž»å¬: äžè¬ç€Ÿå£æ³äºº 人工ç¥èœåŠäŒãäŒè°å: 第34åå
šåœå€§äŒ(2020)
https://www.jstage.jst.go.jp/article/pjsai/JSAI2020/0/JSAI2020_4Q3GS903/_article/-char/ja/
self-attention LSTM
ïŒç¹è𱿿žãã¯ãã«ãçšããäŒæ¥äŸ¡å€è©äŸ¡
*è€å å å¹³, æŸæ¬ ç¥ä», è
æå, 髿© 倧å¿
äž»å¬: äžè¬ç€Ÿå£æ³äºº 人工ç¥èœåŠäŒãäŒè°å: 第34åå
šåœå€§äŒ(2020)
https://www.jstage.jst.go.jp/article/pjsai/JSAI2020/0/JSAI2020_3Rin456/_article/-char/ja/
SCDV+
Constructing a Decision-Making System Using Patent Document Analysis
https://link.springer.com/chapter/10.1007/978-981-16-2994-5_23
ïŒç¹è𱿿žæ
å ±ã察象ãšããã³ã³ãã³ãåæã®ææ³ãšåºé¡äººã¿ã€ãå¥ç¹æ§æ¯èŒ
å
¬éæ¥: 2020.06.08
èè
: å°æŽ çïŒæ± å
å¥å€ªïŒå
æ© äžä¹
éèªæ
å ±: STI Horizon, Vol.6, No.2
https://www.nistep.go.jp/activities/sti-horizon%E8%AA%8C/vol-06no-02/stih00210
åæ£è¡šçŸâk-meansâumapïŒç¹åŸŽèªæœåºå·¥çšã¯ååšããŠããªãïŒïŒ
å ±éãã課é¡ãšè§£æ±ºãã¹ã課é¡ã¯äœããªã»ã»ã»å®è£ ãæ¯èŒããŠã¿ããã»ã»ã»
ïŒTopic Modeling with BERT
https://towardsdatascience.com/topic-modeling-with-bert-779f7db187e6
ïŒã·ãããšã«ã³ããã§ãŒã
ç¹å®äŒæ¥ã®æŠç¥çç¹èš±åºé¡ç¶²ã®äºäŸåæã瀟ã®èªè»¢è»éšååžå Žã«ãããç¥çè²¡ç£æŠç¥
https://www.jstage.jst.go.jp/article/infopro/2020/0/2020_1/_pdf/-char/ja

åèïŒããŸæ¬²ããé»åã³ã³ãã¯ã·ãããããšãã¹ã©ã ïŒ ããããã®é·æãšçæã«è¿«ãïŒ
https://blog.cbnanashi.net/2020/10/15316
ç§ã¯ã·ãã&SRAM掟ã§ãããããªã¢ã®æ¯æ°ãã
ïŒãã³ãïŒç¥è²¡åæãœããçºå£²ãç°æ¥çš®é£æºãåŸæŒã
https://r.nikkei.com/article/DGXMZO66223610T11C20A1910M00
ãç¥ç財ç£ã®ããŒã¿ããŒã¹ã®æ
å ±ãªã©ãè§£æãïŒç¹å®ã®åéã§åæ¥ã§ããããªäŒæ¥ãå¹ççã«æ¢ãã
è峿·±ãïŒã»ã°ã¡ã³ããŒã·ã§ã³äžå¿ã®ææ³ãªã®ãã¯ã©ã¹ã¿ãªã³ã°äžå¿ã®ææ³ãªã®ããã€ããªãããïŒ
ã俺ãã¡ãå¿
æ»ã§çšŒãã 1åïŒ1ç§ãITã«äœ¿ãããªãããšãŸã§èšãããããã³ãã®ææŠ
https://www.atmarkit.co.jp/ait/articles/1908/26/news012.html
ãããæãåºãïŒ
ïŒå»è¬ã®æ°é å äžååå»è¬ã®éçºç¶æ³ãç¹èš±ã§èªã¿è§£ã
é
äºãªæªäºé å㌠AIæŽ»çš ãŒ
https://www.jpo.go.jp/news/koho/kohoshi/vol46/01_page3.html
40äžä»¶ïŒtfidf embeddings/cluster visã§ã¯äžã€äžã€ã®ç¹èš±ãçç¥ããªãã®ã§éãããïŒ
ïŒIPRally
https://www.iprally.com/
ç¥èã°ã©ãããŒã¹ã®ç¹èš±æ€çŽ¢
深局åŠç¿ã«å¯Ÿãç¥èã°ã©ãã®çµæãå ããŠ?粟床ãäžããŠããããã ïŒ
ïŒãèµ·èæèæ
®ç¥èãããšã«ããç¥èã°ã©ã
ïŒã審æ»å±¥æŽ
ïŒãïŒïŒïŒãå ããæ·±å±€åŠç¿
ç¥èã®çµã¿èŸŒã¿ãšãããã¬ã³ãéã?ïŒ
ç¥èã°ã©ãããå€ããéšåãã©ããªããæ°ã«ãªãã粟床ã¯é«ãã ããïŒ
æ¢åã®æ§é ããæœåºãã?èŠçŽ ãèæ
®ããã®ã§ïŒå
è¡æè¡èª¿æ»ïŒæš©å©åå¯èœæ§èª¿æ»ïŒç¹ã«ç¡å¹è³æèª¿æ»ã«åŒ·ããïŒãŸãïŒæ§æèŠçŽ éšåãä»»æã«åãåºãææ³ããæŒããå°ãªãããïŒå€äŸã®äºç¹ãã©ãåãåºããæªã ã«æ©ãã§ãããåèãšã§ããã?
ïŒSTNext - æ°æ©èœ ãChemscapeã ãªãªãŒã¹
20210125
https://www.jaici.or.jp/news/news.php?selectedId=2324
ãååŠç©è³ªã®é¡äŒŒæ§ãè§£æãã Chemscape (ã±ã ã¹ã±ãŒã) ãæèŒãããŸããïŒãããå
ã«ã¯é¢é£ç¹èš±ã®ä»¶æ°ã 3D ã§è¡šç€ºãããããïŒç©è³ªã®é¡äŒŒæ§ãšååŠé¢é£ç¹èš±ãèŠèŠçã«è§£æããããšãã§ããŸãïŒã
å
å®ããŠãããªã
ïŒçµ±åæ
å ±éã®æ
å ±å¹Ÿäœçå®çŸ©
https://tripdancer0916.hatenablog.com/entry/2018/08/04/%E7%B5%B1%E5%90%88%E6%83%85%E5%A0%B1%E9%87%8F%E3%81%AE%E6%83%85%E5%A0%B1%E5%B9%BE%E4%BD%95%E7%9A%84%E5%AE%9A%E7%BE%A9
åçš®ã®æ
å ±éã¯ïŒèŠçŽ éã®çžäºäœçšãå
šãŠèæ
®ã«å
¥ãã"full model"ãšããããäžéšã®çžäºäœçšãåãé€ãã"disconnected model"ãšã®è·é¢ãšããŠå®çŸ©ã§ããïŒ
ïŒLegalscape
https://www.legalscape.co.jp/
å€äŸè§£æïŒãè¡ã£ãŠããïŒLegalscapeãšããäŒç€ŸïŒ
ç¹èš±6469919
ããçºæã解決ããããšãã課é¡ã
ãïŒïŒïŒïŒã
ããããªããïŒç¹èš±æç®ïŒã«é瀺ãããèšç·Žçšæç« ã®åé¡åãã§ã¯ïŒæç« æ¯ã«ãããã¯ãŸãã¯ãç¹åŸŽããä»äžãããŠããããïŒæç« å
šäœãæ§æè§£æããçµæãšããŠïŒæç« å
šäœã«ãããã¯ãŸãã¯ãç¹åŸŽããä»äžãããïŒ
ãïŒïŒïŒïŒã
éåžžïŒå€æ±ºæãªã©ã¯ïŒçç±ïŒè«æ±ïŒäž»åŒµïŒãªã©ã®å€ãã®é
ç®ãããªãïŒãã£ãŠïŒããããã®é
ç®ã®ç¹åŸŽã¯ïŒãã®äºä»¶ã«ãã£ãŠå
容ïŒçš®é¡ïŒãç°ãªãïŒãããã®è€æ°ã®ç¹åŸŽãå¹³åããŠïŒã€ã®å€æ±ºæã«ã©ãã«ãä»äžããããšã¯ãŠãŒã¶ã«ãã倿±ºæã®æ€çŽ¢ç²ŸåºŠãäœäžãããŠããæããããïŒããããïŒè€æ°ã®é
ç®ã®ç¹åŸŽãå¹³åããŠïŒã€ã®ã©ãã«ãå°ãåºãããšèªäœã粟床ã®äœãã©ããªã³ã°ææ³ã§ãããšããããïŒ
ãïŒïŒïŒïŒã
æŽã«ïŒå€ãã®å€æ±ºæã¯ïŒãã®å€æ±ºæã詳现ã«èª¬æããè©éãšåŒã°ãã解説æãäœæãããŠããïŒãã®è©éã倿±ºæãšç
§ããåãããŠèªãããšã«ãã£ãŠå€æ±ºæã®çè§£ãæ·±ãŸãïŒãããïŒå€æ±ºæãšè©éã®æç« äžã®å¯Ÿå¿é¢ä¿ã«ã€ããŠã¯ïŒèªè
ã«ãã£ãŠèŠåºã仿¹æ³ã¯ãªãïŒå€æ±ºæãè©éã®èªè§£ãå°é£ã«ããŠããïŒ
ãïŒïŒïŒïŒã
ãŸãïŒå€æ±ºæã¯ïŒèšŽç¶ã«èšèŒããã蚎ãã®å
容ã«å¯Ÿç
§çã«äœæããããã®ã§ããïŒå€æ±ºæãèªã¿æ
£ããŠããã°ïŒå€æ±ºæãèªãéã«ïŒèšŽç¶ãšç
§ããåãããŠèªãå¿
èŠã¯ãªãïŒããªãã¡ïŒèšŽç¶ãšå€æ±ºæãšã¯ïŒå¯Ÿå¿é¢ä¿ã«ããã«ãé¢ãããïŒèšŽç¶ã®äœæã«ã¯è«å€§ãªæéãšè²»çšããããããšãéåžžã§ããïŒäžè¬äººã蚎èšãèµ·ããããšã®åŠšããšãªã£ãŠããå¯èœæ§ãããïŒ
ãïŒïŒïŒïŒã
ããã§ïŒæ¬çºæã¯äžèšäºæ
ã«éã¿ãŠãªããããã®ã§ããïŒå€æ±ºæãé©åã«å°æç« ã«ç« ç«ãŠãïŒåœè©²å°æç« ã®é
ç®åããŸãã¯çš®é¡åããè¡ãæè¡ãïŒãã®é
ç®åããŸãã¯çš®é¡åããããå°æç« ãã倿±ºæã«é¢é£ããä»ã®å€æ±ºæã®å°æç« ãè©éãªã©ã®å°æç« ã®æ€çŽ¢ãè¡ãæè¡ãïŒãã®é
ç®åããŸãã¯çš®é¡åããããå°æç« ãã蚎ç¶ãªã©ã®æãäœæããæè¡ãæäŸããããšãç®çãšããïŒã
ãïœåèšåŠçéšãåèšèšæ¶éšã«åèšç¬¬ïŒã®æïŒåèšç¬¬ïŒå°æç« ïŒããã³åèšç¬¬ïŒé
ç®ããããã察å¿ä»ããŠèšæ¶ãïŒåèšç¬¬ïŒã®æïŒåèšç¬¬ïŒå°æç« ïŒããã³ç¬¬ïŒé
ç®ããããã察å¿ä»ããŠèšæ¶ãïŒ
æŽã«åèšç¬¬ïŒã®æããã³åèšç¬¬ïŒã®æã察å¿ä»ããŠèšæ¶ãããèšæ¶ã¹ããããšïŒ
ãå«ãããšãç¹åŸŽãšãã倿±ºæããŒã¿ããŒã¹äœææ¹æ³ïŒã
åºããã ãªïŒåºæ¬ã«ãŒã«ããŒã¹ïŒ
ãã®æ³ä»€ã«ç²Ÿéããå°éå®¶çšã§ã¯ãªãïŒ
è€æ°ã®é
ç®ã®ç¹åŸŽãå¹³åããŠïŒã€ã®ã©ãã«ãå°ãåºãããšèªäœã粟床ã®äœãã©ããªã³ã°ææ³ã§ãããšããããïŒãšããã®ã¯ç¢ºãã«ïŒèªåã¯äºç¹ãé
ç®åäœã§åºåãããšãããïŒ
ïŒã¹ããã¯ããŒã¯ãWikipediaãçšããæ¥æ¬èªã®åºæè¡šçŸæœåºããŒã¿ã»ãã
https://github.com/stockmarkteam/ner-wikipedia-dataset
ïŒLanguage-agnostic BERT Sentence Embedding
https://arxiv.org/abs/2007.01852
å€èšèªBERTãã¯ãã«è¡šçŸLaBERTïŒBERTãã¯ãã«ã§å·¥å€«ãããªãã°ãã¡ãããªïŒ
"vocab_size": 501153ïŒèªåœã¯è¶³ããã ãããïŒ
https://huggingface.co/setu4993/LaBSE
å ŽãæŽããŠãããŠããhuggingfaceã«ã¯è¶³ãåããŠç ããªããªïŒ
ïŒsentence BERT
https://medium.com/swlh/transformer-based-sentence-embeddings-cd0935b3b1e0
æç« ãã¯ãã«ã®åŠ¥åœãªè§£ã®äžã€ãïŒ
ïŒThe Dynamic Embedded Topic Model
https://deepai.org/publication/the-dynamic-embedded-topic-model
ïŒçºæã®æœåºçèŠçŽãšçæçèŠçŽïŒæç®æ€çŽ¢çšäžäœäžäœæŠå¿µåèŠçŽïŒå®æœç©æ€çŽ¢çšè¡šå±€åèŠçŽãåæã«ã§ãããšããªã䜿ããïŒ
ãã¹ãŠtransformerãçšãã詊è¡ïŒèª²é¡æœåºäžã ãïŒã©ããŸã§ã§ããããªïŒ
æç®æ€çŽ¢çšèŠçŽã§ã¯è峿·±ãæ€çŽ¢çµæãèªåã§åŸããæ¥œããã£ãïŒ
æå€§ã®èª²é¡ã¯å€ååŠçæéïŒã©ããã§ããªããâŠ
HuggingFace model hubã«ã¢ãã«ã眮ãAPIãçšããŠçµæãåŸããšããæ¹æ³ããããïŒïŒéãããªãïŒ
ãµãŒããŒã欲ãããªâŠ
ïŒGoogleAILanguage Interpretability Tool (LIT
https://pair-code.github.io/lit/
https://colab.research.google.com/github/PAIR-code/lit/blob/main/lit_nlp/examples/notebooks/LIT_sentiment_classifier.ipynb
ãèšèªè§£éå¯èœæ§ããŒã«ïŒLITïŒã¯ïŒNLPã¢ãã«ãèŠèŠåããŠçè§£ããããã®ãªãŒãã³ãœãŒã¹ãã©ãããã©ãŒã ã§ãã
ïŒRuishi Liang; Yilong Li; Xiaomiao Chen; Junwen Chen
2020 International Symposium on Computer Engineering and Intelligent Communications (ISCEIC)
Patent Trend Analysis through Text Clustering based on K-Means Algorithm
https://ieeexplore.ieee.org/document/9325336
TFIDF k-means
ïŒããŒãã«ããŒã¿åãã®èªç¶èšèªç¹åŸŽæœåºè¡
https://zenn.dev/koukyo1994/articles/9b1da2482d8ba1
çŸç¶åºç€æŽç
ïŒïŒ¡ïŒ©ã§ç¹èš±ã»è«æã®é¡äŒŒåºŠãåæïŒNISTEPãããŒã«éçº
https://newswitch.jp/p/27622
ãæéšç§åŠçã®ç§åŠæè¡ã»åŠè¡æ¿çç ç©¶æïŒïŒ®ïŒ©ïŒ³ïŒŽïŒ¥ïŒ°ïŒã®å°æŽç䞻任ç ç©¶å®ã¯ïŒäººå·¥ç¥èœïŒïŒ¡ïŒ©ïŒæè¡ã§ç¹èš±ãç§åŠè«æã®é¡äŒŒåºŠãæ±ããåæããŒã«ãéçºããïŒåºç¯ãªç 究掻åãåé¡ãããïŒäŒŒãç ç©¶ãæ€çŽ¢ãããã§ããïŒæ€çŽ¢ã·ã¹ãã ã®ãããªããŒã«ãšãïŒïŒ®ïŒ©ïŒ³ïŒŽïŒ¥ïŒ°ã«å ããŠæç§çã®è·å¡ã䜿ããïŒç§ææ¿çã®ç«æ¡ãåŠè¡çã®ä¿¯ç°ïŒãµããïŒçåæãæ¯æŽããïŒèªç¶èšèªåŠçæè¡ã§ç ç©¶éçºã®é¡äŒŒåºŠãç®åºããïŒç¹æ®ãªã€ã³ããã¯ã¹ãçšããŠïŒé¡äŒŒããããªãã®ã ãã«çµã£ãŠé¡äŒŒåºŠãèšç®ããïŒæ¢åæè¡ã¯å
šãŠã®å
容ã«ã€ããŠé¡äŒŒåºŠãæ±ããŠããïŒå
ã«çµã蟌ãããšã§èšç®è² è·ãæããïŒäŸãã°ïŒä»¶ã®ç¹èš±ã«å¯ŸããŠïŒïŒïŒä»¶ã®é¡äŒŒç¹èš±ãæ±ããã®ã«ïŒããªç§ãšé«éã§èšç®ãçµããïŒéå»ïŒïŒå¹Žåã®ç¹èš±ïŒïŒïŒäžä»¶ã«å¯Ÿãé¡äŒŒåºŠãæ±ããã®ã«ïŒæéïŒïŒåã§æžãïŒ
æè¡ãšããŠã¯ïŒç¹èš±ã®æç« ãæ°å€ãã¯ãã«ã®åæ£è¡šçŸã«ããŠïŒé«æ¬¡å
ãã¯ãã«è¿åæ¢çŽ¢ãšããæè¡ã§è¿ããç¹èš±ãæ¢ãïŒç¹èš±å
šäœã§æ°å€ãã¯ãã«ãçŽïŒïŒïŒæ¬¡å
ã«æããããé«éåŠçã§ããïŒç¹èš±å
šäœã俯ç°ããŠåæã§ããïŒã
å
·äœçãªäœ¿çšäºäŸãå
¬éããŠããããããã®ïŒ
ã©ããïŒããæéå
ã®å
šç¹èš±ããã¯ãã«åããŠä¿åãæ€çŽ¢å³ç€ºããŠããããã ãïŒãã®å Žåã¯ããã®æ¯éå£ã«ãããããšããèŠç¹ïŒäŸãã°å
±é䟡å€ïŒæå³ïŒç¶æ
ïŒå±æ§ãªã©ïŒãå
¥ããªããªãïŒãããŠä»»æã«æ¯éå£ãèŠå®ããããšã«ãã£ãŠïŒäœã«ã€ããŠåæãããã®ãã®æ
å ±ãäžããããïŒããèŠç¹ã§ã¯ãã2ã€ã®ç¹èš±ã¯é ãã¹ãã ãïŒããèŠç¹ã§ã¯è¿ãã¹ãã ïŒãã®æ
å ±ã倱ã£ãŠãããã®ææ³ã§ïŒç®çãéæã§ããã®ããª?ïŒ
ããã®æ¯éå£ã«ããããã¯éåžžã«éèŠïŒvaluebexãæ¯éå£ã«ãããŠåçãªtfidfã䜿ãç¶ããŠããããšã«ã¯æå³ãããïŒãšæã£ãŠããïŒïŒä»ã®ããŒãžã§æžããŠããããã«ïŒèªåã¯ãã®ãèªç¥ããéåžžã«éèŠããŠããïŒïŒ
ïŒããèŠç¹ãäœãã«ã¯å¹Ÿäœãå¿
èŠïŒ
ãéåãïŒç®çã«åããã©ã®ãããªå¹Ÿäœã«æŽãããããæéèŠã§ããã®ã«ïŒãããæ°ã«ããŠããªãå³ç€ºãå€ãããã«æããïŒè¿ãããããã§ãããïŒã§ã¯å³ç€ºãè§£æã«çšããã«æããã«è¶³ããªãïŒãšæãïŒïŒamanogawaã¯éºäŒåèŠç¹ã«åºå®ããŠããããããç®çã«ç¹ãã£ãŠããåé¡ãªãïŒæ±çšç®çãªã®ã«åçãªèŠç¹ã§ãªããã®ã«ã€ããŠïŒãã®å³ç€ºã§äœããããã®ãçè§£ã§ããªãïŒçµ¶å¯Ÿçãªè¿ããªãã°ïŒ1次å
ã®ãªã¹ã圢åŒã§ã®åºåãããè¯ããšæãã®ã ãïŒäŸãããªãïŒPatenrfieldã®ææ³ãè¯ããšæãïŒïŒ
ïŒã¹ããã¯ããŒã¯ã補ååã»äºæ¥åãå éãã-ç ç©¶éçºéšéã®ãã¥ãŒã¹æŽ»çšè¡
https://webinalist.com/events/buisiness/18089/
ïŒAnewsã®è£åŽã§åãïŒèªç¶èšèªåŠçãæŽ»çšããããžãã¹ãã¥ãŒã¹ã®æšèŠã·ã¹ãã
https://tech.stockmark.co.jp/blog/20210601_anews_recommendation/
Anewsã®æ§ææè¡ãštfidf/mbart/mt5 embeddings cluster visã®åºæ¬æè¡ã¯èãæ¹ãã»ãŒåãããªïŒ
èªåã¯ä»ã®ææ®µã§ã¯éã¿ä»ãããŠããã®ã§tfidf/mbart/mt5 embeddings cluster visã§ã¯åãæšãŠãŠããŸã£ãïŒãã®ç¹ãç°ãªãïŒãŸãïŒtfidf/mbart/mt5 embeddings cluster visã§ã¯æ¯èŒçãã¡ã€ã³ãéå®ãããããšãããïŒåãæšãŠãåŸã«æŽã«ã¹ãããã¯ãŒãåŠçãè¡ããã粟床ãé«ããªãããã«ããŠããïŒ
A Simple but Tough-to-Beat Baseline for Sentence Embeddings
https://openreview.net/forum?id=SyK00v5xx
SIFãå²ãåœãŠããããã¯ãã«ãåèªã®éèŠåºŠã«æºããŠéã¿ä»ããïŒè¶³ãåããã
ïŒãTechã®éãäžæ©ããã第40åãTexthero ã§æ¥æ¬èªãè§£æããã
https://buildersbox.corp-sansan.com/entry/2021/06/29/110000?utm_campaign=Weekly%20Kaggle%20News&utm_medium=email&utm_source=Revue%20newsletter
ãTF-IDF è¡åã PCA ã§å¯èŠåãããããã 2次å
ã«æ¬¡å
åæžãïŒK-means çµæãšãšãã«ããããããŸãïŒã
TextheroïŒçããŸãšãŸã䟿å©ããïŒ
ïŒæåæã泚åãããç¥è²¡ã®DXãïŒç¹èš±åæã§ç«¶åä»ç€Ÿã®æŠç¥ãèªã¿è§£ã
https://active.nikkeibp.co.jp/atcl/act/19/00313/091400007/
䜿çšäŸã®è£è¶³
ïŒäŸã§ã¯ã€ã³ã¿ã©ã¯ãã£ãHTMLããé¡äŒŒç¹èš±ãæ¢ããŸãããïŒãèšäºæç« ãå
¥åããŠãããïŒèšäºæç« ãã¯ãã«ãšåç¹èš±æç« ãã¯ãã«ãšã®é¡äŒŒåºŠã®æ°å€ããé¡äŒŒç¹èš±ãæ¢ãããšãå¯èœã§ãïŒ
ïŒæŠå¿µæ€çŽ¢æ©èœã远å ïŒãèšäºæç« ãäºãå
¥åããŠããæ©èœãã¯æŠå¿µã®äžäœäžäœã¬ãã«ã®éžæãã¢ã«ãŽãªãºã ã«å§ããå Žåã«äœ¿çšãïŒãæŠå¿µæ€çŽ¢æ©èœãã¯äºåŸçã«ç¢ºå®çãªæŠå¿µãçšããŠçµã蟌ã¿ãè¡ãããå Žåã«äœ¿çšãããšè¯ãïŒïŒïŒãattentionæŠå¿µæ€çŽ¢ãæ€èšäžã»ã»ã»ïŒ
ïŒé¡äŒŒç¹èš±ãå€ããå³äžã§å€æãé£ãïŒã©ã®ç¹èš±ã確èªãããå¿ããŠããŸãïŒæ¬¡å
åæžåã®é¡äŒŒåºŠã§æ¯èŒãããïŒãªã©ã®å Žåã«ã¯ïŒãé¡äŒŒåºŠãªã¹ããã䜿ããšè¯ãã§ãããïŒ
ïŒçµãèŸŒã¿æ€çŽ¢ãè¡ã£ãåŸã«é¡äŒŒåºŠãªã¹ããäœæããããšã«ããïŒãã粟床ã®é«ããªã¹ããšããããšãã§ããŸãïŒïŒïŒé¡äŒŒåºŠãªã¹ããšãããšç¹æ§ã®æ¹åæ§ãªã©å€é¢çãªé¡äŒŒæ
å ±ã倱ãããŠããŸãããšã«æ³šæããŠãã ããïŒïŒ
ïŒæ¢ãããç¹æ§ã瀺ãã¯ã©ã¹ã¿ãç¡ãã©ã®æ¹åãæ¢ãã°ããã®ãããããªãå Žåã«ã¯ïŒãããŒã¯ãŒãéå®ããåé¡éå®ãïŒãæ¯éå£ã«ãã®ç¹æ§ã瀺ãç¹èš±çŸ€ããããŒïŒåŒã³æ°ŽïŒãšããŠå ããŠãã¯ãã«èšç®ãããã®ã¡ãããŒãæ¶ããïŒãªã©ã®å¯Ÿå¿ãè¡ããšè¯ãã§ãããïŒ
ïŒãããå©çšãäŸãã°ïŒè£œè¬äŒç€Ÿå
šäœã®æ
å ±ã調æ»å¯Ÿè±¡è£œè¬äŒç€Ÿã®æ
å ±ã«æ··ããã¯ãã«èšç®ãã補è¬äŒç€Ÿå
šäœã®æ
å ±ãé€ãããšã«ããïŒèª¿æ»å¯Ÿè±¡è£œè¬äŒç€Ÿãæ¡çšããŠããªãæŠç¥ãæ¥æãæœåºããïŒãªã©ã®äœ¿ãæ¹ãããåŸãã§ãããïŒãã©ã¡ãŒã¿ã倿Žãé¡äŒŒã®æœè±¡åºŠãå€ãããäžäœæŠå¿µã§æ··åããããšã«ããèŠãããã®ããããããããŸããïŒïŒïŒãããã§ãããããã©ïŒç¹èš±é èœäŒç€Ÿã®äºæž¬ã«ã䜿ããã§ãããïŒïŒïŒã¯ã©ã¹ã¿ãªã³ã°ã«ãããŠã¯æ¯éå£ã®å
容ã¯éåžžã«éèŠãšãªããŸãïŒéåžžã®ç¹èš±åæãªã©ã®ã»ã°ã¡ã³ããŒã·ã§ã³ãšåãæèŠã§æ¯éå£ããšãããŠã¯ãããŸããïŒåšèŸºæè¡ãå«ãã æ¯éå£ãšããããšã¯ïŒéåžžã«éèŠã§ãïŒïŒïŒèªç±ã«ä»»æã®èŠç¹ãšã§ããããã«ããããšã¯éèŠã§ãïŒä»»æã®èŠç¹ã®ã¿ãšããŠããŸãããšã¯é¿ããã»ããè¯ããšæããšããã§ãïŒïŒïŒé¢é£æ¯éå£ã远å ããããšã§ïŒæ°ããªèŠç¹ã®ã¯ã©ã¹ã¿ãäœæãïŒæ°ããªã»ã°ã¡ã³ãã«ã€ããŠã®ç¥èŠãåŸãïŒãªã©ãé¢çœããšãhiãã§ãïŒïŒ
ïŒããŒã¯ãŒãã§ãåé¡ã§ãèŠã€ãã¥ããç¹èš±ã«ãããŠç¹ã«æå¹ã§ãããïŒè¿å¹ŽéèŠã®é«ãæ°èŠåéã®ç¹èš±ã§ã¯ããŒã¯ãŒããåé¡ãå®å®ããŠããªãã§ãããããæå¹ãšãªããããã§ãããïŒ
ïŒæªç¥ããŒã¯ãŒãã§ããããšååŸé¢ä¿ãèæ
®ããä»åèªãšã®é¡äŒŒæ§ããé©åã«å³ç€ºãããŸãïŒåæ£è¡šçŸã䜿ã£ãŠããã°åœããåã§ããïŒãã ãïŒããŸãã«æªç¥ã§ããå Žåã«ã¯ïŒæªç¥ããŒã¯ãŒãå士ã§é¡äŒŒæ§ããããšèšç®ãããŠããŸãããšããããŸãã®ã§æ³šæå¿
èŠã§ãïŒïŒ
ïŒããããã®ç¹ã¯ïŒèšå®ãããã©ã¡ãŒã¿ã«äŸåããé¡äŒŒææšã«åºã¥ãè¿æ¥ããŠããããïŒãã®é¡äŒŒææšããå€ããèŠç¹ã«ãããŠã¯ïŒè¿æ¥ããŠããç¹ã§ãã£ãŠãé¢é£æ§ãäœãããã«èŠããããšããããŸãïŒäžäœæŠå¿µãå
±éãããäžäœæŠå¿µã¯ç°ãªããªã©ïŒïŒ
ãã®å Žåã«ãããŠãïŒãé¢é£æ§ãäœãèŠãããã©ãã«å
±éæ§ãããã®ãæ¢çŽ¢ãããããšããèŠç¹ã§èŠãããšã«ããïŒé£æ³èšæ¶ã«åºã¥ããæã蟌ã¿ãé²ãïŒæ°ããªã€ã³ãµã€ããçã¿åºããããšãã§ãïŒãæè¡ã®äœ¿ããæ¹ã®é©æ°ã§ãããã€ãããŒã·ã§ã³ããã«å¯äžããïŒãå
±éã®ç®çã«åãã£ãŠå調ããã«èªç«ãè¡åãããå
±åµããã«å¯äžããïŒåµé æ§ãçã¿åºãåµé ã«å¯äžããïŒãããããŸããïŒ
ïŒãé¢é£æ§ã®äœã2ã€ã®å¯Ÿè±¡éã§ã®å
±éç¹ã®æ¢çŽ¢ã¯å¯Ÿè±¡ã®ãç®ç«ããªããç¥èã®æŽ»æ§åãä¿é²ããïŒããåºå®çãªèŠæ¹ã®è§£æ¶ãšã¯ãç®ç«ããªããç¥èãæŽ»æ§åãããããšïŒããå±±å·çç±ã»æž
河幞åãé¢é£æ§ã®äœã察象éã®å
±éç¹æ¢çŽ¢ããã»ã¹ãŒã«ããŽãªå€æèª²é¡ãšã®é¢é£ã«ããæ€èšãæ¥æ¬èªç¥ç§åŠäŒç¬¬37å倧äŒãJCSS2020
httpss://www.jstage.jst.go.jp/article/jcss/27/4/27_2020.002/_pdf/-char/ja ïŒïŒç®ç«ããªãç¹åŸŽã掻æ§åããïŒããã«èªãã®ãã€ã¢ã¹ããç¹åŸŽã®ééãè£ã£ãŠïŒçæããŠïŒè¶³ãããšãããïŒæ€çŽ¢è
ã調æ»è
ã®åœ¹å²ïŒãšïŒèªç¥èŠç¹ã®æ€çŽ¢è«ã«åºã¥ãèšãåã£ãŠããããšæãïŒïŒïŒããã«ããçæïŒãDRMãã©ãã€ã ãšã¯â¢çžäºã«æå³çé¢é£ã®ããåèªãªã¹ããå瀺ãïŒåŸã«èªç±åçãããããšïŒå®éã«ã¯å瀺ãããŠããªãæå³çé¢é£ã®åŒ·ãåèªã誀ã£ãŠåçããã(Deese, 1959) ïŒïŒéçºè
ã¯ã€ãããŒã·ã§ã³ã«åããªãïŒçŒã®åã«ããå®çŸå¯èœæ§ã®é«ãã«ææãããããïŒããšã¯æ§ã
ãªããŒã¿ã«ãã瀺ãããŠããïŒããã§äŒç»ãå©çšãããããšããããïŒäŒç»ã¯ãã¬ãŒã ã«ææããããã€ãããŒã·ã§ã³ã«ãã©ãçããªãããšããããïŒçµå±ïŒçŒã®åã«ããå®çŸå¯èœæ§ã®é«ãã¢ã€ãã£ã¢ããæ¡çšããã«ããïŒæ§ã
ãªå£ãã³ã¹ãïŒçã¿åºããåŸã䟡å€ãèæ
®ãããšïŒéçºè
èªããã€ãããŒã·ã§ã³ãçãããšãéèŠãšãªãïŒïŒABCDEFã®åæ
ãšããèãæ¹ããããïŒåæ
ããã»ã©äŸ¡å€èгåãããªã©ã®å£ãé«ãããªãçã¿åºããã䟡å€ãå°ãããŸãšãŸãåŸãã®ã§çµ±åã§ãããšããã¯çµ±åããã»ããè¯ãïŒå€æ§æ§ããèªå·±å€æ§æ§ïŒã«äŒŒãïŒïŒïŒéçºè
ã¯ã€ãããŒã·ã§ã³ã«åããªããšèªèããããã§ããã«ã€ãããŒã·ã§ã³ãèµ·ãããïŒãã®çãã®äžã€ã¯ïŒèªåŸçã«å€æ§ãªæ
å ±ãåéããããšã§ããïŒããã«ç€ºãããŒã«ã¯ïŒéçºè
ãèªäž»çã«å€æ§ãªæ
å ±ãåéãã€ã³ãµã€ããåŸãããšãã§ãïŒéçºè
ã«ããã€ãããŒã·ã§ã³çæã«å¯äžããã ããïŒïŒïŒé¢çœããšã¯ãã¿ãŒã³ã®çºèŠã§ããïŒé¢çœããå ±é
¬ãšããŠç¥ç奜å¥å¿ã圢æãããïŒç¥ç奜å¥å¿ã¯èªç«ã®æºæ³ã§ããïŒã€ãããŒã·ã§ã³ã®æºæ³ãšèããŠããïŒïŒïŒãã³ãããã¢ã«ãŽãªãºã ïŒæ¢çŽ¢ãšæŽ»çšã®ãã¬ãŒããªããïŒã€ãããŒã·ã§ã³ããã¿ãéçºã®äžèª å®ããšéçºããèŠãã€ãããŒã·ã§ã³ã®äžèª å®ãã®ãžã¬ã³ããšå¯Ÿå¿ããŠããããã§é¢çœãïŒãžã¬ã³ãããã¬ãŒããªãã«å¯Ÿå¿ããããïŒã€ãããŒã·ã§ã³ã«ãå ææ¢çŽ¢æšè«ã®ã¢ã«ãŽãªãºã ãå°å
¥ãã¹ããªãããªäººä»»ãã§ã¯ãªãïŒããŠã©ãçµã¿èŸŒããïŒïŒïŒç§ã¯ïŒåæè
ã®æ§é åèœåã«åŒ·ãäŸåããããã¹ããã€ãã³ã°ã«ããå€ç©ºéçºæ³æ³ã®ãããªã¢ã€ãã£ã¢çæææ³ãããïŒé²èЧè
ãããããèªèº«ã®èŠç¹ãšèœåã«åºã¥ãã¢ã€ãã£ã¢ãçæããã¯ã©ã¹ã¿ãªã³ã°ç³»ã®ã¢ã€ãã£ã¢çæææ³ã奜ãïŒåŸè
ã®ã»ããããåºãå¯èœæ§ãå«ããšæãããïŒé²èЧè
ã®èœåã«äŸåãããªãã»ããè¯ãïŒåæè
ã®ã»ããåªç§ã ãšè«Šãããšãã®ã¿ïŒåè
ãæ€èšããã ããïŒïŒïŒçè§£äžå¯èœãšèªããäžã§ãã®ãŸãŸåãã ãïŒã»ã°ã¡ã³ããŒã·ã§ã³ã§ã¯ãªãïŒã¯ã©ã¹ã¿ãªã³ã°ææ³ãïŒè€éãªé ãæ
å ±ã䟡å€ã®ããããåãåºãæå¹ãªææ®µã ïŒãšããèŠç¹ã¯ïŒã€ãããŒã·ã§ã³ã«é¢é£ãããšæãããŸãïŒæ¬ææ³ã¯é ãæ
å ±ããã®ãŸãŸäœ¿ã£ãããã§ã¯ãªãã§ããïŒïŒ
ïŒéåãããŒã¯ãŒããåé¡ã«ããã©ãå€åããã®ãæ¯èŒããããšã«ããïŒã粟床ã®é«ãéšåéåãçµã¿åããåçŸçã®é«ãéåãäœãã«ã¯ã©ãããã°ããã®ãã«ã€ããŠãã¬ãŒãã³ã°ããããŒã«ãïŒæ€æ»èœååäžèšç·ŽããŒã«ãšããŠã䜿ãããããããŸããïŒ
ïŒæ¯éå£ã®ç¹æ§ãéèŠãšã®èгç¹ã«ãããŠïŒãããŒãå©çšãã€ã€å°çšåããã°ã·ã¹ãããã£ãã¯ã¬ãã¥ãŒïŒã¡ã¿åæã«ã䜿ãããããããŸããïŒç¹èš±SDIçšAIãšåãããã«ä»æè¡ãšçµã¿åãããŠã¿ãããâŠïŒ
äœ¿ãæ¹ã®è£è¶³
ïŒæç®ããã¥ãŒã¹ã»å€äŸã»ä»èšèªã«ã察å¿ããŠããŸãïŒ
ãã ãïŒé©åœãªååŠçãè¡ãïŒé©åœãªã¹ãããã¯ãŒãã远å ããå¿
èŠãããã§ãããïŒ
ïŒã¹ãããã¯ãŒãã¯çµæã確èªãã€ã€è¿œå ããã»ãããããããããšåããŸãïŒå人ççµéšããè¿°ã¹ããšïŒæåããåè©ã®ã¿ã«éå®ãããããªéå®ã¯ãããã»ããè¯ãã§ãããïŒã¢ã«ãŽãªãºã ãšæ¯éå£ã«é©åããäžå¯§ãªã¹ãããã¯ãŒãèšå®ã«ããããå€ãã®äŸ¡å€ãæ®ããã§ãããïŒïŒ
ïŒè€æ°èšèªåæè¡šç€ºã«ã¯å¯Ÿå¿ããŠããŸããïŒ
Aylienã«æåŸ
ãããšãããããããšæŸãæãã
https://t.co/7LkXI0viQY?amp=1
ïŒmbartãmT5ã®ãããªend-to-endã®å€èšèªã¢ãã«ãèŠãããããã«ãªã£ãŠããŸããïŒããããåŸããã¯ãã«ã¯è€æ°èšèªåæè¡šç€ºã«å¯Ÿå¿ããã§ãããïŒãšããããmbartè§ŠãäžïŒã©ããããã¯ãã«ãåãããããããŸãããïŒç¿»èš³èšèªãåºå®ãå
šç¿»èš³åèªã®åèªåœåºçŸç¢ºçã®èŠçŽ åãåãã°å
ã®èšèªã«äŸåããªããã¯ãã«ãšããŠåãåºãããšãã§ãããšæã£ãŠããŸãïŒå°ã
次å
ã倧ããããã®ã§æ¬¡å
åæžã¯å¿
èŠã§ãããããã©ïŒïŒïŒè±èªã®åºæãã¯ãã«ã𿥿¬èªã®åºæãã¯ãã«ãªã©å¿
èŠãªèšèªã®åºæãã¯ãã«ãæ±ããŠããã°ïŒçå±ã®äžã§ã¯ïŒèšèªéäŸåã®ãã¯ãã«ãæ±ãããïŒè€æ°èšèªåæè¡šç€ºãã§ããã¯ãã§ãïŒretrofitting?ïŒã©ããã§è©ŠããïŒïŒ
ïŒèŠçŽéçšãçµãããïŒããçšåºŠé·ãæç« ã«åããŠããŸãïŒç®å®ãšããŠã¯400åèªä»¥äžã§ãïŒ
ïŒç¹èš±ã§çšããããšããŸãæ³å®ããŠããŸãã®ã§ïŒïŒããã¹ãå
ã§å
容ãåäžæ§ã瀺ãããšãä»®å®ããŠããŸãïŒ
䜿çšãããããã¹ãããã®ä»®å®ã«åèŽããªãå Žåã¯ïŒãããããåäžæ§ã瀺ãã»ã³ãã³ã¹ã«åãåããŠãããªã©å¯Ÿå¿ãã ããïŒ
ïŒæ§ã
ãªãœãŒã¹ããéããããã¹ããæ··åããå ŽåïŒãããããã¹ãéã§ã¯æç« ã®è§£å床ãããé¢ããŠããããšãå€ãããïŒé©åã«å³ç€ºããé£ããªããããããŸããïŒ
å¯Ÿå¿æžã¿ã§ããïŒå®å
šãšã¯èšãé£ãã§ãããïŒ
ïŒè§£å床ã®éããåžåããæç« ãã¯ãã«äœæææ³ãææ¡ãããã®ãåŸ
ã¡ãŸãããïŒtransfomerããæ±ããæç« ãã¯ãã«ã§ããã°ïŒãã§ã«è§£å床ã®éããåžåã§ããŠãããšèšãããããããŸããïŒé©åœã«çœ®ãæããŠãã ããïŒïŒïŒLDAãçšããããã¯ãä»ããŸãšããããšã§ç¹èš±ãšæç®ã®çšèªã®å·®ã¯åžåã§ãããšã®ããšïŒvaluenexãæ°æ©èœãšããŠæç€ºãããã§ãïŒåŸ
ã¡æ¡çšããã®ãè¯ãã§ãããïŒïŒ
ïŒè€æ°ã®æœåºææ³ãçµã¿åãããææ³ãããè¯ãã§ããããïŒ
ã³ãŒãæŠèŠè£è¶³
ïŒæœåºçæç« èŠçŽãšããåéã®ææ³ãšããããããïŒ
ïŒæåž«ãªãããŒã¿ããåŠç¿ããåèªã¯ã©ã¹ã¿ãªã³ã° (Brown clustering)ã®ç¹åŸŽã远å ããŠæåž«ããåŠç¿ããâœ
æ³è«ãææ¡
https://www.aclweb.org/anthology/N04-1043/NAACL-2004
Name Tagging with Word Clusters and Discriminative Training
https://www.aclweb.org/anthology/N04-1043/
ïŒword2vecãštf-idfãçµã¿åãããäŸ
ïŒïŒïŒ
ïŒããçããæ±ããã«ãããïŒæ°çŸã®ïŒè€æ°ã®è§£æ±ºææ®µãèãããããšãïŒãã®ã«ãŒããããããã«éå®ããããšãããïŒèªç¥é©åãšåªç§ããäž¡ç«ããïŒç¥èœã®åœ¹å²ã ãšããã°ïŒïŒã®ãããªã¯ã©ã¹ã¿ãªã³ã°ãããæ¬è³ªã«è¿ããšèããŠããïŒ
ïŒçµæãšããŠã¯ã©ã¹ã¿ãªã³ã°ãããå ŽåãšäŒŒãã«ãŒããã¯ã©ã¹ã¿ãªã³ã°ããŠããå Žåã®å·®ã¯âŠïŒïŒãã€ãºã§éå®ããŠãè¯ããæ§é ãé©çšããŠéå®ããŠãè¯ãïŒéåžžã«é¢çœãïŒïŒ
ïŒæžç®ãšçž®çŽã®èŠç¹ãããããšçž®çŽã«è©²åœãããšæããïŒæ¬äœã®ç¹èš±SDIçšAIã§ã¯æžç®ãè¡ã£ãŠããïŒãšæãïŒã®ã§ïŒæ°ã«ã¯ããŠããªãïŒïŒæžç®ãšçž®çŽã«ã€ããŠïŒæªã ã«çè§£ããŠããããã§ã¯ãªããïŒå¯éå§çž®ãšéå¯éå§çž®ãšèãããšããã ããã?ïŒïŒïŒçž®çŽã¯ããã°ç®ç«ããªãç¹åŸŽãæ¶ãå»ã£ãŠãããšãèšããããã ãïŒãããŠæ¶ãå»ãããšã§ïŒæ€çŽ¢è
ã«ããçæãä¿ã广ã匷ããããŠããïŒãšãæããªãããªãïŒïŒ
ïŒåæ§æã«ããïŒtfidfã§ã¯é¿ããããªãèè
ã®çã«ãã圱é¿ãäœæžãããŠããïŒã¹ã¿ã€ã«ã»æäœãæ¶ãïŒç°ãªãè§£å床ãäž»èŠåèªã«éå®ããïŒãšè¡šçŸããŠããããããããªããhttps://qiita.com/kzuzuo/items/b6875441d7103ee515c1 ïŒ
ïŒåæ§æã®èãæ¹ã¯ïŒãµã³ããªã³ã°äžã®æã代衚çãªåèªãæ¡çšãããšãããµã³ãã«äžã®ååžãèæ
®ããminimun bayes risk MBRãšããææ³ã«äŒŒãŠããããã ïŒself-attention LSTMãªã©ã§æ¡çšãããŠãããšã®ããšïŒ https://arxiv.org/abs/1805.04833 ãåŸåã¯reformerãšåãèãæ¹ã?ïŒãŸãšãããšïŒMBR&LSHãšããã¢ã«ãŽãªãºã ã ãšè¡šçŸã§ãããïŒïŒ
ïŒïŒäžè¬åã§ããªãã®ã§ãã®ä»ãžç§»åïŒãæåž«ãããã«ããæç« ãã¯ãã«çæã¯ïŒè·é¢ãããæ¹åã«ãããŠç©æ¥µçã«è¿ã¥ãïŒããã«ããä»ããé¢ããïŒé¢ãéã«ïŒå
åšçãªæ¹åæ§ã瀺ãæ
å ±ãã倱ããããšãããïŒ
ããæåž«ãªããã«ããæç« ãã¯ãã«çæã¯ïŒè·é¢ãæ¹åã®éãããèŠåºããïŒå
åšçãªæ¹åæ§ã瀺ãæ
å ±ã¯å€±ãããŠãããïŒåã«èŠãã ããªãã£ããšè§£éãããïŒ
ãæ¬ææ³ã®ãããªæç« éé¡äŒŒå¯èŠåã«ãããŠïŒäºãå®ããä»»æã®æ¹åãæç€ºããããšãå¿
èŠãªã®ã§ã¯ç¡ãïŒä»®èª¬ãåæãšãã€ã€ãèªç±ãªã€ã³ãµã€ããå¿
èŠãªã®ã§ããã°ïŒæåž«ãªããæé©ã§ãããšèããŠããïŒãã ãïŒãã®å Žåã¯èŠåºãããã®é³åïŒçªïŒãããžã§ã¯ã·ã§ã³ïŒãªããšåŒãã§ãè¯ãããã©ïŒãåé¡ãšãªãïŒ
ã³ãŒããšæ¹è¯
ïŒ3000è¡+
ïŒMITã©ã€ã»ã³ã¹ã§Githubã«äžããïŒãŸãã¯æ¬²ãããšææè¡šæããäººã«æ¹è¯æã©ã€ã»ã³ã¹ããã¯å¥çŽã§éãïŒäºå®ïŒïŒäžèšã³ãŒãæŠèŠãåèã«èªäœããã»ããè¯ããšãæãïŒçé¢ç®ã«æžãã°2000è¡ãããã§æžããã¯ãïŒïŒ
ïŒæç« ãã¯ãã«ãšããŠBERTãã¯ãã«ïŒCLSã§ã¯ãªãsentence-BERTïŒïŒã¯ã©ã¹ã¿ãªã³ã°ææ³ãšããŠDBSCANïŒPHATEãéžæã§ããããã«æ¹è¯äžïŒ
ïŒTfidf embeddings/cluster visã¯ããããBERTãå«ãèªäœAIã®è£æ£ã®ããã«äœæããAIã·ã¹ãã ã®äžéšïŒéè€ãè£æ£ã®åœ¹ã«ç«ããªãã ããBERTãã¯ãã«ãæ¡çšããã¢ãããŒã·ã§ã³ã¯ããŸããªãïŒãšã¯ããéžæã§ãããšäŸ¿å©ã ããïŒïŒ
ïŒçµ±åæŠå¿µæ€çŽ¢æ©èœã远å ïŒå
¥åããåèªçŸ€ãšè¿ãæŠå¿µã®ç¹èš±ãæœåºå¯èœãšããïŒïŒé¡çŸ©èªãšèšèŒã®å€æ§æ§ãåžåïŒïŒ
ïŒç¥è²¡é«è£å€æ±ºãäºç¹åäœã§æ€çŽ¢ãããã¢ãé£äŒäžã«äœã£ãŠã¿ãã
âå€äŸäºç¹åäœæŠå¿µæ€çŽ¢ã·ã¹ãã
ã倿±ºPDFâããã¹ãâäºç¹åºåãâæç« ãã¯ãã«åâhtml
ãã»äžéãå®è¡å®äºïŒ
ãã»äºç¹ãããŒã¯ãŒãã§éå®ããã€æŠå¿µã§ãéå®ã§ããããã«ããïŒäŸãã°ïŒå»è¬ã«ãããã鲿©æ§ã«ã€ããŠã®äºç¹ã«éå®ãé¡äŒŒåºŠãåºåã§ããããã«ããïŒäžå³åç
§ïŒ
ãæ®ããã課é¡
ãã»åºåäœçœ®èŠçŽã
ãã»å€æ±º12000ä»¶åäºæ³èšç®æé15æ¥
ãã»äžèŠèªé€å»
ïŒç䌌self-attentionããšãïŒãã®äžäœæ°åèªãæŽã«è¶³ãåãããã®ãããããªïŒããããã«å®è£
ã§ããã詊ããŠã¿ããïŒ
ïŒäžèšã®å€äŸäºç¹åäœæŠå¿µæ€çŽ¢ã·ã¹ãã ã調æŽããŠããŠæã£ãã®ã ãïŒkeyword attention圢åŒãšããã°ïŒäºç¹åºåããããå¿
èŠãç¡ããªãããåºåãäœçœ®ãã¹ã®åœ±é¿ãåããã«é¡äŒŒèšç®ãã§ããããã«ãªãïŒããã«self-attentionãã軜ãã§ãïŒããè¯ãã®ã§ã¯ãªããïŒ
æ§æèŠçŽ åºåãäœçœ®ãåé¡ãšãªãæ§æèŠçŽ åäœé¡äŒŒæ€çŽ¢ã§ãåã課é¡ããããªâŠå®è£
ããŠã¿ãã
âãattentionæŠå¿µæ€çŽ¢ã
ïŒãªãïŒæ§æèŠçŽ åäœé¡äŒŒæ€çŽ¢ã§åŸãããã®ã¯ãããŸã§ããã¢ã«ãŽãªãºã ã«åŸã£ãæ§æèŠçŽ åäœéã®é¡äŒŒã®ã¿ïŒæ§æèŠçŽ ãé¡äŒŒããŠãããããšèšã£ãŠãããçŽã¡ã«é²æ©æ§ãªã©ã«é¢é£ããå
è¡æç®ãšãªãåŸãããã§ã¯ãªãïŒããïŒé²æ©æ§ãªã©ã«é¢é£ããå
è¡æç®ãèŠã€ããããšãããªãã°ïŒé¡äŒŒãšã¯å¥ã®ãããªãã¢ãããŒããå¿
é ãšãªãïŒãã®ã¢ãããŒãããªãå Žåã¯ïŒé²æ©æ§ãªã©ã«é¢é£ããå
è¡æç®ã®çºèŠã¯æ€çŽ¢è
ã®åœ¹å²ãšãªãïŒïŒ
åå¥attentionæŠå¿µæ€çŽ¢ã«ä»ãïŒæŠå¿µãç¥èã°ã©ãããæã£ãŠããã°ïŒããæå³zero-shotåŠç¿ã«ããæ€çŽ¢ãã§ãããšãããã®ããªïŒïŒåŒã³æ¹ã¯ãšãããå®çŸå¯èœã ãªã»ã»ã»
ïŒæç¢ºãªé¡äŒŒãšã¢ã«ãŽãªãºã äžã®é¡äŒŒãšã®çžéç¹ã®åèçãšããããïŒåŒçšè¢«åŒçšã®é¢ä¿ã«ããç¹éã«ç·ã远å ããäºå®ïŒ
åèãhttps://stackoverrun.com/ja/q/11636832
â被åŒçšå®è£
ïŒå€æ¹åããç·ãåŒãããŠããç¹èš±ã¯å€åéã«åœ±é¿åãããç¹èš±ã ãšãããã®ã ããïŒç¹å®æ¹åãžã®ç·ã®éäžã¯åééã®é¢é£ã®æ·±ãã衚ããŠããã®ã ããïŒäžèšå³ã¯ããç¹èš±ã«é¡äŒŒããç¹è𱿝éå£ãäœæããããå³ç€ºãããã®ïŒé¡äŒŒããç¹èš±çŸ€ã§ããããåœç¶ã«åŒçšé¢ä¿ã«ãªãããšãå€ãïŒç·ã®å¯åºŠãé«ãã倿ãé£ããïŒãã°ãã芳å¯ããŠã¿ãïŒ

ïŒæ··åã¬ãŠã¹ã¢ãã«ãæ¡çšãæœåºå¯èœãšããäºå®ïŒ
ïŒmBARTãªã©ã®èŠçŽæè¡ãå©çšããŠïŒãã©ãããªæŠå¿µã§ãå³ç€ºå¯èœãšããäºå®ïŒã€ãã§ïŒç¹èš±æç« ãšæç®ïŒååæ
å ±ïŒå€äŸã®ïŒããé©åãªåææå¹³é¢å³ç€ºã®æ€èšã詊ã¿ãïŒ
ïŒéèšèŒåºæè¡šçŸæœåºã£ãœãæ©èœãåæã§å®è£
ã詊è¡ïŒ5000ä»¶ã»ã©ã®åŠçã«100æéã»ã©ãããã©ããããã®ãïŒ
ïŒæŠå¿µè¿åæ€çŽ¢è¿œå ïŒ
ïŒè¿œå äºå®
DSOCïŒãžã§ã³ãºãããã³ã¹å€§åŠã®Angelo Meleåææãšå
±åç ç©¶ãããããã¯ãŒã¯è§£æã®ã¢ã«ãŽãªãºã ããªãŒãã³ãœãŒã¹ã§å
Ž
https://prtimes.jp/main/html/rd/amp/p/000000210.000049627.html
lighthergm
https://github.com/sansan-inc/lighthergm
ïŒNLP2022
ã»C3-3 補åç¹åŸŽã«åºã¥ã補åçºè¡šãã¬ã¹ãªãªãŒã¹ã®é¢é£ç¹èš±èªåå€å®
ããâäžå±±åªèŒ, é
äºæµ©ä¹, 氞䞊å¥åŸ (æè¹å€§)
éèŠèªãæœåºãåæ£è¡šçŸãšããïŒåºæ¬çãªèãæ¹ã¯åæããïŒ
ãã ïŒãã¬ã¹ãšç¹èš±ã§é¡äŒŒããèªã䜿çšãããŠãããšèšãçç±ã§ïŒç¹èš±ã®å¹æã宿œäŸãããŒã¿ãšããã®ã§ããããïŒç¹èš±å
šæãç°¡åã«å
¥æããããšã¯æ¯èŒçé£ããããå®çšãšããŠã¯ã©ãã§ããããïŒãŸãåéã«ãã£ãŠã¯å¿
ããã广ã宿œäŸã®çšèªããã¬ã¹ã®çšèªã«é¡äŒŒããããã§ããªãïŒè¡šçŸã®äžäœåäžäœåãè¡ã£ãŠããªãã£ãç¹ã¯æ®å¿µïŒ
FIãªã©åé¡ãã¢ãã«ã«å«ããªãã£ãç¹ã¯æ£ããã ããïŒã¢ãã«å€å®ã䜿ã£ãŠããåé¡ã䜿çšããã°è¯ãïŒåé¡ãã¢ãã«ã«å«ãããšïŒãã®åœ±é¿ã倧ãããïŒåé¡ã®ééãã«ã倧ããåŒãã¥ãããŠããŸãïŒæ°å€äžã®ç²ŸåºŠçãè¯ããªããšããŠãïŒåé¡ã§èŠããªããã®ãèŠããããã«ããå¯èœæ§ãç¡ãããŠããŸãïŒããã®åé¡ç¥èãç©æ¥µçã«ç Žæ£ã§ããã·ã¹ãã ãå ããªãå Žåã¯ãäŸãã°ãåé¡ããããã¯ãŒã¯åãïŒééããè¿äŒŒããŒãæ
å ±ã«ããè£æ£ã§ããããã«ããŠããªãå Žåã¯ãïŒããã¯ãªãïŒïŒåé¡ã䜿çšããå Žåã¯ç¥èãããã¯ãŒã¯ã«å€æããŠããå
¥ãããšè¯ãïŒåé¡ãééããŠãããšããŠãïŒå
šãééããŠããå¯èœæ§ã¯äœãïŒééãã¯é£ãåãããŒãã«çŸããã¯ãã§ïŒãããã¯ãŒã¯ã¯ãããè£æ£ããŠãããïŒïŒ
ã»ã»ã»selfã§ãªãïŒäžè¬ã³ãŒãã¹ã§åŠç¿ããã¢ãã«ã«éãçŽãïŒãã¯ãã«ã®é·ãã§éžæãïŒç¹èš±ãšäŸµå®³åã«å
±éããçšèªãæœåºããïŒãšèšãããæ¹ãããããªïŒCLIPéããããåŠçãæ©ãã ããïŒ
ã»A5-1 åèªãã¯ãã«ã®é·ãã¯æå³ã®åŒ·ãã衚ã
â倧山çŸã
å¢ (京倧/çç ), 暪äºç¥¥ (æ±å倧/çç ), äžå¹³è±å¯¿ (京倧/çç )
éåžžã«éèŠïŒ
ã»B6-1 察ç
§åŠç¿ã«ããæãã¯ãã«ã䜿çšããé害ã¬ããŒãã®ã¯ã©ã¹ã¿ãªã³ã°
âå°æåç, å±±äžéæµ·, 岡ç
§æ, å°çºå® (éœç«å€§), çéç« , è°·æ¬æé (å¯å£«é»æ©)
é害ã¬ããŒããæ
éã®ãç¶æ³ãåå ãæªçœ®
ããŸãéçŽããã
ãããããïŒæãã¯ãã«å
ããã¯ã©ã¹ã¿ãªã³ã°ããŠéçŽ
simCSEã®åŠç¿ãšéçŽïŒäºäŸåæ
ã»BERTãçšããæåž«ãªãæè¡šçŸ
https://tech.retrieva.jp/entry/2021/10/12/130850
ïŒtfidf/mbart/mt5/CLIP embeddings cluster visã¯ïŒã»ãŒçµ±èšãšååžä»®èª¬ã«ã®ã¿ã«åºã¥ããããattentionããããã¡ã€ã³äŸåãå°ãªããããïŒããèªç¥ã«è¿ããšæããïŒã€ã¡ãŒãžãšããŠã¯ïŒèªåãå«ãŸãªãä»è
ã®èãã§ã¯ãªãèãŸã£ãŠãããèªåã®èããå«ãããèªç¥ã«è¿ããšæããïŒãšãããŸããããïŒNLP2022ãèªç¶èšèªãè€éç³»ãšããŠæãã詊ã¿ãã®æ°æã¡ã«äŒŒãŠãããšãããŸããããïŒãŸãã¯ïŒããã¯éç·åœ¢ãªæèŠã«åºã¥ããŠãã®ã®è¯ãæªããå€å¥ããŠããïŒæ°äžã®ãã¡7ã€ã»ã©ãéžã¶ïŒãããŠãããçè§£ããããšãããšç·åœ¢ã§æããããšããŠããŸãïŒ7ã€ã®ã¿ãã倿ããããšããïŒãšããïŒããèªç¥ã«è¿ã¥ããã«ã¯ãã®åŸè
ã®7ã€ã®ã¿ããæ§æãããããªã¢ã«ãŽãªãºã ãè¯ãïŒãšãããŸããããïŒãããªã€ã¡ãŒãžïŒ https://note.com/4bata/n/n0a44276a0ef1 ïŒã
ãªãïŒã¢ãã«ãšããèªç¥ãšã®é©åãã©ãæ€èšŒããã°ãããæªã ã«ããã£ãŠããŸããïŒåºæ¬çã«ã¯è·é¢ã®ç·åã§è©äŸ¡ããã°è¯ããšã¯æããŸããbinæ¯ã®å·®ãéèŠãšãªãããâŠ
ïŒBM25
tfidfé²åç³»ïŒæç« ã®é·ãè£æ£ïŒæ€çŽ¢ã§ããçšããããïŒ
æç« å
ã§åèªã®ã©ã³ãã³ã°ãããéã«ã¯é¢ä¿ãªãïŒ
ïŒRe-evaluating Word Mover's Distance
https://arxiv.org/abs/2105.14403
https://github.com/joisino/reeval-wmd

æ£ããæ£èŠåãããšbag-of-wordsãWord Mover's Distanceã«è¿«ãããšãçºèŠ
Top2Vec learns jointly embedded topic, document and word vectors.
https://github.com/ddangelov/Top2Vec
2. UMAPã䜿çšããŠïŒããã¥ã¡ã³ããã¯ãã«ã®äœæ¬¡å
ã®åã蟌ã¿ãäœæããŸãïŒ
3.HDBSCANã䜿çšããŠããã¥ã¡ã³ãã®å¯éããé åãèŠã€ããŸãïŒ
âŠ
æ¯éå£ããšã«å¯å€ãªãã¯ãã«ã䜿çšããªãã®ã§ç®çã«åããªã
ïŒåæçµæãå ±æããããŒã«ã¯å€ã ãããŸããïŒããçšåºŠä»¥äžè€éãªè¡šçŸãšãªããšåãæã®èªç¥ã®åé¡ã«ããå ±æããŠãçè§£ãããªãããšãå€ãã§ãïŒKH Coderã®å¯Ÿå¿åæããããéçããšæããŠããŸãïŒïŒåæçµæãããïŒæ¢çŽ¢çãªä»®èª¬ãçå圢åŒã§æç€ºããã»ããçè§£ãããããããšããããŸãïŒèªäœã¯ãã®ãããªå·¥å€«ããããããšããç¹ã§åªããŠããã§ãããïŒïŒåãæã®èªç¥ãžã®å¯Ÿå¿ææ®µãšããŠã¯ïŒTableauã®ãããªBIããŒã«ã®æšæºããã·ã¥ããŒãã«è€éãªè¡šçŸãæŒãèŸŒã¿æ £ãããïŒåãæã®èªç¥ãåäžãããæ¹æ³ãè¯ããšã¯æããŸãïŒãã ãïŒããã¡ãã£ãã§ããïŒïŒïŒå人çãªå°è±¡ã§ããïŒãçžæã«ç¡é§ãªæèããããªãããªã©ãšïŒåãæã«ãããåçŽãªè¡šçŸã®ã¿éžæããŠããã°ïŒããããªãããšãããããªããã®ãå¢ããŠãããŸãïŒãããããããšã¯çæçãªå©ã ãé·æçãªæ¯ã ïŒãšæã£ãŠããŸãïŒããè¡ããããè€éãªè¡šçŸãã宿ã«åãæã®èªèãèæ ®ãåçŽã«ãããããªéããããã¹ãããïŒå Žåã«ãã£ãŠã¯ãã¡ãã§è¯ããšãããããŸãïŒå¥ã®ããŒãžã§ã¯ãã¡ããäž»çŒã«ãããŠæ€èšããŠããŸãïŒããã®ã§ã¯ãªãïŒãè€éããç¶æãããŸãŸè¡šçŸãå€ãåãæã®èªç¥ãåäžãããããã«èª¬æããåªåãããã¹ãã§ãïŒåçŽåã§ããªããã®ãŸã§ãã¹ãããåçŽåããæèã¯ïŒæšä»ã®ã€ãããŒã·ã§ã³åæ»ã®å€§ããªåå ã®äžã€ã§ããå€§äŒæ¥ç ã®ç åã®äžã€ããšãæããŠããŸãïŒæ³åŸããšãã¹ããŒãã·ã¹ãã ã«ããããšããŠå€±æããŸããããïŒãã¹ãããåçŽåã®åé¡ã¯ããã倱æã«ããåå瀺åãããŠããŸãïŒïŒ
SAO2Vec: Development of an algorithm for embedding the subject-action-object (SAO) structure using Doc2Vec
https://pubmed.ncbi.nlm.nih.gov/32023289/
è¿ããããããªãïŒé³åã®åãæ¹ãã§ã¯ãªãç¹ã¯ç°ãªã
ïŒAIåµè¬ã¯Alphafold2å çã®ç¬å£å Žã«ãªãã€ã€ããæ°ãããŸãïŒæ¢ç¥ã®å®æž¬ããŒã¿çŸ€ããé¡äŒŒãäºæ³ããããïŒæ§é èªäœã粟床é«ãäºæž¬ããŠé¡äŒŒïŒçµåã確èªããã»ããïŒããŒã¿äžè¶³ã®æªç¥ã«ã察å¿ã§ãããè¯ãã§ããããïŒïŒçŸå®ã®åæ ãèæ ®ããæ§é ã§ã¯ãªãïŒç¹å€ç°ã®äºæž¬ã«åŒ·ããªãïŒïŒãªã©èª²é¡ã¯ããããã§ããïŒïŒïŒ ã¿ã³ãã¯è³ªç«äœæ§é æ å ±ãçšããè¬å€ããŒãã£ã«ã¹ã¯ãªãŒãã³ã°(2021) https://www.jstage.jst.go.jp/article/jsbibr/2/1/2_jsbibr.2021.9/_html/-char/ja ïŒïŒå人çã«ã¯ïŒç¹å€ç°ãšçŸå®ã®æ§é æ å ±ãèæ ®ãããªã¬ã³ãçãšã®çµåã®é¡äŒŒãçŽæ¥æ¯èŒã§ãããã¯ãã«äœããªãããªãšèå³ããããŸãïŒèšç®éå°ãªãïŒïŒïŒ AlphaFold2ããåŸãããããå šã¿ã³ãã¯è³ªç«äœæ§é ãšã®çµåèŠªåæ§ã«åºã¥ãå»è¬åååç©ã®å¯äœçšäºæž¬(2022) https://confit.atlas.jp/guide/event-img/pharm142/26J-am-11/public/pdf?type=in æ¥æ¬è¬åŠäŒåå ããŠããŸãããã©ããªè©±ã ã£ããã§ãããããïŒïŒ
ïŒãã€ãºçµ±èšåŠå
¥é ãé »åºŠäž»çŸ©ãããã€ãºäž»çŸ©ãžã
https://speakerdeck.com/ueniki/beizutong-ji-xue-ru-men-pin-du-zhu-yi-karabeizuzhu-yi-he
ïŒç¹èš±ã«èªåçæåæã¡ã¢
æåã¯ããŒã¿äžè¶³ã§èªç¶ãªæç« ãšã¯ãªããªãããšãå€ãã§ããããããã§ããšãŠãé¢çœãæç« ãçæã§ããã§ãããïŒïŒããŒã¿äžè¶³ã¯ç¹èš±åé¡ã®èª¬ææãPPPTã³ã³ãäžäœè§£æ³ãåèã«ããã¡ããããã°è¯ãããã§ãïŒã³ã³ãã®ç¥èã¯æå€ãšåœ¹ã«ç«ã€ãã§ãããïŒïŒãŸã æ¬ããŒã«ã«å
èµããŠããŸãããå
èµãããäºå®ã§ãïŒïŒãã¯ãã«ç©ºéã¯èªåœã§è¡šçŸãããããªãèªåœãäžè¶³ããå Žåã«ã¯æ°ããç¹èš±ã衚çŸãé£ãããæ§é çéçãããïŒãšæã£ãŠããã®ã§ããïŒæ£èŠè¡šçŸã埮åå¯èœã§çµã¿èŸŒããããã§ãïŒå¯èœæ§ãåºãããŸããïŒïŒ
ïŒåºæ¬çãªèãæ¹ã¯æ¬¡ã®éãïŒãã¯ãã«ç©ºéã«ããã宿
ã¯èªåœãããªããšããïŒèªåœããªã空éã«ãäŸãã°['brain','electrod','chip']ãšããè€æ°èªåœã§è¡šçŸãããç¹ãããããã§ãïŒãã®ç¹ã¯æç« ã§è¡šãããšãã§ããŸãïŒã€ãŸãvec2seqã§ã äŸãã° https://ieeexplore.ieee.org/document/9263191
['brain','electrod','chip']ã§æŠå¿µã圢æããå Žåã¯ïŒ'brain','electrod','chip'ããå«ãæç« ãçæãããããã§ãããïŒæç« çæã®æå³ã¯ããŸããªããããããŸãããïŒãã ïŒäžèšçŸ€ããæ±ãããæŠç¥æŠå¿µãã®å Žåã¯ã©ãã§ãããïŒã空çœãã®å Žåã¯ã©ãã§ãããïŒåèªããæŠå¿µãæ±ããŠããªãã®ã§ã©ã®åèªããçæããããããããŸãããïŒåšèŸºããèªåœãéããŠãããããããªããããããŸããïŒãããªãšãã«vec2seqã¯åœ¹ã«ç«ã€ã§ãããïŒçæã®ããšãšãªãååžèªäœã¯æ£ãããšç¢ºä¿¡ãšãããæ£ãããšæ³å®ã§ããã®ã§ïŒåååµé çãšèªèããããšãã§ãïŒåãæã®ã€ã³ãµã€ããåºæ¿ãããŸãïŒ
ããªãïŒèªåã®ããæ¹ã§ã¯ïŒãã¯ãã«ç©ºéãæ¯åäœãçŽããŠããŸãã®ã§äºååŠç¿ã«ãã転移ã䜿ãé£ãïŒäœããã®å·¥å€«ããããã»ããè¯ãã§ãïŒãã¯ãã«ãããææ³ã§ãã¹ã¯ããå Žåã¯æå³ãã»ãŒå€ãããŸããã§ãããããããçšãããããããŸããïŒãŸãïŒæç« ãencoderã«éããå Žåã®åºåãç»åã¢ãã«ã®äžéå±€ã®åºåãšã¯ç°ãªãïŒencoderã«å
¥åããmemoryã¯å®å
šã«äœçœ®æ
å ±ã倱ã£ãŠããŸããïŒäœçœ®æ
å ±ã¯ã©ã®ããã«çæãããããšã«ãªãã§ããããïŒã©ããŸã§ã§ããã§ããããïŒãCNN-transformer decoderã«ããimage captioningã®äºåå®éšã§ã¯6000ä»¶ãåŠç¿ããŒã¿ãããã°è¶³ãããã§ãïŒãã ïŒèšå®ãããã©ã¡ãŒã¿ãšcpuã§ã¯åŠç¿æéã8hã»ã©ããããŸããã»ã»ã»ã¡ã¢ãªã¯4GBã»ã©ãšèš±å®¹ã§ããŸããïŒmemoryåŽã®æ
å ±éãè¶³ããªãã®ã§é©åœã«å ããŠæ€èšŒããŠãã¿ãŸãããããïŒåé¡ãåŒçšæ°ãåŒçšå
ãªã©ã»ã»ã»ïŒmemoryã€ãŸãk,vã®æ
å ±éãå°ãªãåïŒæ¯éå£ã®æç« å
šäœããqãååã«åŠç¿ã§ããããã«ããŠãããšè¯ããããããŸããïŒ
ãã空çœãéšåã¯ä»»æã®èšç®ã§æ±ããã®ã§ã¯ãªãããããžãŒã§æ±ããŠãè¯ãã§ãããã©ïŒç¹èš±ã«ããã空çœã¯æ°èŠã ãã§ãªãåã«å®çŸäžå¯èœã§ãã空çœããããŸããã»ã»ã»ããŒãïŒãšããããç·åœããã§ç©ºçœããçæãããŠã¿ãŠãé¢çœãã§ããããïŒããçšåºŠå®æ
ãèæ
®ããqãããšã«ããçæãã§ããã¯ãã§ããã»ã»ã»ç¹èš±ã®èªåçæã«ã¯ããŸãèå³ã¯ãããŸãããïŒãªããªãé¢çœããã§ã
ãimage2seqã§ãã¹ãäžã»ã»ã»
ããããŸãããã¯ããã¯ãèè² ã£ãŠããªãããšãŸã§ãããããã§ãããã»ã»ã»pretrainãããç»ååé¡ã«ããã¯ããã¯ã¯å«ãŸããŠããã§ããããâèŠåãããããã¯ãã«ã圢æãããŠãããšèšããã§ããããã»ã»ã»
n02769748: backpack, back pack, knapsack, packsack, rucksack, haversack
https://image-net.org/challenges/LSVRC/2012/browse-synsets.php
å«ãŸããŠã¯ããŸãããïŒåé¡å€ã®è¡šçŸãçšããæ¯èŒçåµé çãšèšããããªæç« ãçæããäŸãèŠã€ãããããã©ã»ã»ã»
ããïŒããããªãããšãã衚çŸã¯ãããããïŒã¯ã©ã€ãã³ã°æã«ã¯ããã¯ããã¯ãèè² ã£ãŠããããšãæ®éã§ããïŒãããšãäžè¬ã§ããå Žåã«ããèªç¶ã«èªããªãã¯ãã§ãããïŒããã¯ãããã確信æ§ãé«ããšããæå³ã§åµé çãªã®ãããããŸããïŒ
ïŒmT5ã®ãããªencoder-decoderãèããã®ã§ããïŒencodeéšåã¯æŒç®å¯èœã§ãããã¯ãã«ç©ºéã§ãªããšç©ºçœé åã®æå®ãé£ããã®ã§ããã
Technology Fitness Landscape for Design Innovation: A Deep Neural Embedding Approach Based on Patent Data
https://arxiv.org/abs/2110.13624
åŒçšã°ã©ããšåèªåã蟌ã¿ããæ°ããªåã蟌ã¿ãåŠç¿ïŒ
é¢çœãç¹åŸŽããšãããŠãã
GraphSAGE
http://snap.stanford.edu/graphsage/
æªç¥ã®ããŒãã®embeddingãèšç®ããããšãã§ãã
https://github.com/kzh22/graphsage-simple
ïŒæèŠçã«ã¯ïŒattentionãçšãã粟床ãé«ããªããããææ³ãããïŒçµ±èšçãªtf-idfãååžä»®èª¬ãçšããææ³ã®ã»ããïŒããèªç¥ã«è¿ããšæããŠããŸãïŒãŸãïŒãããçµ±èšçãªææ³ã¯ïŒæ¯éå£ã®ç¹æ§ã«å¿ãå€åãããšããéœåã®è¯ãæ§è³ªãæã£ãŠããŸãïŒ
ããŸãïŒããšç¹èš±æç« ã«ãããŠã¯ïŒattentionãçšãã粟床ãé«ããªããããææ³ããããŠããŸãããæèã«åºã¥ããåèªã®æå³ããèŠåããæ©èœã¯ïŒäžéå端ã§éªéã«ããªããŸãïŒç¹èš±ã«ãããæèã¯ïŒæçްãèªã¿èŸŒã¿ããã«èª¿æ»ããããšã«ããåããŠå€æãïŒãã®äžã§åèªã®æå³ã確å®ãããªã©ïŒã©ããŸã§ãã£ãŠãããŒã¿å€ã®ããŒã¿ãå¿
èŠãšãªããªã©è€éã§ç¹çްã§ããããšãå€ãïŒattentionææ³ãåŠç¿ãããæ¯èŒçäžè¬çãªæèãèæ
®ããåèªã®æå³ããšã¯ïŒfine-tuningãããšããŠãïŒç°ãªãããšãå€ãããã§ãïŒcommon sence reasoningãŸã§ã¯è§£ããããã©ç¹èš±è§£éã«å¿
èŠãªæè¡åžžèãŸã§ã¯åŸãããªãïŒãšããã€ã¡ãŒãžã§ãïŒïŒçµ±èšçãªææ³ã¯ãã®ãããªéªéãããŸããïŒç¹èš±ã«ãããŠã¯ïŒãæèã«åºã¥ããåèªã®æå³ãã¯ãçµ±èšçãªææ³ãèŠãåŸã«ããã倿ããïŒhuman-machine collaborationãã®æ¹éãšããŠåŸãã»ããè¯ãçµæãšãªãããšãå€ãã§ãããïŒãïŒself attentionãçšããææ³ã¯queryã§ãã®æãã©ã®ãããªèгç¹ã§èªã¿åããæ±ºå®ããŠãããšèšããŸãïŒmulti headãçšããããšã§ãè€æ°ã®èгç¹ã§æãèªã¿åã£ãŠãããããã§ãïŒããŠïŒheadæ°ã¯largeã§16åã»ã©ã§ããïŒãã®åæ°ã§ãŸãã¯ãã®ã¢ã«ãŽãªãºã ã§ïŒãå¿
èŠååãªèгç¹ããæã£ãŠãããšèšããã§ããããïŒèгç¹ã¯æäœã§ãææ³èªèã«å¿
èŠãªåæ°ïŒ5w1hã®6åïŒäººãèªç¥ããæèã«å¯Ÿå¿ããåæ°ïŒæŽã«æäžã®éæ©èœèªããšã«1ã€æ¬²ãããšããã§ãïŒãããïŒqãkãXããæ±ãã以äžXã«èгç¹ãå«ãŸããŠããå¿
èŠãããïŒãã€ãåããèæ
®ããã«ããŠãïŒãå¿
èŠååãªèгç¹ããæããã«ã¯éçãããã§ãããïŒAttention Is All You Needè«æã®è©Šéšçµæãããããšheadæ°ãåçŽã«å¢ããã°ããããã§ã¯ãªãããã§ããïŒããã¯Xã®æ
å ±éã«éçãããheadãå¢ãããããŠã解決ã§ããªããšããæå³ã§ãããå€åïŒãŸãéçŽã®ä»æ¹ã¯concatãšéã¿ä»ããšãªã£ãŠããŸããïŒããã§è¶³ããã§ããããïŒãå¿
èŠååãªèгç¹ããæããããã®æ¡ãæãã€ãã§é©åœã«åºãã°ïŒheadã®å€æ§æ§ã確ä¿ããããã«concatåã«å
ç©ãåãé¡äŒŒãã芳ç¹ã¯éã¿ä»ããã€ã€ç¬ç«ã§ã¯ãªããšããŠé€å»ããŠããïŒãŸãã¯qã®å€æ§æ§ã確ä¿ãããããŸãã¯äž»é¡ã«æ²¿ã£ãåŠç¿ããããããããŒã¿å€ã®æé»ç¥ãšããŠä»»æã®subjectãäžããŠããããã«è¿ã芳ç¹ãåªå
çã«æ®ãïŒãªã©ãããããããããŸããïŒåŸè¿°ã®ãããã¯ãŒã¯ã¯ãããèæ
®ããŠåœ¢æã§ããããã«ããŠããŸãïŒïŒåŸè¿°ã®ç¹èš±çæã¯attentionææ³ãçšããçæã§ããïŒãããèæ
®ããŠçæã§ããããã«ããŠããŸãïŒ
ãè€éãªèгç¹ãå¿
èŠãšãªãå Žåã«ã¯ïŒä»ã®ãšããã¯ãŸã ïŒhuman-machine collaborationãçãã ãšæããŸãïŒããæ°çŸGBã®GPUã¡ã¢ãªãçšããææ°ã¢ãã«ã§ã¯ã©ããªã£ãŠããã®ãããããŸããããã©ãïŒçå±ã®äžã§ã¯ç¡éã®Xã«ã¯ãã¹ãŠã®èгç¹ãå«ãŸããŸãããïŒ
ããŸãïŒç¹èš±æç« ã®å Žåã¯ïŒsequenceãããŸãéèŠãšãªãã«ãããããŸãã¯ésequenceæ
å ±ãããéèŠã§ããããšãå€ãããïŒsequenceãèæ
®ããææ³ã¯éå®çãªèª²é¡ä»¥å€ã§ã¯æ¯èŒçéèŠã§ã¯ãããŸããïŒæ©æ¢°ç¿»èš³ãåºæè¡šçŸæœåºãææ
åæãšã¯ç°ãªãïŒå
¥åé ãç¡èŠããŠãæ¯èŒçååãªæ
å ±ãåŸãããïŒãšè¡šçŸããã°ããã§ããããïŒãšãããŸããã©ã¡ãããšããã°é çªã¯ç¡èŠããŠã»ãããããã§ã¯ãããŸãïŒäŸãã°ïŒããå«ããšããå«ãŸãªããšããé åºãèæ
®ããã°èŠåãããã衚çŸã¯èŠåããåã䟡å€ã§è¡šçŸããŠæ¬²ããã§ãå人çã«ã¯ïŒ
ãããç¹æ§ãéèŠãïŒäœæãããã¯ãã«ã¯çµ±èšçææ³ã§åŸããããã®ãã骚ããšããŠããŸãïŒ
ã»åºæ¬çã«ã¯ã¯ã©ã¹ã¿ãªã³ã°ãè¡ãã³ãŒãã§ããïŒç¹å®ã®æ¡ä»¶ã«åãç¹èš±çŸ€ã®è²ãä»»æã«å€ãããªã©å¯Ÿå¿ããããšã«ããã»ã°ã¡ã³ããŒã·ã§ã³ã«ã䜿ããã§ãããïŒåŸã»ã©è¿°ã¹ãããã«ãåŒã³æ°Žããå©çšããã°ïŒè»žã«ä»»æã®æå³ãããããããšãã§ããŸãïŒ
ïŒæ¢ç¥ã®ç¥èãæŽ»çšããã«ã¯ã»ã°ã¡ã³ããŒã·ã§ã³ïŒæªç¥ã®çºèŠãæ±ãããªãã¯ã©ã¹ã¿ãªã³ã°ãããé©åã§ãïŒã¯ã©ã¹ã¿ãªã³ã°ã¯æªç¥ã®æ§é ã®çºèŠãå¯èœãšããŸããïŒããçšåºŠã®ç¥èã®åæ§æãå¯èœãšããŸãïŒæ¬ã³ãŒãã¯ãã®ç¹ãéèŠããŠããŸãïŒ
ïŒäºæ³ä»¥äžã«äœ¿ãåæãè¯ãã£ãããïŒããšããšã¯å¥éäœæããŠãããç¹èš±SDIçšAIã·ã¹ãã ãã®äžéšã§ãããç¬ç«ãããŸããïŒ
https://qiita.com/kzuzuo/items/d41327433c9cdc6a5fd3
https://qiita.com/kzuzuo/items/4670b5ff7526319680f4
ïŒAãæ©æ¢°åŠç¿ã¯èªååå«ããå€ãã®æ±ºå®ãããéã«æ¡çšãïŒ
Bãçµ±èšã¯ããããã®æææ±ºå®ãããéã«æ¡çšãïŒ
Cãåæã¯æªç¥ã«å¯Ÿããã€ã³ã¹ãã¬ãŒã·ã§ã³ãåŸããå Žåã«æ¡çšãã
ãšãããŸãïŒ
ãç¹èš±SDIçšAIã·ã¹ãã ãã¯Aãæ
ãïŒæ¬ããŒã«ã¯äžéšã®BãšCãæ
ã£ãŠããŸãïŒ
ïŒæ±äº¬å€§åŠ ç·åç ç©¶æ©æ§ ã€ãããŒã·ã§ã³æ¿çç ç©¶ã»ã³ã¿ãŒãæ±äº¬å·¥æ¥å€§åŠå€§åŠé¢ ã€ãããŒã·ã§ã³ãããžã¡ã³ãç ç©¶ç§ æ¢¶å·ç 究宀
åŠè¡ä¿¯ç°ã·ã¹ãã
https://academic-landscape.com/page/about
202302远èšãchatGPTã«èããŠã¿ã
ãchatGPTã¯ïŒã³ã³ã»ã³ãµã¹ã®ããåçã«ã€ããŠã¯ã¯çã£åœã§ïŒã³ã³ã»ã³ãµã¹ã®ãªããšããžã®å¹ããåçã«ã€ããŠã¯èŠæãšããŠããå¿è±¡ã§ãïŒæ¬¡ã®è³ªåã¯ã³ã³ã»ã³ãµã¹ã®ãªãåçãšãªã質åãšæããŸããïŒè峿·±ãåçã§ã¯ãããããããŸããïŒ
- ã¢ãã«èªäœã®åµé æ§ãšã¯äœã§ãã
- ã¢ãã«ã®ç€ºãåµé æ§ãšã¯ãæ°ããã¢ã€ãã¢ãæŠå¿µãææ¡ããæ¢åã®ãã®ãšã¯ç°ãªãæ¹æ³ã§åé¡ã解決ããããšãã§ããããšãæããŸããã¢ãã«ã®åµé æ§ãè©äŸ¡ããããã«ã¯ã以äžã®ãããªææšãèããããŸãã
æ°èŠæ§ïŒã¢ãã«ãææ¡ããè§£æ±ºæ¹æ³ãæŠå¿µããæ¢åã®ãã®ãšæ¯ã¹ãŠæ°ãããã®ã§ãããã©ãããããã¯ãäŸãã°ãå°éå®¶ãäžè¬ã®äººã ãããã®ææ¡ã«ã€ããŠã©ã®çšåºŠç¥ã£ãŠãããã調æ»ãããããæ°èŠãã©ãããè©äŸ¡ããããšã§æž¬å®ããããšãã§ããŸãã
æçšæ§ïŒã¢ãã«ãææ¡ããè§£æ±ºæ¹æ³ãæŠå¿µããåé¡ã解決ããããã«å®çšçã§ãããã©ãããããã¯ãäŸãã°ãå®éã«ãã®ææ¡ãå®è£ ãããã®å¹æãè©äŸ¡ããããšã§æž¬å®ããããšãã§ããŸãã
æå€æ§ïŒã¢ãã«ãææ¡ããè§£æ±ºæ¹æ³ãæŠå¿µãããã®åé¡ã解決ããããã®äºæ³å€ã®æ¹æ³ã§ãããã©ãããããã¯ãæ¢åã®æ¹æ³ãšæ¯ã¹ãŠç°ãªãè§£æ±ºæ¹æ³ãææ¡ããããã©ãããè©äŸ¡ããããšã§æž¬å®ããããšãã§ããŸãã
äžè¬æ§ïŒã¢ãã«ãææ¡ããè§£æ±ºæ¹æ³ãæŠå¿µãããã®åé¡ã解決ããã ãã§ãªããä»ã®åé¡ã«ãé©çšå¯èœã§ãããã©ãããããã¯ãäŸãã°ãææ¡ãããè§£æ±ºæ¹æ³ãæŠå¿µããä»ã®åé¡ã«ãé©çšå¯èœã§ãããã©ãããè©äŸ¡ããããšã§æž¬å®ããããšãã§ããŸãã
以äžã®ãããªææšãç·åçã«è©äŸ¡ããã¢ãã«ã®åµé æ§ãè©äŸ¡ããããšãã§ããŸãã
ïŒïŒsimCSEã®ãããªïŒè·é¢åŠç¿ã®ã¿ãè¡ã£ãå Žåã«ã¯ïŒããã®ãããªå¶çŽããªãå ŽåïŒé ãã®å¶çŽãã§ããŠãªãéãïŒããŒã¿ã®å¶ç¶ã«å·Šå³ããããšæããïŒïŒentangleãããè€æ°ã®è»žããã¯ãã«ã嫿ããããšãã§ãããïŒçåã«æã£ãŠããŸãïŒãŸãïŒè·é¢åŠç¿ã¯å°æ°ã®ããŒã¿ãšã¯çžæ§ãæªããšããããïŒç¹å®æ¯éå£ã®ç¹åŸŽãèæ
®ããåŠç¿ã¯é£ããã§ãããïŒãŸãèªåãç¥ãéãã§ã¯ïŒäœ¿çšäŸïŒ.ïŒã®ãããªåèªãã¯ãã«ãšã®æŒç®ã¯é£ããã§ãïŒäžèšã«äžããããã€ããããã§ããïŒïŒè·é¢åŠç¿ãšentangleãŸãã¯æŒç®å¯èœæ§ã«ã€ããŠã¯èª¿æ»äžã§ãïŒçŸç¶ã§ã¯ïŒè·é¢åŠç¿ã¯è£æ£ã«ã®ã¿äœ¿ããŸãïŒãªãïŒè£æ£ããå Žåã«ã¯ïŒè£æ£ã®äžéãèšå®ãããšè¯ãããã§ãïŒïŒ
ç¡å¹è³æèª¿æ»ã®ãã¡äž»åŒçšäŸã®ã¿ã®æ€çŽ¢ãç¹èš±ã®ã¿ããè¡ãïŒæç« ã¬ãã«ã®ã¿æ±ãïŒåŸã«è¿°ã¹ããããªåèªã¬ãã«ã®æŒç®ãããŠå¯åŒçšäŸãæ¢ããããªããšãããªããªãïŒè·é¢åŠç¿ã¯è¯ãããã§ãïŒ
ããã»ã»ã»å¯Ÿè±¡åŠç¿ã§ãããCLIPã§æŒç®ã§ããã®ã ããæŒç®èªäœã¯ã§ããã®ãïŒ
察象åŠç¿ã¯ããããå°ãããšæèåŠç¿ãã§ããªãåé¡ã¯ãã
åèªãå ããæŒç®ãããéã«ã¯ã»ã»ã»ãã®åèªãè€æ°ã®æŠå¿µãæã£ãŠããå Žåã¯ã»ã»ã»
æç« åäœã§ãã¯ãã«ãçæãïŒåèªã¬ãã«ã®æŒç®ãããå ŽåãšïŒåèªã¬ãã«ã§ãã¯ãã«ãçæãïŒåèªã¬ãã«ã®æŒç®ãããå ŽåïŒã©ã¡ããè¯ãã®ã ãããïŒåè
ã¯transformer以åã¯åŸè
ããæªããšèšããã®ã ãã»ã»ã»ãåèªã®æèèæ
®ã®å¿
èŠæ§ã§èããã°ããã®ãïŒ
chatgptãããããïŒ
ã察ç
§åŠç¿ãšword2vecã¯ãäž¡æ¹ãšãåèªã®åæ£è¡šçŸãåŠç¿ããããã®æ¹æ³ã§ããããã®ç¹æ§ã«ã¯ããã€ãã®éãããããŸãã
ãŸãã察ç
§åŠç¿ã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ãçšããŠåŠç¿ãè¡ããããããŒã¿ã»ããã«å¿ããŠåŠç¿ããããã¯ãã«ãç°ãªãå¯èœæ§ããããŸããããã«å¯ŸããŠãword2vecã¯ãã¥ãŒã©ã«ãããã¯ãŒã¯ãçšããŠåŠç¿ãããããåŠç¿ããã»ã¹ãããäžè²«æ§ãããããšãç¥ãããŠããŸãã
ãŸãã察ç
§åŠç¿ã¯ãèšå€§ãªéã®ããã¹ãããŒã¿ãå¿
èŠãšããŸãããword2vecã¯æ¯èŒçå°ãªãããŒã¿ã§å¹æçãªåŠç¿ãã§ããŸãã
ããã«ã察ç
§åŠç¿ã¯ãåèªã®å
±èµ·æ
å ±ãèæ
®ããªããåŠç¿ãè¡ããããæèã«ãã£ãŠç°ãªãæå³ãæã€åèªã®ãã¯ãã«è¡šçŸãç²åŸã§ããŸããäžæ¹ãword2vecã¯ãåèªã®åšèŸºã®åèªã®åºçŸç¢ºçãèæ
®ããŠåŠç¿ãè¡ããããåèªã®æå³ããã粟確ã«è¡šçŸã§ããããšãç¥ãããŠããŸãã
以äžã®ãããªç¹æ§ã®éãããããããã©ã¡ãã®ææ³ãé©ããŠãããã¯ã¿ã¹ã¯ã«äŸåããŸããã
ãç¹èš±æç« 矀ã«ã€ããŠïŒsimCSEãªã©å¯Ÿç
§åŠç¿ãçšããŠäœæãããã¯ãã«ç©ºéãšïŒword2vecãçšããŠäœæãããã¯ãã«ç©ºéã®ïŒç¹æ§ã®éãã¯ãªãã§ããïŒ
simCSEãword2vecãªã©ã®èªç¶èšèªåŠçã«ãããã¢ã«ãŽãªãºã ãçšããŠäœæããããã¯ãã«ç©ºéã«ããç¹èš±æç« 矀ã®åæã«ã€ããŠèãããšã以äžã®ãããªç¹æ§ã®éããæããããŸãã
åŠç¿ã®æ¹æ³ã®éã
simCSEã¯ã察ç
§åŠç¿ãçšããŠãã¯ãã«ç©ºéãåŠç¿ããŸããäžæ¹ãword2vecã¯ãã¥ãŒã©ã«ãããã¯ãŒã¯ã«ããæåž«ãªãåŠç¿ãçšããŠãåèªã®åæ£è¡šçŸãåŠç¿ããŸãããããã£ãŠãsimCSEã¯å¯Ÿç
§åŠç¿ã«ãã£ãŠãåèªãæç« ã®æå³çé¢ä¿ãåŠç¿ããããšãã§ããŸãããword2vecã¯ãåèªã®åºçŸé »åºŠã«åºã¥ããçµ±èšçãªæ
å ±ããåèªã®æå³çé¢ä¿ãåŠç¿ããããšãã§ããŸãã
衚çŸã®ç²ŸåºŠã®éã
simCSEã¯ãåèªãæç« ã®æå³çé¢ä¿ãããæ£ç¢ºã«è¡šçŸããããšãã§ããå¯èœæ§ããããŸããããã¯ãsimCSEã察ç
§åŠç¿ã«ãã£ãŠãå矩èªãé¡çŸ©èªãªã©ã®åèªã®æå³çãªé¡äŒŒæ§ãåŠç¿ããããšãã§ããããã§ããäžæ¹ãword2vecã¯ãåèªã®å
±èµ·é¢ä¿ã«åºã¥ããŠåèªã®æå³çãªé¡äŒŒæ§ãæšå®ããããã衚çŸã®ç²ŸåºŠãäœäžããå¯èœæ§ããããŸãã
èšç®ã³ã¹ãã®éã
simCSEã¯ã察ç
§åŠç¿ãçšããŠãåèªãæç« ã®ãã¯ãã«è¡šçŸãåŠç¿ãããããèšç®ã³ã¹ããé«ãåŸåããããŸããäžæ¹ãword2vecã¯ãåèªã®å
±èµ·é¢ä¿ãåºã«ãã¯ãã«è¡šçŸãåŠç¿ãããããèšç®ã³ã¹ããæ¯èŒçäœãã§ãããããã£ãŠãå€§èŠæš¡ãªç¹èš±æç« 矀ã®åæã«ãããŠã¯ãword2vecã®æ¹ãå¹ççã§ããå¯èœæ§ããããŸãã
以äžã®ããã«ãsimCSEãword2vecãªã©ã®ãã¯ãã«è¡šçŸåŠç¿ã¢ã«ãŽãªãºã ã«ãã£ãŠçæããããã¯ãã«ç©ºéã®ç¹æ§ã«ã¯éãããããŸããã©ã¡ããåªããŠãããã¯ãåæã®ç®çãããŒã¿ã®æ§è³ªã«ãã£ãŠç°ãªããŸãã
ããããã®åé¡ç¹ã«ã€ããŠæããŠãã ãã
simCSEãword2vecãªã©ã®èªç¶èšèªåŠçã«ãããã¢ã«ãŽãªãºã ã¯ãé«ã衚çŸåãæã€ãããå€ãã®å¿çšãå¯èœã§ããããããã«ã¯ããã€ãã®åé¡ç¹ããããŸãã以äžã«ãããããã®åé¡ç¹ã説æããŸãã
é ã倿°ã®è§£éã®å°é£æ§
simCSEãword2vecãªã©ã®ã¢ã«ãŽãªãºã ã¯ãåèªãæç« ã®æå³çãªç¹åŸŽã衚çŸããããããã¯ãã«ç©ºéãçæããŸãããããããã®ãã¯ãã«ç©ºéã®å次å
ãäœã衚ããŠããã®ããè§£éããããšã¯å°é£ã§ãããããã£ãŠãè§£éæ§ãäœããããå®éã«ã©ã®ãããªç¹åŸŽã衚çŸããŠããããæ£ç¢ºã«çè§£ããããšãé£ãããªããŸãã
Out of Vocabulary (OOV) åé¡
simCSEãword2vecãªã©ã®ã¢ã«ãŽãªãºã ã¯ãäºåã«åŠç¿ããåèªãæç« ã®ã¿ã«å¯ŸããŠããã¯ãã«è¡šçŸãçæãããããæªç¥èªã«å¯ŸããŠã¯å¯Ÿå¿ã§ããŸããããã®åé¡ããOut of Vocabulary (OOV) åé¡ããšåŒã³ãŸããç¹ã«ãå°éçšèªãªã©ã®ã¬ã¢ãªåèªã«å¯ŸããŠã¯ãOOVåé¡ãæ·±å»ã«ãªããŸãã
ããŒã¿ã®ã¹ããŒã¹æ§ã®åé¡
åèªã®å
±èµ·æ
å ±ã«åºã¥ããŠåŠç¿ãè¡ãword2vecãªã©ã®ã¢ã«ãŽãªãºã ã§ã¯ãããŒã¿ã®ã¹ããŒã¹æ§ãåé¡ãšãªããŸããã€ãŸããããåèªãšä»ã®åèªã®å
±èµ·é »åºŠãäœãå Žåããã®åèªãšä»ã®åèªãšã®æå³çãªé¢é£æ§ãæ£ç¢ºã«æããããšãã§ããŸããããã®ãããåèªã®æå³çãªè¡šçŸãäžååã«ãªãããšããããŸãã
ãã¬ãŒãã³ã°ããŒã¿ã®åãã®åé¡
simCSEãword2vecãªã©ã®ã¢ã«ãŽãªãºã ã¯ããã¬ãŒãã³ã°ããŒã¿ã«äŸåããããããã¬ãŒãã³ã°ããŒã¿ãåã£ãŠããå Žåãçæããããã¯ãã«è¡šçŸã«ãåããçããŸãããã®ããããã¬ãŒãã³ã°ããŒã¿ã®è³ªãåäžãããããšãéèŠã§ãã
以äžããsimCSEãword2vecãªã©ã®èªç¶èšèªåŠçã«ãããã¢ã«ãŽãªãºã ã«é¢ããåé¡ç¹ã®äžéšã§ãããããã®åé¡ã解決ããããã«ãããé«åºŠãªææ³ãæ¹ã
ãããã£ãŠãããã©ããã¯å¥ãšããŠïŒèª²é¡ãæŽçã§ããŠããã§ããã
ã³ã¡ã³ã
衚çŸã«ã€ããŠïŒ
åèªã®æèãèæ
®ããŠã»ããã¯ãããŸããïŒåäžåèªãè€æ°ã®æå³ã«åããŠã»ãããããŸããïŒéã«åèªããã®é¡çŸ©èªãå«ããŠåããããªãã¯ãã«ã§è¡šçŸããŠã»ããã§ãïŒæ¬ä»¶çšéã§ã¯ïŒword2vecã®ã»ããé©åã§ãããïŒãã ãïŒå¯Ÿç
§åŠç¿ã§ãïŒåäžåèªãšé¡çŸ©èªãåãã¯ã©ã¹ã¿ã«å
¥ããäžã§ïŒåäžåèªãè€æ°ã®æå³ã«åããŠããããšã¯æãããŸãïŒ
åèªãå«ããæŒç®ãããéã«ã¯ïŒå¯Ÿç
§åŠç¿ã§ã¯åäžåèªã§ãå¥ã®æå³ãšè§£éãããããéœåãè¯ãã¯ãªãã§ãããïŒ
æ¯éå£ã«åºã¥ããæèè§£éã¯ïŒäœ¿çšè
ã®èªç¥ã«å«ãŸããŸããïŒç¹ã«ç¹èš±ãªã©æªç¥ã®æç®ãæ¯éå£ãšããå Žåã¯é¡èã§ãããïŒãã®å ŽåïŒäœ¿çšè
ã®èªç¥ã«è¿ãã®ã¯ïŒword2vecã®æ¹ã«ãªãã®ã§ã¯ãªãã§ããããïŒ
OOVã«ã€ããŠ
word2vecã§ã¯åé¡ãšãªããŸããïŒã€ãåŠç¿ããŠããïŒèªåœã¯èœãšããŠããªãããã§ãïŒ
äžæ¹simCSEã§ã¯ïŒäžåºŠå€§ããåŠç¿ããã¢ãã«ãæµçšããããïŒOOVã®åé¡ããããŸãïŒç¹ã«ç¹èš±ãªã©æ°åèªãçŸãããæç« ã§ã¯åé¡ãšãªãããã§ãïŒ
å
±èµ·é »åºŠã«ã€ããŠ
word2vecã§ã¯åé¡ãšãªãããã§ãïŒå¯Ÿç
§åŠç¿ã§ã¯è² äŸã«ãã驿£åã§ããããããããŸããïŒ
èšç·ŽããŒã¿ã®åãã«ã€ããŠ
æ¬ç®çã«ãããŠã¯ïŒæ¯éå£ã®åããéèŠãšãªã£ãŠããŸãïŒsimCSEã§ã¯ãã倧ããªèšç·ŽããŒã¿ã§èšç·Žããããšã«ãªãïŒåãã¯ãªããªã£ãŠããŸããŸãïŒããã¯æ¬ç®çã«ãããŠã¯éœåãæªãã§ãïŒsimCSEã¯ïŒå€§ããªä»£è¡šååžãåŸãããŠããå Žåã«ã¯é©åãšæãããŸããïŒç¹èš±ãªã©ååžãå°æ¥å€§ããå€ããããå Žåã«ã¯ïŒããŸãé©åãªææ³ãšã¯ãããªãã®ã§ã¯ãªãã§ããããïŒãåŸãåãã®æ€èšŒã¯ããè¯ããªãåŸããšã¯æããŸããã»ã»ã»
ãŸãïŒèªåã¯ïŒattentionãããïŒæ±çšã§ããååžä»®èª¬ã远æ±ããããŒã«ãäžã€æã£ãŠãããããšæã£ãŠãããŸãïŒ
ãååžæå³è«ä»®èª¬ã¯äºæ³ãè¶
ããŠ(ãã£ã±ã)匷åã ã£ããããšã¯å€ã
ç¥ãããŠãããšããã§ãïŒ
word2vecã¯ååžä»®èª¬ãã®ãã®ã§ãããïŒå¯Ÿç
§åŠç¿ãïŒååžä»®èª¬ãåŠãã§ããã§ãããïŒæ£äŸ2ã€ïŒã€ãŸãé¡äŒŒæç« ã«ã¯ïŒåãæ£ã®åèªãçŸããããšãå€ãã§ããããïŒãããããšåãåèªã®åšèŸºã«åºçŸããŠããåèªã®ååžãåŠãã§ããã¯ãã§ãïŒword2vecãšã®å€§ããªéãã¯ïŒå¯Ÿç
§åŠç¿ãè² ã®åèªãåŠã¶ã®ã«å¯ŸãïŒword2vecã¯è² ã®åèªãåŠã°ãªããšããç¹ã«ããã§ãããïŒè² ã®åèªãšã¯ïŒäžçŸ©ã«æ±ºãŸãã®ã§ããããïŒãã©ããŸã§ãã£ãŠãïŒãã¡ã€ã³é©åãå¿
èŠã«æããŸããïŒå¯Ÿç
§åŠç¿ã«ãããŠã¯è² äŸã®èšå®ã倧ããªããŒãã§ãããïŒããã«æ£ææ§ãããïŒå¯Ÿç
§åŠç¿ã§åŸããããã¯ãã«ã«ã¯ïŒå®¢èгçã§ãªãæ£ææ§ãå
¥ãããã«æããŠããŸãïŒãããæ°æã¡æªãã§ãïŒååå€ãã®ããŒã¿ãæããšãã«ã¯ïŒè² äŸã®æ±åãã§ãïŒç¡èŠã§ãããããããŸãããã»ã»ã»
ãšã¯ããïŒæ€èšŒããªããšãããªãã§ãããïŒ
察ç
§åŠç¿ã«ããæãã¯ãã«ã䜿çšããé害ã¬ããŒãã®ã¯ã©ã¹ã¿ãªã³ã°
https://www.anlp.jp/proceedings/annual_meeting/2022/pdf_dir/B6-1.pdf
GPT-4ã®èŠè§£ãSimCSEãšSentence BERT:
ãããã®ææ³ã¯æã¬ãã«ã®åã蟌ã¿ãäž»ãªç®çã§ãããåèªã¬ãã«ã§ã®æŒç®ã«é¢ããŠã¯ä¿èšŒãããŸããããã ãã䜿çšãããäºååŠç¿æžã¿ã®èšèªã¢ãã«ïŒäŸïŒBERTïŒã«ãã£ãŠçæãããåèªãã¯ãã«ã¯ãæèã«äŸåããæå³è¡šçŸãå¯èœã§ãããããåèªéã®é¡äŒŒæ§ãããçšåºŠèšç®ããããšãã§ããŸãã
Word2Vec:
Word2Vecã¯åèªã¬ãã«ã®ãã¯ãã«è¡šçŸãåŠç¿ããããšãç®çãšããŠãããåèªãã¯ãã«åå£«ã®æŒç®ãäžå®ã®æå³ãæã€ããšãç¥ãããŠããŸããå ç®ãæžç®ã«ãã£ãŠãæå³çãªé¢ä¿æ§ãä¿æããããšãã§ããŸããããããã®æŒç®ãåžžã«æ£ç¢ºãªçµæãããããããã§ã¯ãããŸããã
ç·ããŠãSimCSEãSentence BERTã®åèªãã¯ãã«ã¯ãæèã«äŸåããæå³è¡šçŸãå¯èœã§ãããããé¡äŒŒæ§èšç®ã«ã¯æçšã§ããããã¯ãã«åå£«ã®æŒç®ã«ããæ°ããªæå³ã®åœ¢æã«é¢ããŠã¯ä¿èšŒãããŸãããäžæ¹ãWord2Vecã¯åèªã¬ãã«ã§ã®æŒç®ãæå³çãªé¢ä¿æ§ãä¿æããããšãã§ããããšãå€ããããåèªéã®æå³æäœã«é©ããŠããŸãããã ããçµæã®æ£ç¢ºæ§ã¯ã±ãŒã¹ãã€ã±ãŒã¹ã§ç°ãªããŸãã
ïŒèšèªã¯ã¹ããŒã¹ã ã£ãã®ã§ãããã?ãèªåã¯åèªã«æå³ãä»äžãããåæã§ãã¹ããŒã¹ã ãšæããŸãã
ã»Isotropy, Clusters, and Classifiers
https://speakerdeck.com/hpprc/isotropy-clusters-and-classifiers
ãåã蟌ã¿è¡šçŸã®çæ¹æ§åã¯ãSTSãªã©æå³çé¡äŒŒåºŠã¿ã¹ã¯ã§ã¯æçã ããã¯ã©ã¹ã¿ãªã³ã°ã¿ã¹ã¯ã«ãããŠã¯äžå©çã察ç
§åŠç¿ãè¯ãæªãã
äœæã©ããã®çµè«ãšæããŸãã
ïŒãç¥ããªãããšãèªãïŒé²ãã¹ãæ¹åããããããªããšããæ
床ãïŒãã€ãä¿ã¡ã€ã¥ããããšïŒãããããã°ïŒããšãäœãã»ããã®ããç¥ããªããšãïŒçµå±ã¯ãããéæããæ¹æ³ãè²ãŠãã®ã«å¿
èŠãªæèã®å€é©ïŒç¥èãžã®æ°ããè²¢ç®ãšçºèŠãå¯èœã«ãªãã®ã§ãïŒïŒãã¡ã€ã³ãã³ïŒã
èªåãªãã«èšãçŽããªãïŒãã¢ãã«ãæ£ãããšèªèãïŒãã®ã¢ãã«ã瀺ããçµæã«ã€ãèªåã¯äœã«æ°ã¥ããŠããªãã®ãåãç¶ããïŒãã®å§¿å¢ãåãããšãã§ãããã®ã®ã¿ãïŒã€ã³ãµã€ããåŸãããšãã§ãããã®ã ãšæããŸãïŒãããŠïŒã¢ãã«ãæ£ãããšã®ç¢ºä¿¡ãæã€ã«ã¯ïŒãã®ã¢ãã«ãçè§£ããŠããããšãéèŠã§ãïŒ
ãåã¯ãã®ãšãã¯ã£ãšããïŒãªãããªã³ã¹ãã³ã®å®éšå®€ããïŒã©ãã©ãå ±åãåºãŠããã®ãã«æãåœãã£ãããã ïŒåœŒãã¯å®éã«èªåãã¡ã®æã§é ããããè£
眮ã§ç ç©¶ããŠããã®ã ïŒã ããããã©ãã«äœãããïŒäœãã©ãåããŠããããïŒã¡ãããšããã£ãŠããã®ã ïŒïŒãã¡ã€ã³ãã³ïŒã
çµ±èšã¯ããçè§£ããããã§ãïŒn-gramãçšãã確çèšèªã¢ãã«ã¯ãŸã çè§£ã§ããŸã
https://qiita.com/kzuzuo/items/b6875441d7103ee515c1 ïŒããããæ·±å±€åŠç¿ã«ããã¢ãã«ã¯ããçè§£ããããã§ãïŒçµ±èšãçè§£ã§ããªã人ãããã°æ·±å±€åŠç¿ã¢ãã«ãçè§£ã§ãããããããŸãïŒåãæã®çè§£åã®åé¡ã§ãïŒ
ãã ïŒã¢ãã«ãçè§£ãããšãïŒèªèããããšã¯ã§ããŸãïŒçè§£ããã¹ãã§ããïŒèªèããããšã§ãïŒã€ã³ãµã€ããããããšãã§ãããããšãªãã§ãããïŒ
ïŒèª¬æãæ±ããã¹ããªã®ã¯ïŒèªåã®èªèãšããèªå·±ã®å
éšã«å¯ŸããŠãªã®ã§ãã£ãŠïŒã¢ãã«ã«å¯ŸããŠã§ã¯ãªãïŒãšèšãã°ããã§ããããïŒ
ïŒã¢ãã«ãåŠä¿¡ãããšããæå³ã§ã¯ãªãïŒã¢ãã«ãéå°ã«åŠå®ããªããŸãã¯ïŒãŸãã¢ãã«ã®åºåã«æå³ããããšä»®å®ãèããïŒãšããããšïŒ
ïŒéäžåŒãçè§£ã§ããªããŠãïŒçµè«ã¯åãïŒãšããã°ããã§ããããïŒ
ïŒã¢ãã«ã®åµé æ§ãšã¯ãªã«ãã«ã€ããŠå¥ã«æ€èšããŠããŸããã https://qiita.com/kzuzuo/items/756470e6e17c54aa5e2e ïŒçŸæç¹ã§ã®èªåãªãã®é©åœãªçµè«ã¯ãåãæã«ç¢ºä¿¡é«ãèªèãããããšãã§ãïŒåµé çãªäŸ¡å€ã«ãã©ãçããã©ããã¯ããã®åé¡ã§ãïŒã¢ãã«ã«å¿
èŠãªæ©èœèªäœã¯åã«ïŒéãã瀺ãããšãã§ããã°ïŒååãªè¡šçŸèœåãããã°ïŒè¶³ãããšãããŸãïŒçžé¢ã®ãã¡äŸ¡å€ãçããã®ã¯å æã®ã¿ã§ããïŒïŒå®å
šãªããŒã¿é§åãæèããªãéãïŒãããçžé¢ããå æãèªã¿åãã°è¶³ããŸãïŒãããåæãšåºæ¥ããªãïŒã¢ãã«ãå æã瀺ããªããã°è¶³ããªããšããããã§ã¯ãããŸããïŒåæè
ããŸãã¯åæé²èЧè
ãããã®ããããã®ç¥èãå
ã«ããŠã䟡å€ãäœãã®ã§ããïŒåæãã¢ãã«ã䟡å€ã®ãã¹ãŠãäœãã®ã§ã¯ãããŸããïŒ é¡§å®¢ã䟡å€ãäœãã®ã§ããéçºè
ã䟡å€ãäœãããã§ã¯ãªããšã®èãæ¹ã«äŒŒãŠããŸãïŒã¢ãã«ãã䟡å€ãåŒãåºããªãã®ã¯ïŒãã®ã¢ãã«ã®è¡šçŸèœåãå¥ãšããã°ïŒããã®åé¡ã§ãïŒïŒ
ïŒæ
å ±ãšã¯ããŒã¿ã§ããç¥èãšã¯é¢æ°ã§ãããšããïŒåæãšã¯åºåãããæ
å ±ã§ã¯ãªããç¥èãã®ãã®ãèŠããããšã§ããã®ã§ïŒåãæã®ç¥èãšã®ç
§ããåãããå¿
èŠãšãªã£ãŠããïŒãšããã°ããã®ã§ããããïŒ
ïŒçµå±ã®æãããããªãããšãããããªãç¶æ
ãå€åãããããšã課é¡ãšãªãããã®ææ³ãšããŠãæéãããããããåŠç¿ä»¥å€ã®ææ³ã¯ãªãã®ã§ããããã
ç§ã¯ãã¢ãã«ã®èªç¥ã«ãããããã§ãããšä¿¡ããã
ãããŠãããããã䜿ããããªãèšèã ããããããDXã®ãŽãŒã«ãªã®ã ãšæãã
ïŒamplifiedãç¹èš±åºã¯å°ã£ãŠããïŒã¯ãïŒ
https://www.amplified.ai/ja/blog/29062020/why-patent-ai-2/
ãUSPTOã§ãããèªåããŒã«ãäœãã»ã©ã«å
è¡æè¡èª¿æ»ã«èª²é¡ãæããŠããããšããçè§£ããã ããã®ã§ã¯ãªãã§ããããïŒ
ãããŠãUSPTOãAIã䜿ããšããææè¡šç€ºãããŠããŸããããã®ãã©ãã€ã ã·ããã«ãããŠã¯ã調æ»ãããç§ãã¡èªèº«ãAIã䜿ãããªãã¹ãã«ãå¿
èŠã«ãªãããšãããã£ãŠããã®ã§ããã
ïŒããŒã¿ã¢ããªã¹ãã®å§¿å¢ãšéå
·ç®±
https://note.com/valuenex/n/n5acf3384a7a1
ïŒç€ŸäŒã«ããã忣çãã€ãºæšè«ãšããŠã®èšå·åµçºãïœéåçäºæž¬ç¬Šå·åãšããŠã®èšèªèгïœ
https://www.youtube.com/@anlpyoutubechannel7888/streams
https://www.dropbox.com/s/2kbu0m5jc1lo1xc/230315nlp23-taniguchi_cut_still_small.pdf?dl=0
ïŒçšèªã¯å匷ãã€ã€çµ±äžããŠãããŸã
ïŒæ£®äžè£äž (æ¡å±±åŠé¢å€§)ãèšéçãªèªã®æå³åæããèŠç¹ãšäž»èŠ³ã¯æããããã®ã
ïŒå€§èŠæš¡èšèªã¢ãã«ã®ç¥èãè£å®ããããã® Retriever ã®ç޹ä»
https://tech.acesinc.co.jp/entry/2023/03/31/121001
ïŒãã¯ã¹ãŠã§ã«èŠãå©çšããç¶²èæåœ±åã¢ã€ãŠã§ã¢æè¡ãåçŒéšçŸæ£å¯Ÿå¿æè¡ãšããŠããããªãã®ã§ãããç¶²èãªã©åŸçŒéšçŸæ£ã®éèŠæ§ãèæ
®ããã«ãå人ç課é¡ã«ãããŠè¶³ããªããšããã§ãã
ïŒæŠå¿µãã¯ãã«ã粟å¯ãªè³æ³¢ã«çœ®ãæããŠãè³æ³¢ããçŽæ¥æç« çæã§ããããã«ãªã£ããé¢çœãã§ãããåèªã¬ãã«ã«è§£èªãèšèªã¢ãã«ã§ã¯ä¿®æ£ããæ¹æ³ãããããã§ããçµæ§æåŸ
ã§ãããšæããŠããŸãã
ïŒã»ãŒåãææ³ã§ãããè¡åãã©ããªæ³ä»€ãã«ãŒã«ã«æµè§Šããããæç€ºãããã³ã³ãã©ã€ã¢ã³ã¹éµå®ããŒã«ãäœãããã§ãããã
æ³ä»€ãã«ãŒã«ã«åŸãããšæã£ãŠããŠããå¿ããŠãããåŠç¿è² è·ãé«ãã察åŠã§ããªãããšã¯ãããé¢é£ããã«ãŒã«çãæç€ºããããŒã«ã¯éåžžã«æçšã«æããç¹ã«ãããŒã«ã«ã«ãŒã«ã«æè»ã«å¯Ÿå¿ã§ããã°ãªãæçšã ããã
ã»èšèªã®åºææ¬¡å
ãæž¬ã
âäžç°äº® (æ±å€§), 暪äºç¥¥ (æ±å倧/çç )
https://www.anlp.jp/proceedings/annual_meeting/2024/pdf_dir/E6-1.pdf
æ¹åãä¿ã¡ããåºææ¬¡å
ã¯æå€ãšå°ããïŒ10次å
ãããïŒãword2vecã®çµéšè«ããèšããš200次å
ã¯å¿
èŠã ãïŒ12次å
çšåºŠãŸã§ãªãæ
å ±æå€±ãªãå§çž®ããããããªç€ºåãåŸãŠãè¯ãã®ããªïŒ
ãŸãïŒçµéšè«ã®äœåãªæ¬¡å
ã¯ãæŒç®ã«ããå¹³åã誀ã£ãŠé¡äŒŒããªãããã«ããã«ååãªæŒç®å¯èœã§ãããã¯ãã«ç©ºéããäœãã«å¿
èŠãªã®ããïŒ
äœå°æ¬¡å
ã¯ïŒæå³ç©ºéãé©åã«è¡šçŸããã®ã«å¿
èŠãªïŒãã现ããææ
ã®éããæèã®ãã¥ã¢ã³ã¹ããããã¯ç¹å®ã®ç¶æ³ãã³ã³ããã¹ãã«é¢é£ããæ
å ±ãå«ãã®ãããããªãïŒ
ã»å¹³åããŒãªã³ã°ã«ããæåã蟌ã¿ã®åæ€èš: å¹³åã¯ç¹çŸ€ã®èŠçŽãšããŠååã?
âåç¥æ£, æ ç°å®äºº (æ±å倧), 暪äºç¥¥ (æ±å倧/çç ), 也å¥å€ªé (MBZUAI/æ±å倧/çç )
ãhttps://www.anlp.jp/proceedings/annual_meeting/2024/pdf_dir/A10-4.pdf
次å
ãå¢ããã°å¹³åã¯é©åã«é¢ãããããªãïŒèªåçã«æã®èª€ã£ãæ··åã¯å°ãªããªãã®ã§ã¯ãªãã®ã ãããïŒïŒãã ïŒæ¬¡å
ãå¢ããã ãã§ã¯åé¡ã解決ããªãããšã¯ïŒword2vecã®éå»å®éšã§äºæ³ã§ãããã ãïŒåé¡ã®å²åã¯æžãã®ã§ã¯ïŒïŒ
word2vecã®å¹³åããŒãªã³ã°ã§ã¯ïŒäžå¿
èŠãªåèªã®éã¿ïŒåå²ããªã衚çŸã§èšãã°ãã«ã ïŒãèæ
®ããŠããŸãããïŒèª€ã£ãå¹³åã«å°ããããããºã¬ã倧ããã®ã§ã¯ïŒãdoc2vecã®æ§èœã®äœããšåãã§ïŒïŒäŸãã°å©è©ã®ã¿ãå¹³åããã°äŒŒãŠããã ããïŒ
æè¿ã®ã¢ãã«ã§åé¡ãèµ·ããªãçç±ã¯ïŒãã髿¬¡å
ã䜿ã£ãŠããïŒãã€éã¿ãããæé©ã«éžãã§ããããïŒã§ã¯ãªããïŒ
ã»éšåç©ºéæ³ã«çæ³ãåŸãTransformerã®ã¢ãã³ã·ã§ã³ãããã«ãããç¹åŸŽæœåº
âåç°æåŒ (JAIST), é³¥å±
æéЬ (æ±äº¬é»æ©å€§), æ¥é«æå¹³ (JAIST), 倧颿Žå¹³ (æ±å€§)
ã»RLHFã匷ååŠç¿ã«ãã èªç¥ãã£ãŒãããã¯
èªç¥ãã£ãŒãããã¯ïŒçŒçéåã»è³æ³¢ã«ããå€§èŠæš¡èšèªã¢ãã«ã®åŒ·ååŠç¿
âåç°å®¥éœ, 倧颿Žå¹³ (æ±å€§)
https://www.anlp.jp/proceedings/annual_meeting/2024/pdf_dir/D7-2.pdf
èªç¥ãã¡ã€ã³ãã¥ãŒãã³ã°ïŒçŒçéåã«ããå€§èŠæš¡èšèªã¢ãã«ã®ãã¡ã€ã³ãã¥ãŒãã³ã°
æè°·å€§æ²³, â倧颿Žå¹³ (æ±å€§)
https://www.anlp.jp/proceedings/annual_meeting/2024/pdf_dir/D7-5.pdf
ãèªã¿ããããèªåæ¡ç¹ãªã©
ã»å·¥åŠçæ§èœãšäººéãããã®é¢ä¿ã¯ããŒã¯ã³åå²ã«äŸåãã
âäžèŒªæ¬å€ª, åç°éŒ, 倧颿Žå¹³ (æ±å€§)
ãhttps://www.anlp.jp/proceedings/annual_meeting/2024/pdf_dir/D7-6.pdf
人éããããšïŒåèªåå²ã®é·ãã®é¢ä¿
















