AIãšããŒã¿ãµã€ãšã³ã¹ãããžãã¹ã«é©åœãèµ·ãããŠããä»ããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ãã®äžå¿çæè¡ã§ããããããããŸã ã«å€ãã®ããžãã¹ããŒãœã³ã«ãšã£ãŠããã®åºæ¬æ§é ãå¿çšç¯å²ã¯è¬ã«å ãŸããŠããŸãããã®èšäºã§ã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯å ¥éããå§ããŠãåºæ¬çãªæ§é ãåŠç¿ããã»ã¹ããããŠãã®å¹ åºãå¿çšç¯å²ã«è³ããŸã§ããããããã解説ããŠãããŸãããŸããä»åŸã®èª²é¡ãšãã¥ãŒã©ã«ãããã¯ãŒã¯æè¡ã®æªæ¥ã«ã€ããŠãèå¯ããŸããç»åèªèããèªç¶èšèªåŠçãé³å£°èªèã«è³ããŸã§ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®å¯èœæ§ãæ倧éã«æŽ»ããããã®ç¥èãæäŸããŸããããããAIãããžãã¹ã«åãå ¥ããããšèããŠããçããã«ãšã£ãŠã貎éãªç¬¬äžæ©ãšãªãã§ãããã
ãã¥ãŒã©ã«ãããã¯ãŒã¯å ¥é
ãã¥ãŒã©ã«ãããã¯ãŒã¯ãšã¯äœãïŒ
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ã人éã®è³ã®åããæš¡å£ããŠèšèšãããã³ã³ãã¥ãŒã¿ã¢ã«ãŽãªãºã ã®äžçš®ã§ããããã¯ãæ©æ¢°åŠç¿ãšäººå·¥ç¥èœã®åéã§åºãçšããããŠãããããŒã¿ã®ãã¿ãŒã³ãé¢ä¿æ§ãåŠç¿ããããšãã§ããŸãããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ãæ å ±åŠçã®ããã®å€æ°ã®ããŒãïŒãã¥ãŒãã³ãšåŒã°ããïŒãçžäºã«çµåããŠæ§ç¯ãããŠããŸãã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®äž»ãªç¹åŸŽã¯ãããŒã¿ã»ããããèªåçã«åŠç¿ããäºæž¬ãåé¡ãªã©ã®åé¡è§£æ±ºã«å¿çšããããšãã§ããç¹ã§ããè€éãªé¢æ°ãè¿äŒŒããèœåãæã€ãããç»åèªèãèšèªç解ãã²ãŒã ãã¬ã€ãªã©æ§ã ãªã¿ã¹ã¯ã§åè¶ããæ§èœãçºæ®ããŸãã
ãã®æè¡ã®é²åã«ããã以åã¯äžå¯èœã ãšæãããŠããå€ãã®åé¡ãçŸåšã§ã¯è§£æ±ºå¯èœã«ãªããŸããããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ãæ¥ã ã®ç掻ããã䟿å©ã«ããå€ãã®ç£æ¥ã«é©åœããããããŠããŸãã
æŽå²ãšçºå±ã®æŠèŠ
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®ç 究ã¯ã1940幎代ã«é¡ããŸããåæã®æ®µéã§ã¯ãçç©åŠçãªç¥çµç³»ã®ä»çµã¿ãæ°åŠçã«ã¢ãã«åããŠç解ããããšããè©Šã¿ãå§ãŸããŸãããåœæã®æè¡ãèšç®èœåã®éçã«ãããç 究ã¯é²å±ã«ä¹ããã£ããã®ã®ãåºç€çãªçè«ãã¢ãã«ã圢æãããŸããã
1980幎代ã«å ¥ããšãããã¯ãããã²ãŒã·ã§ã³ã¢ã«ãŽãªãºã ã®ææ¡ãããŒããŠã§ã¢ã®é²åã«ããããã¥ãŒã©ã«ãããã¯ãŒã¯ç 究ã¯å€§ããåé²ããŸãããããã«ãããããæ·±ãå±€ãæã€ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®èšç·ŽãçŸå®çã«ãªãããã®å¿çšç¯å²ãåºããå§ããŸããã
è¿å¹Žã§ã¯ãããŒã¿ã®ççºçå¢å ãšGPUïŒã°ã©ãã£ãã¯åŠçãŠãããïŒã®é²åã«ããããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ããã«é«åºŠãªåœ¢æ ãžãšé²åããŠããŸãã倧éã®ããŒã¿ãåŠçããåŠç¿ããèœåãåäžããå€ãã®é©æ°çãªã¢ããªã±ãŒã·ã§ã³ãå®çŸãããŠããŸãã
åºæ¬çãªæ§é ãšä»çµã¿
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®åºæ¬çãªæ§é ã¯ãå ¥åå±€ãé ãå±€ãåºåå±€ã®3ã€ã®äž»èŠãªå±€ããæãç«ã£ãŠããŸããå ¥åå±€ã«ããŒã¿ãäŸçµŠããããšãé ãå±€ãéããŠåŠçãããæçµçã«åºåå±€ã§çµæãåºåãããŸããåå±€ã¯è€æ°ã®ãã¥ãŒãã³ã§æ§æãããŠããããããããç¬èªã®ãŠã§ã€ãïŒéã¿ïŒãæã£ãŠããŸãã
åŠç¿éçšã§ã¯ãããŒã¿ã»ããå ã®å ¥åãšããã«å¯ŸããæåŸ ãããåºåãåºã«ããã¥ãŒãã³éã®ãŠã§ã€ãã調æŽãããŸãããã®éçšãéããŠããããã¯ãŒã¯ã¯ããŒã¿å ã®ãã¿ãŒã³ãåŠç¿ããäºæž¬ãåé¡ãè¡ãããã®ç¥èãèç©ããŸãã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã§æãéèŠãªããã»ã¹ã®äžã€ãããšã©ãŒãæ倱é¢æ°ãæå°åããããã®ããã¯ãããã²ãŒã·ã§ã³ã§ããããã¯ãåºåå±€ã§åŸãããçµæãšæåŸ ãããçµæãšã®å·®ç°ãé ãå±€ãéããŠéæ¹åã«äŒæããããŠã§ã€ããæé©åããããã»ã¹ã§ãã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®çš®é¡
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã«ã¯è€æ°ã®çš®é¡ãååšãããããããç¹å®ã®ã¿ã¹ã¯ãåé¡ã«æé©åãããŠããŸãããã£ãŒããã©ã¯ãŒããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯æãåºæ¬çãªã¿ã€ãã§ãããããŒã¿ãå ¥åå±€ããåºåå±€ãžãšäžæ¹åã«æµããŸãã
ãªã«ã¬ã³ããã¥ãŒã©ã«ãããã¯ãŒã¯ïŒRNNïŒã¯ãæéã«äŸåããããŒã¿ãã·ãŒã±ã³ã·ã£ã«ãªæ å ±ãæ±ãã®ã«é©ãããããã¯ãŒã¯ã§ããRNNã§ã¯ãéå»ã®æ å ±ãã«ãŒããéããŠé ãå±€ã«ä¿æããããããããã¹ããé³å£°ãªã©ã®é£ç¶ããŒã¿ã®åŠçã«æå¹ã§ãã
ãŸãã倧èŠæš¡ãªç»åèªèã¿ã¹ã¯ã«åºãå©çšãããç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ïŒCNNïŒã¯ãç»åå ã®å±æçãªãã¿ãŒã³ãæããç¹åããå±€ãæã£ãŠããŸããããã«ãããCNNã¯ç»åããŒã¿ã®ç¹åŸŽãå¹ççã«åŠç¿ããé«ç²ŸåºŠãªåé¡ãèªèãå®çŸããŸãã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®åºæ¬æ§é
ãã¥ãŒãã³ïŒåºæ¬åäœã®ç解
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®äžçã«ãããŠããã¥ãŒãã³ã¯ãã®æåºæ¬ã®åäœã§ããçç©åŠã«ãããè³çŽ°èã«è§Šçºããããã®ã¢ã€ãã¢ã¯æ å ±åŠçã®åºæ¬çãªèŠçŽ ãšããŠæ©èœããŸããåãã¥ãŒãã³ã¯è€æ°ã®å ¥åãåãåããããããå å·¥ããŠåºåãçæããŸãããã®ããã»ã¹ã¯ãè³å ã®ãã¥ãŒãã³ãæ å ±ãåŠçããä»çµã¿ã«éåžžã«äŒŒãŠããŸãã
ãã¥ãŒãã³å ã®åŠçã¯åçŽãªãããéåããããšã§è€éãªèšç®ãå¯èœã«ãªããŸããå ¥åä¿¡å·ã¯éã¿ä»ãããããããéã¿ä»ãä¿¡å·ã¯åç®ãããŸãããã®çµæããã¥ãŒãã³ã®åºåãšãªãã次ã®å±€ãžãšäŒéãããæµãã§ãããã®äžé£ã®æµããããã¥ãŒã©ã«ãããã¯ãŒã¯å ã§æ å ±ãäŒæããåºæ¬çãªã¡ã«ããºã ãæ§ç¯ããŸãã
éèŠãªã®ã¯ãåäžã®ãã¥ãŒãã³ã§ã¯ãªãããã¥ãŒãã³ã®è€éãªãããã¯ãŒã¯ãçã¿åºãçžäºäœçšã§ãããã®çžäºäœçšããæ©æ¢°åŠç¿ã¢ãã«ã«ãšã£ãŠã®åŠç¿èœåãçã¿åºããŸãããã¥ãŒãã³äžã€äžã€ã®åãããæçµçãªã¢ãã«ã®æ§èœãå·Šå³ããèŠå ãšãªãã®ã§ãã
掻æ§åé¢æ°ïŒãã¥ãŒãã³ã®åºåãã©ã決ããã
次ã«ããã¥ãŒãã³ã®åºåã決å®ããããã«äœ¿çšãããã®ã掻æ§åé¢æ°ã§ãããã®é¢æ°ã¯ãå ¥åã®åèšãäžå®ã®éŸå€ãè¶ ãããšãã«ããã¥ãŒãã³ãçºç«ãããã©ããã決ãã圹å²ãæ ããŸããäžè¬çã«ã¯ãã·ã°ã¢ã€ãé¢æ°ãReLUïŒRectified Linear UnitïŒé¢æ°ãtanhïŒåæ²ç·æ£æ¥é¢æ°ïŒãªã©ãããçšããããŸãã
掻æ§åé¢æ°ã¯ãéç·åœ¢æ§ããããã¯ãŒã¯ã«å°å ¥ããéèŠãªåœ¹å²ãæã¡ãŸããããã«ããããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ç·åœ¢åé¢å¯èœã§ãªãããè€éãªåé¡ãæ±ãããšãå¯èœã«ãªããŸããã€ãŸããåçŽãªçŽç·ãå¹³é¢ã ãã§ãªããæ²ç·ãæ²é¢ãçšããŠããŒã¿ãåºåãããããšãã§ããããã«ãªãã®ã§ãã
掻æ§åé¢æ°ã¯ãããããç°ãªãç¹æ§ãæã¡ãŸããããšãã°ãReLUé¢æ°ã¯è² ã®å€ã«å¯ŸããŠã¯0ãåºåããæ£ã®å€ã«å¯ŸããŠã¯ãã®å€ããã®ãŸãŸåºåãããããèšç®ã®é«éåãšå¹ççãªåŠç¿ãæåŸ ã§ããŸãããããã®é¢æ°ãé©åã«éžæããããšã§ãã¢ãã«ã®ç²ŸåºŠãšåŠç¿é床ãæé©åããããšãå¯èœã§ãã
å±€ã®æ§é ïŒå ¥åå±€ãé ãå±€ãåºåå±€
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ãå€æ°ã®ãã¥ãŒãã³ãå±€ãæããŠæ§æãããŠããŸããå ·äœçã«ã¯ãå ¥åå±€ãäžã€ãŸãã¯è€æ°ã®é ãå±€ããããŠåºåå±€ããæãç«ã£ãŠããŸããå ¥åå±€ã§ã¯ãå€éšããã®ããŒã¿ããããã¯ãŒã¯ã«éã蟌ã¿ãŸããé ãå±€ã¯å éšã§è€éãªç¹åŸŽãæœåºããæçµçã«åºåå±€ãåé¡ã®è§£çãæäŸããŸãã
é ãå±€ã®éèŠæ§ã¯ããããã¯ãŒã¯ãããŒã¿ããããæœè±¡çãªç¹åŸŽãåŠç¿ããèœåã«ãããŸããé ãå±€ãå€ããã°å€ãã»ã©ããããã¯ãŒã¯ã¯è€éãªåé¡ãç解ãã解決ããèœåãé«ãŸããŸããããããå±€ãå€ããããšåŠç¿ãå°é£ã«ãªã£ãããéåŠç¿ãèµ·ããå¯èœæ§ããããŸãã
åºåå±€ã§ã¯ããããã¯ãŒã¯ã®æçµçãªå€æãäºæž¬ãè¡ãããŸãããã®å±€ã®èšèšã¯ã解決ãããåé¡ã®æ§è³ªã«ãã£ãŠå€§ããç°ãªããŸããåé¡åé¡ã§ã¯ãåºåå±€ã®ãã¥ãŒãã³æ°ã¯ã¯ã©ã¹ã®æ°ã«çžåœããååž°åé¡ã§ã¯äžã€ã®å€ãåºåãããŸããåé¡ã«å¿ããé©åãªæ§é èšèšããé«ç²ŸåºŠãªã¢ãã«æ§ç¯ã®éµãšãªããŸãã
éã¿ãšãã€ã¢ã¹ïŒåŠç¿ã®ããŒãšãªãèŠçŽ
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®åŠç¿éçšã§ã¯ãéã¿ãšãã€ã¢ã¹ãæãéèŠãªèŠå ã§ãããããã¯ãåãã¥ãŒãã³ã®å ¥åä¿¡å·ã«å¯Ÿããåå¿ã®åŒ·ããšæ¹åã調æŽãã圹å²ãæã¡ãŸããåŠç¿ã¯ãäžããããããŒã¿ã«å¯ŸããŠæé©ãªéã¿ãšãã€ã¢ã¹ãèŠã€ããéçšãšèšããŸãã
ãã®åŠç¿éçšã¯ãäž»ã«èª€å·®éäŒææ³ïŒããã¯ãããã²ãŒã·ã§ã³ïŒãšåŒã°ããã¢ã«ãŽãªãºã ãçšããŠè¡ãããŸãããã®æ¹æ³ã§ã¯ãåºåå±€ããå ¥åå±€ã«åãã£ãŠã誀差ãéäŒæãããªããéã¿ãšãã€ã¢ã¹ã®æŽæ°ãè¡ããŸãããã®éçšãç¹°ãè¿ãããšã§ããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ã¿ã¹ã¯ãããæ£ç¢ºã«å®è¡ããããã«ãªããŸãã
ãã€ã¢ã¹ã¯ãå ¥åããã¹ãŠ0ã§ãã£ããšããŠããã¥ãŒãã³ã掻æ§åãããã©ããã決å®ããŸããéã¿ãšãã€ã¢ã¹ã®é©åãªèª¿æŽã«ããããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ããè€éã§åŸ®åŠãªããŒã¿ã®ãã¿ãŒã³ãæããèœåãç²åŸããŸããçµå±ããããã®èŠçŽ ããããã¯ãŒã¯ã®åŠç¿èœåãšæ§èœã倧ããå·Šå³ããã®ã§ãã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®åŠç¿ããã»ã¹
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ã人éã®è³ãæ å ±ãåŠçããä»çµã¿ã«ãã³ããåŸãŠéçºãããèšç®ã¢ãã«ã§ãããã®åŠç¿ããã»ã¹ã¯ãããŒã¿ãå¹ççã«è§£æããäºæž¬ãè¡ãããã®ããŒãšãªããŸãããã¥ãŒã©ã«ãããã¯ãŒã¯ã®åŠç¿ããã»ã¹ãç解ããããšã§ããããã®ã·ã¹ãã ãã©ã®ããã«ããŠè€éãªåé¡ã解決ããŠããã®ããç解ããããšãã§ããŸãã
é äŒæãšéäŒæã®ããã»ã¹
åŠç¿ããã»ã¹ã¯å€§ããåããŠäºã€ã®ã¹ããããé äŒæ(Forward Propagation)ãšéäŒæ(Backpropagation)ããã»ã¹ãããªããŸããé äŒæã§ã¯ãå ¥åããŒã¿ããããã¯ãŒã¯ãéããŠäŒéãããåå±€ãéãããšã«éã¿ä»ãã®å ç®ãšéç·åœ¢ã®æŽ»æ§åé¢æ°ãçµãŠãæçµçãªåºåãåŸãããŸãã
äžæ¹ã§ãéäŒæã¯åŸãããåºåãšæ£è§£ã©ãã«ãšã®å·®ãããªãã¡èª€å·®æ å ±ãçšããŠããããã¯ãŒã¯ã®éã¿ã調æŽããŸãããã®ããã»ã¹ã§ã¯ãåŸé éäžæ³ãçšããŠã誀差ãæå°éã«ãªãããã«ãããã¯ãŒã¯ã®éã¿ãæŽæ°ããŸãã
ãããã®ããã»ã¹ãç¹°ãè¿ãããšã§ããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ããŒã¿ãããã¿ãŒã³ãåŠç¿ããäºæž¬ãæ¹åããŠãããŸãã
æ倱é¢æ°ïŒããã©ãŒãã³ã¹ã®æž¬å®
æ倱é¢æ°ïŒLoss FunctionïŒã¯ããããã¯ãŒã¯ã®åºåãšæ£è§£ã©ãã«ãšã®èª€å·®ãèšç®ããŸãããã®å€ã¯ãã¢ãã«ãã©ãã»ã©ã®ããã©ãŒãã³ã¹ãçºæ®ããŠãããã枬å®ããããã®éèŠãªææšã§ããããçšããããæ倱é¢æ°ã«ã¯ãå¹³åäºä¹èª€å·®(MSE)ã亀差ãšã³ããããŒèª€å·®ããããŸãã
æ倱é¢æ°ã®å€ãå°ãããªãã»ã©ãã¢ãã«ã®äºæž¬ç²ŸåºŠã¯é«ããšè©äŸ¡ãããŸããéäŒæããã»ã¹ã§ã¯ããã®æ倱é¢æ°ã®åŸé ãå©çšããŠããããã¯ãŒã¯ã®éã¿ã®æŽæ°ãè¡ããŸãã
æ倱é¢æ°ã®éžæã¯ã察象ãšããåé¡ã®çš®é¡ã«äŸåããŸããäŸãã°ãäºå€åé¡åé¡ã«ã¯äº€å·®ãšã³ããããŒèª€å·®ããé£ç¶å€ã®äºæž¬ã«ã¯å¹³åäºä¹èª€å·®ãé©ããŠããŸãã
æé©åã¢ã«ãŽãªãºã ïŒåŠç¿ããã»ã¹ã®æ¹å
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®åŠç¿å¹çãåäžãããããã«ã¯ãæé©åã¢ã«ãŽãªãºã ã®éžæãéèŠã§ããæé©åã¢ã«ãŽãªãºã ã¯ãæ倱é¢æ°ã®æå°åãå¹ççã«è¡ãææ³ã§ãã代衚çãªã¢ã«ãŽãªãºã ã«ã¯ã確ççåŸé éäžæ³(SGD)ãã¢ãŒã¡ã³ã¿ã ãAdamãªã©ããããŸãã
ããããã®ã¢ã«ãŽãªãºã ã¯ãéã¿ã®æŽæ°æ¹æ³ãåŠç¿çã®èª¿æŽæ¹æ³ã«ç¹åŸŽããããåé¡ã®çš®é¡ãããŒã¿ã®ç¹æ§ã«å¿ããŠé©åãªãã®ãéžæããããšãéèŠã§ãã
æé©åã¢ã«ãŽãªãºã ãé©åã«éžæããããšã§ãåŠç¿ããã»ã¹ã®é床ãåäžãããã ãã§ãªããåŠç¿ãåæãããããªããããé«ãããã©ãŒãã³ã¹ãéæããããšãã§ããŸãã
éåŠç¿ãšæ±åæ§èœ
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®åŠç¿ã«ãããŠã¯ãéåŠç¿(Overfitting)ãé²ãããšãéèŠã§ããéåŠç¿ã¯ãèšç·ŽããŒã¿ã«å¯ŸããŠéå°ã«é©å¿ããããããšã§ãæ°ããããŒã¿ã«å¯Ÿããæ±åæ§èœãäœäžããçŸè±¡ãæããŸãã
éåŠç¿ãé²ãããã®ææ³ã«ã¯ãæ£ååïŒL1ãL2æ£ååïŒãããããã¢ãŠããæ©æçµäºãªã©ããããŸãããããã®ææ³ãé©åã«å©çšããããšã§ãã¢ãã«ã®æ±åæ§èœãé«ããããšãã§ããŸãã
ãŸããããŒã¿ã®å€æ§æ§ãå¢ããããšããããŒã¿æ¡åŒµ(Data Augmentation)ãè¡ãããšããæ±åæ§èœãåäžãããæå¹ãªæ段ã§ããé©åãªææ³ãéžæãããã©ã¡ãŒã¿ã調æŽããããšã§ãããå ç¢ãªãã¥ãŒã©ã«ãããã¯ãŒã¯ã¢ãã«ãæ§ç¯ããããšãå¯èœã«ãªããŸãã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®å¿çšç¯å²
ç»åèªèãšã³ã³ãã¥ãŒã¿ããžã§ã³
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ãç»åèªèãšã³ã³ãã¥ãŒã¿ããžã§ã³ã®åéã§å€§ããªé©åœããããããŠããŸãããããã®æè¡ã¯ãããžã¿ã«ç»åã解æããå 容ãèå¥ããããã«äœ¿çšãããŸããããšãã°ãé¡èªèã·ã¹ãã ã¯ãåçããããªãã人ã®é¡ãèå¥ããããã«ãã¥ãŒã©ã«ãããã¯ãŒã¯ãçšããŠããŸãã
ãã®å¿çšç¯å²ã¯ãã»ãã¥ãªãã£ããœãŒã·ã£ã«ã¡ãã£ã¢ãå»çç»å蚺æãªã©ãå€æ¹é¢ã«ããããŸããå»çåéã§ã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ã䜿çšããŠãXç·ç»åãMRIããç°åžžãæ€åºããããšãã§ãã蚺æã®ç²ŸåºŠãé£èºçã«åäžãããŠããŸãã
èªåé転è»ã®éçºã«ãããŠããã³ã³ãã¥ãŒã¿ããžã§ã³ãšãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯äžå¯æ¬ ãªæè¡ã§ããè»äž¡ã¯ãåšå²ã®ç°å¢ãæ£ç¢ºã«ç解ããé©åãªè¡åãåãããã«ããããã®æè¡ã«å€§ããäŸåããŠããŸãã
èªç¶èšèªåŠçïŒNLPïŒ
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ãèªç¶èšèªåŠçïŒNLPïŒã®åéã§ãé¡èãªææãäžããŠããŸããNLPã¯ãã³ã³ãã¥ãŒã¿ã人éã®èšèªãç解ããçæããæè¡ã§ããããã«ãããããã¹ãããŒã¿ããã®æ å ±æœåºãæ©æ¢°ç¿»èš³ã質åå¿çã·ã¹ãã ãªã©ãå¯èœã«ãªããŸãã
ç¹ã«ããã©ã³ã¹ãã©ãŒããŒã¢ãã«ã®åºçŸã¯ãææžèŠçŽãäŒè©±åAIã®ç²ŸåºŠãåçã«åäžãããããžãã¹ãæè²ãªã©å€ãã®åéã§ã®æŽ»çšãé²ãã§ããŸãããã¥ãŒã©ã«ãããã¯ãŒã¯ãçšããNLPæè¡ã¯ããŠãŒã¶ãŒã€ã³ã¿ãŒãã§ãŒã¹ã®é©æ°ãšããŠãæåŸ ãããŠããŸãã
ãŸããææ åæã«ãããŠãããã¥ãŒã©ã«ãããã¯ãŒã¯ãçšããããšã§ãããã¹ããããŠãŒã¶ãŒã®ææ ãèªã¿åãããšãã§ãã顧客æºè¶³åºŠã®åäžãåžå Žåæã«åœ¹ç«ãŠãããŠããŸãã
é³å£°èªèãšé³å£°åæ
é³å£°èªèã§ã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ããã®åšåãçºæ®ããŠããŸããé³å£°ã¢ã·ã¹ã¿ã³ããé³å£°ã³ãã³ãã·ã¹ãã ã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ãåºç€ãšããŠããŠãŒã¶ãŒã®é³å£°æ瀺ãããã¹ãã«å€æããŸãããŸããé³å£°åæã®åéã§ã¯ãèªç¶ãªçºè©±ãçæããããã«ãã¥ãŒã©ã«ãããã¯ãŒã¯ã䜿çšãããŠããŸãã
æè¿ã§ã¯ããªã¢ã«ã¿ã€ã é³å£°ç¿»èš³ããã¥ãŒã©ã«ãããã¯ãŒã¯ãçšããŠå®çŸãããŠãããç°ãªãèšèªã話ã人ã ã®éã®ã³ãã¥ãã±ãŒã·ã§ã³ã®éå£ãåãé€ããŠããŸããããã«ãããåœéäŒè°ãå€èšèªå¯Ÿå¿ãµãŒãã¹ã®å質ãåäžããŠããŸãã
ããã«ããã¥ãŒã©ã«ãããã¯ãŒã¯ã«ããé³å£°åæã¯ãå»çåéã§ã®å¿çšãèŠåºãããŠããŸããäŸãã°ãç¹å®ã®å£°ã®ç¹åŸŽããå¥åº·ç¶æ ãåæããããšãå¯èœã«ãªããæ©æçºèŠã蚺æã®æ°ããæ段ãšãªã£ãŠããŸãã
äºæž¬ãšããŒã¿åæ
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ãäºæž¬ãšããŒã¿åæã®ç²ŸåºŠãå€§å¹ ã«æ¹åããŠããŸããéèåžå Žã®ååäºæž¬ãåšåº«ç®¡çããšãã«ã®ãŒæ¶è²»äºæž¬ãªã©ãå€æ§ãªæ¥çã§æŽ»èºããŠããŸãã
深局åŠç¿æè¡ã®é²æ©ã«ããã倧éã®ããŒã¿ããè€éãªãã¿ãŒã³ãæœåºããããæ£ç¢ºãªäºæž¬ãå¯èœã«ããŠããŸãããã®æè¡ã¯ããªã¹ã¯ç®¡çãææ決å®ããã»ã¹ããµããŒãããããã«ç¹ã«äŸ¡å€ããããŸãã
ãŸããããŒã¿åæã®åéã§ã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ãæ°ããªç¥èŠãçºèŠããæ段ãšããŠå©çšãããŠããŸããäŸãã°ã顧客è¡åã®åæãéããŠãããŒã±ãã£ã³ã°æŠç¥ã補åéçºã«çããããã€ã³ãµã€ããæäŸããŠããŸãã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®èª²é¡ãšæªæ¥
解éå¯èœæ§ã®åé¡
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®æè¡ãé²åããã«ã€ããŠãããã©ãã¯ããã¯ã¹ãã®åé¡ã泚ç®ãããŠããŸããã€ãŸãããã¥ãŒã©ã«ãããã¯ãŒã¯ãã©ã®ããã«ããŠæ±ºå®ãäžããŠããããç解ããããšãé£ãããšããåé¡ã§ããããã¯ãæ©æ¢°åŠç¿ã¢ãã«ãè€éã«ãªãã»ã©ãå éšã®åŠçã¡ã«ããºã ã人éã解éãã«ãããªãããã§ãã
äŒæ¥ãç 究è ãã¡ã¯ããã®åé¡ã«å¯ŸåŠããããã«éææ§ãšè§£éå¯èœæ§ãé«ããææ³ãéçºããŠããŸããããšãã°ãç¹å®ã®å€æãã©ã®ãããªããŒã¿ãšåŠçéçšã«åºã¥ããŠããããæããã«ããããšã§ãäºæž¬ã®æ ¹æ ãç解ããããããè©Šã¿ããããŸãã
ãããã解éå¯èœæ§ãé«ããããšã¯ãå¿ ãããã¢ãã«ã®ç²ŸåºŠãå¹çæ§ãåäžãããããã§ã¯ãããŸããããã®ããã解éå¯èœæ§ãšæ§èœã®ãã©ã³ã¹ãåãããšã¯ãä»åŸã®ç 究éçºã«ããã倧ããªèª²é¡ã®äžã€ã§ãã
ããŒã¿ãšãã©ã€ãã·ãŒã®æžå¿µ
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®åŠç¿ãšæ§èœåäžã«ã¯èšå€§ãªéã®ããŒã¿ãå¿ èŠãšãããŸããããã¯ãå人ã®ãã©ã€ãã·ãŒãå±éºã«æãå¯èœæ§ããããŸããç¹ã«ãå人æ å ±ãå«ãããŒã¿ãæ±ãå Žåã«ã¯ãæªæã®ããå©çšãæžå¿µããããããå³æ ŒãªããŒã¿ç®¡çãšãã©ã€ãã·ãŒä¿è·ãæ±ããããŸãã
ãã®åé¡ã«å¯Ÿå¿ãããããããŒã¿ãå¿ååããæè¡ããã©ã€ãã·ãŒãä¿è·ããããã®åŠç¿ææ³ãªã©ãç 究ãããŠããŸãããŸããããŒã¿ã®å©çšã«é¢ããæ³èŠå¶ãã¬ã€ãã©ã€ã³ã®æŽåãé²ãã§ããŸãã
ããããããŒã¿ã®åéãšå©çšæ¹æ³ã«é¢ããå«ççãªè°è«ã¯ãæè¡ã®é²æ©ã«äŒŽãåžžã«é²åããŠãããããé¢é£ããå šãŠã®ã¹ããŒã¯ãã«ããŒã«ãšã£ãŠç¶ç¶çãªè°è«ãšå¯Ÿçãå¿ èŠã§ãã
èšç®è³æºãžã®äŸå床
é«åºŠãªãã¥ãŒã©ã«ãããã¯ãŒã¯ã¢ãã«ã¯ãèšå€§ãªèšç®è³æºãå¿ èŠãšããŸããããã¯ã倧éã®ããŒã¿ã»ãããåŠçããã¢ãã«ãåŠç¿ãããããã«é«éãªåŠçèœåãšå€§å®¹éã®ã¡ã¢ãªãå¿ èŠã§ããããã§ãããã®ãããªèšç®è³æºãžã®äŸåã¯ãã¢ãã«ã®éçºãšéçšã³ã¹ããé«ããèŠå ãšãªããŸãã
èšç®è³æºãžã®äŸåã軜æžããããã«ãå¹ççãªã¢ãã«ã®èšèšãåŠç¿ã¢ã«ãŽãªãºã ã®æé©åãç 究ãããŠããŸãããŸããã¯ã©ãŠãã³ã³ãã¥ãŒãã£ã³ã°ãå°çšã®ããŒããŠã§ã¢ã®äœ¿çšã«ããããã¥ãŒã©ã«ãããã¯ãŒã¯ã¢ãã«ã®èšç®èœåãåäžãããè©Šã¿ããããŸãã
ããã«ãçãšãã«ã®ãŒãæèããã¢ãã«èšèšã«ãããç°å¢è² è·ã®äœæžãéèŠãªèª²é¡ãšãªã£ãŠããŸãããããã®åãçµã¿ã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯æè¡ã®æªæ¥ã«ãããŠãæç¶å¯èœæ§ãšå¹çæ§ã®äž¡æ¹ãè¿œæ±ããæ¹åæ§ã瀺ããŠããŸãã
次äžä»£ãã¥ãŒã©ã«ãããã¯ãŒã¯æè¡ã®å±æ
ãã¥ãŒã©ã«ãããã¯ãŒã¯æè¡ã¯ãåŒãç¶ãæ¥éã«é²åããŠããŸãã次äžä»£ã®æè¡ãšããŠã¯ãéåã³ã³ãã¥ãŒãã£ã³ã°ããã¥ãŒãã¢ã«ãã£ãã¯ã³ã³ãã¥ãŒãã£ã³ã°ãªã©ã泚ç®ãããŠããããããã®æè¡ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®èšç®é床ãšå¹çæ§ãé£èºçã«åäžãããå¯èœæ§ããããŸãã
ããã«ã深局匷ååŠç¿ããžã§ãã¬ãŒãã£ãã»ã¢ãããŒãµãªã¢ã«ã»ãããã¯ãŒã¯ïŒGANïŒãªã©ãæ°ããã¢ã«ãŽãªãºã ã®éçºãé²è¡äžã§ãããããã®æè¡ã¯ããã人éã«è¿ãåŠç¿èœåãåµé æ§ããã¥ãŒã©ã«ãããã¯ãŒã¯ã«äžããããšãæåŸ ãããŠããŸãã
æåŸã«ããšããžã³ã³ãã¥ãŒãã£ã³ã°ã®çºå±ã¯ãããŒã¿ãã©ã€ãã·ãŒã®ä¿è·ãšäœé 延ã®ãªã¢ã«ã¿ã€ã åŠçã®äž¡æ¹ãå®çŸããããã®ããŒãšãªããŸãã次äžä»£ãã¥ãŒã©ã«ãããã¯ãŒã¯æè¡ã®å±æã¯ããããã®æ°ããæè¡ãšã¢ã«ãŽãªãºã ã®çµ±åã«ãã£ãŠãããã¹ããŒããªããã€ã¹ãšã·ã¹ãã ã®å®çŸãç®æããŠããŸãã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ãåŠã¶äžã§ã®ãªãœãŒã¹
ãªã³ã©ã€ã³ã³ãŒã¹ãšãã¥ãŒããªã¢ã«
ãã¥ãŒã©ã«ãããã¯ãŒã¯ãç解ããæ·±ãåŠã¶ããã®æåã®ã¹ãããã¯ã質ã®é«ãææãèŠã€ããããšã§ãã幞ããªããšã«ãå€ãã®ãªã³ã©ã€ã³ãã©ãããã©ãŒã ãç¡æãŸãã¯ææã®ã³ãŒã¹ãæäŸããŠããããããã¯åå¿è ããäžçŽè ãŸã§ãå¹ åºãã¬ãã«ã®ããã°ã©ããç 究è ã察象ã«ããŠããŸãã
代衚çãªãã©ãããã©ãŒã ã«ã¯CourseraãedXãUdacityãããããããããã®ãŠã§ããµã€ãã§ã¯ã¹ã¿ã³ãã©ãŒã倧åŠãMITãªã©ãäžçææ°ã®å€§åŠãæäŸãã質ã®é«ãã³ãŒã¹ãåè¬ã§ããŸãããããã®ã³ãŒã¹ã§ã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®åºç€ããã深局åŠç¿ãç³èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ãååž°åãã¥ãŒã©ã«ãããã¯ãŒã¯ãªã©ãææ°ã®æè¡ã«è³ããŸã§ãæ±ã£ãŠããŸãã
ãŸããYouTubeãGitHubã«ã¯ãæ§ã ãªãã¥ãŒããªã¢ã«ããã¢ãå ¬éãããŠãããå®éã®ãããžã§ã¯ããéããŠåŠç¿ãæ·±ãããæ¹ã«ã¯ç¹ã«åœ¹ç«ã¡ãŸãããããã®ãªãœãŒã¹ã掻çšããããšã§ãçè«ã ãã§ãªããå®è·µçãªã¹ãã«ã身ã«ã€ããããšãã§ããŸãã
éèŠãªç 究è«æãšæžç±
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®çè«çãªèæ¯ãšé²åãæ·±ãç解ããã«ã¯ãç 究è«æãšå°éæžç±ãèªãããšãäžå¯æ¬ ã§ããç¹ã«éèŠãªåºç€ãšãªãè«æã«ã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®åŠç¿ã¢ã«ãŽãªãºã ãã深局åŠç¿ã®å¯èœæ§ãåºããç 究ãå«ãŸããŸãã
Google ScholarãarXivã®ãããªãªã³ã©ã€ã³ããŒã¿ããŒã¹ã«ã¢ã¯ã»ã¹ããããšã§ãæ©æ¢°åŠç¿ããã¥ãŒã©ã«ãããã¯ãŒã¯ã«é¢ããææ°ã®ç 究ææãæ€çŽ¢ããé²èŠ§ããããšãã§ããŸããããããã©ã®ç 究ãæãéèŠã§ããããèŠæ¥µããã«ã¯ãããçšåºŠã®å°éç¥èãšçµéšãå¿ èŠã§ãã
å ¥éæžãšããŠã¯ãIan GoodfellowãYoshua BengioãAaron Courvilleã«ãã”Deep Learning”ãåºãæšèŠãããŠããŸãããã®æžç±ã¯ãçè«ããå®è·µãŸã§ãç¶²çŸ ããŠããããã¥ãŒã©ã«ãããã¯ãŒã¯ãåŠã¶ããã®åºç€ç¥èãåºããã®ã«é©ããŠããŸãã
ãœãããŠã§ã¢ãšã©ã€ãã©ãª
çŸä»£ã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ãå¹çè¯ãèšç·Žããå®è£ ããã«ã¯ã匷åãªãœãããŠã§ã¢ãšã©ã€ãã©ãªãå¿ èŠã§ããTensorFlowãPyTorchã¯ããã®åéã§åºã䜿çšãããŠãããªãŒãã³ãœãŒã¹ã®ã©ã€ãã©ãªã§ã䜿ãããããšæè»æ§ãå ŒãåããŠããŸãã
TensorFlowã¯Googleã«ãã£ãŠéçºããã倧èŠæš¡ãªãããã¯ãŒã¯ãå¹ççã«ãã¬ãŒãã³ã°ããããšãã§ããäžæ¹ã§ãPyTorchã¯Facebookã«ãã£ãŠæäŸãããŠãããåçãªèšç®ã°ã©ãããµããŒãããç 究è ã«äººæ°ããããŸãããããã®ã©ã€ãã©ãªã«ã¯åºç¯ãªããã¥ã¡ã³ããšãã¥ãŒããªã¢ã«ããããåŠç¿ã®æåŒããšããŠæ©èœããŸãã
å ããŠãKerasã®ãããªé«æ°Žæºã®APIã䜿çšããããšã§ãããè€éãªãããã¯ãŒã¯ãç°¡åã«æ§ç¯ããããšãå¯èœã§ããKerasã¯TensorFlowã®äžã«æ§ç¯ãããŠãããããçŽæçãªAPIãæäŸããŠããŸãã
ã³ãã¥ããã£ãšãã©ãŒã©ã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ãåŠã³ãç 究ããéçšã§çåãåé¡ã«çŽé¢ããæãã³ãã¥ããã£ããã©ãŒã©ã ã¯éåžžã«æå¹ãªãªãœãŒã¹ãšãªããŸããStack OverflowãRedditã®ãµãã¬ãã£ããããããŠç¹å®ã®ãã¬ãŒã ã¯ãŒã¯ã®å ¬åŒãã©ãŒã©ã ã¯ã質åãæçš¿ããå°é家ãåãéãæ©ã仲éããã¢ããã€ã¹ãåŸãæé©ãªå Žã§ãã
ãããã®ãã©ãããã©ãŒã ã¯ãæ©æ¢°åŠç¿ããã¥ãŒã©ã«ãããã¯ãŒã¯ã«é¢ããåºç¯ãªè©±é¡ãã«ããŒããŠãããææ°ã®ç 究çµæããã¯ããã¯ã®å ±æããããžã§ã¯ãã®ãã£ãŒãããã¯ã®ããã«ãå©çšãããŠããŸãããŸããããã«ãœã³ããªã³ã©ã€ã³ã³ã³ãã¹ãã«åå ããããšããç¥èãæ·±ãããããã¯ãŒã¯ãåºããè¯ãæ©äŒã§ãã
å ããŠãGitHubã¯ãããžã§ã¯ããã³ãŒãã®å ±æã«æé©ãªãã©ãããã©ãŒã ã§ããããªãŒãã³ãœãŒã¹ãããžã§ã¯ããžã®è²¢ç®ã¯ãå®åçãªæè¡ãšååã®ã¹ãã«ãåäžããã絶奜ã®ãã£ã³ã¹ãšãªããŸããã³ãŒãã¬ãã¥ãŒãåããããšã§ãããæŽç·Žãããã³ãŒãã£ã³ã°æè¡ã身ã«ã€ããããšãã§ããŸãã
ãŸãšã
ããžãã¹ããŒãœã³ã®çæ§ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®å¯èœæ§ãæ倧éã«æŽ»ãããŸãããããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ãè³ã®ç¥çµçŽ°èïŒãã¥ãŒãã³ïŒããã³ããšãªã£ãŠèšèšãããæè¡ã§ããåºæ¬æ§é ãšããŠãå ¥åå±€ãé ãå±€ãåºåå±€ããæããããããéããŠæ å ±ãåŠçãããŸããåŠç¿ããã»ã¹ã«ã¯ãé äŒæãšéäŒæããããŠæ倱é¢æ°ãšæé©åã¢ã«ãŽãªãºã ãäžå¯æ¬ ã§ãããã®æè¡ã¯ãç»åèªèãèªç¶èšèªåŠçïŒNLPïŒãé³å£°èªèãäºæž¬ãªã©ãå¹ åºãå¿çšãå¯èœã§ãããããã解éå¯èœæ§ãããŒã¿ãã©ã€ãã·ãŒãªã©ã解決ãã¹ã課é¡ããããŸããæªæ¥ãžåãããã¥ãŒã©ã«ãããã¯ãŒã¯ã®é²åãšå¿çšç¯å²ã®æ¡å€§ã«æåŸ ãå¯ããããŠããŸããç¶ç¶çãªåŠç¿ãšæ¢ç©¶å¿ãããã®é©æ°çãªæè¡åéã§æåããéµãšãªãã§ãããã