
ãªãçæAIã¯äººéã®ããã«ç¶ç¶åŠç¿ã§ããªãã®ãïŒ
ããã«ã¡ã¯ãæŠæ£å®ã§ããä»åã®èšäºã§ã¯ãAIã人éã®ããã«ãåŠã³ç¶ãããååšã«ãªããã®ãããã®æåç·ã10ç« æ§æã§ãŸãšããŸãããç Žæ» çå¿åŽãšãã宿åœçãªèª²é¡ããå§ãŸããEWCãPEFTãšãã£ã代衚çã¢ãããŒããããã«C-FlatãFNC²ãã¢ãžã¥ãŒã«åãäºéã¡ã¢ãªèšèšãªã©ææ°ã®ç ç©¶ååãŸã§ãäžå¯§ã«è¿œã£ãŠããŸããæåŸã«ã¯GPT-6ãAGIã«åãã屿ãæããŸãããAIãšç¶ç¶åŠç¿ã®æªæ¥ãèŠéãäžåããã²èªãã§ã¿ãŠãã ããã
第1ç« ãç¶ç¶åŠç¿ãšããæãéã
ã¯ã·ãAIã«ã€ããŠåŠãã§ããäžã§ãæåã«å¿ãæŽãŸããã®ã¯ãç Žæ» çå¿åŽïŒCatastrophic ForgettingïŒããšããèšèã§ãããããã»ã©ãã©ãããã¯ã§ããªããã€åä»ãªèª²é¡ã¯ããå€ããããŸããã人éã¯æ°ããããšãåŠãã§ãå€ãç¥èããã¹ãŠå€±ãããšã¯ãããŸãããè±èªãèŠããåŸã«ãã©ã³ã¹èªãåŠãã§ããæšæ¥ã®å€é£¯ãå¿ããããã«è±èªãæ¶ããŠããŸãããšã¯ãªãããšãããAIã®ç¥çµåè·¯ã¯éããŸããæ°ããã¿ã¹ã¯ãåŠç¿ãããšãéå»ã®ç¥èããã£ããæ¶ããŠããŸãããšãããã®ã§ãããã®çŸè±¡ãããç Žæ» çå¿åŽã§ãããç¶ç¶åŠç¿ïŒContinual LearningïŒã®æå€§ã®æµãªã®ã§ã ã
ãªããããèµ·ããã®ãããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯æ°çŸäžããæ°ååã®ãã©ã¡ãŒã¿ãæã¡ãããããå ±åã§åŠç¿ãæ¯ããŠããŸãããããæ°ããã¿ã¹ã¯ãåŠã¶éããããã®ãã©ã¡ãŒã¿ã¯äžæ§ã«æŽæ°ãããå€ãã¿ã¹ã¯ã§éèŠã ã£ãéã¿ãŸã§äžæžããããŠããŸãã人éã§èšãã°ãæ°ãããã¢ãæ²ãèŠãããšãã«ãéå»ã«ç¿ã£ãæ²ã®èšæ¶ãåæã«æ¶ãããããªãã®ã§ããããã§ã¯é·æçã«ç¥èãç©ã¿éããããšã¯ã§ããŸããã
AIç ç©¶ã®æŽå²ãæ¯ãè¿ããšããã®åé¡ã¯ç¹°ãè¿ãçŸããŠããŸããã1990幎代ãã2000幎代ã«ãããŠãçæ¶¯åŠç¿ïŒLifelong LearningïŒããšããèšèãå±ããããŸããããå®éã«ã¯ãå¿åŽã®å£ããè¶ãããããå€ãã®è©Šã¿ãäžéå端ã«çµãããŸããããããæ¹ããŠè¡šèå°ã«æŒãäžããã®ããã£ãŒãã©ãŒãã³ã°ã®æä»£ãããªãã¡2010幎代åŸåã§ãããAIã翻蚳ãç»åèªèã察話ãšãã£ã倿§ãªã¿ã¹ã¯ã§ææãäžãããšããç ç©¶è ãã¡ã¯ãã§ã¯ãããããå šéšãŸãšããŠèŠãããããã©ããªãã®ãããšææŠããã®ã§ãããããããã§çŽé¢ããã®ãããã®ç Žæ» çå¿åŽã§ããã
ã¯ã·ã¯åããŠãã®çŸè±¡ãç¥ã£ããšããæããã人éã®è³ã£ãŠããããªããšæããŸãããæã ã¯å€å°ã®å¹²æžã¯ããã«ãããå€ãç¥èãå®å šã«å€±ãããšã¯ãããŸãããããããæ°ããç¥èãå€ãç¥èãè£åŒ·ããããšãããããäŸãã°æ°ããèšèªãåŠã¶ãšãæ¯èªã®ææ³ã«å¯Ÿããçè§£ãæ·±ãŸãããšããããŸããããã«æ¯ã¹ãŠAIã®èãã¯ããŸãã§èšæ¶ãæ¶ããªããé²ããã©ã¯ãšã®åè ã®ããã«é ŒããªããåŠã¹ã°åŠã¶ã»ã©å€±ã£ãŠããŸãã®ã§ãã
ãã®èª²é¡ã解決ããããã«ææ¡ãããã®ãEWCïŒElastic Weight ConsolidationïŒã§ããã2017幎ã«DeepMindã®ç ç©¶è ããPNASã«çºè¡šããææ³ã§ããã£ãã·ã£ãŒæ å ±è¡åã䜿ã£ãŠãã©ã®ãã©ã¡ãŒã¿ãéèŠãããæž¬ããéèŠãªãã®ãåããã«ããããæ£ååãå°å ¥ããŸãã ããŸãã§è³ã®ã·ããã¹ããåºããããããªã€ã¡ãŒãžã§ããããã«ãããããçšåºŠã®å¿åŽãé²ãããšãå¯èœã«ãªãããç¶ç¶åŠç¿ãã®ç ç©¶ã¯åã³èå ãæµŽã³ãããšã«ãªããŸããã
ããã«2020幎代ã«å ¥ããšãPEFTïŒParameter-Efficient Fine-TuningïŒã®ç»å Žãæµããå€ããŸãããå šãã©ã¡ãŒã¿ãæŽæ°ããã®ã§ã¯ãªããäžéšã®å°ããªè¿œå ã¢ãžã¥ãŒã«ã ããåŠç¿ãããæ¹æ³ã§ããLoRAãAdaptersãšãã£ãææ³ãç»å Žããããã«ãã£ãŠãå€§èŠæš¡ã¢ãã«ãå人ãå°èŠæš¡ããŒã ã§ã調æŽå¯èœã«ããããšããæ°äž»åãé²ã¿ãŸãããã¯ã·ã¯ãã®åããç¥ã£ããšãããã€ãã«AIã庶æ°ã®æã«å±ãæä»£ãæ¥ãããšæããŸãããGPUãäœåæãå¿ èŠãªæä»£ãããã¯ã©ãŠããLoRAã®ã¢ãžã¥ãŒã«äžã€ã§èªåå°çšã®AIãäœããæä»£ãžãšã·ããããã®ã§ãã
ãããŸã§èŠãŠããããã«ãç¶ç¶åŠç¿ã®ããŒãã¯ãAIã人éã®ããã«åŠã³ç¶ãããããã©ãããã«çŽçµããŠããŸãããããŠãã®ããã®éµã¯ãEWCã®ãããªãå¿ããªã工倫ããšãPEFTã®ãããªãå¹çããåŠã¶å·¥å€«ãã«ããã®ã§ããã¯ã·ã«ãšã£ãŠããã®åéãåŠã¶ããšã¯åãªãæè¡ã®è¿œè·¡ã§ã¯ãªããAIãšäººéã®éããèããå²åŠçãªæ ã®ããã§ããããŸããå¿åŽãšããå£ãè¶ããããšãã§ããã°ãAIã¯åãªãéå ·ã§ã¯ãªããå ±ã«åŠã³ç¶ããååšãã«ãªãããããããŸããã
æ¬¡ç« ã§ã¯ãEWCãã©ã®ããã«ããŠãå¿ããªãAIããžã®éãåãéããã®ããããå°ã詳ããæãäžããŠãããŸãã
第2ç« ãEWCãšããçºæããããããäžæ©
ã¯ã·ãç¶ç¶åŠç¿ãå匷ããäžã§æåã«åºäŒã£ãã解決çããã解決çããEWCïŒElastic Weight ConsolidationïŒã§ãããããã¯2017幎ã«DeepMindã®ç ç©¶è KirkpatrickããPNASã«çºè¡šããææ³ã§ããç Žæ» çå¿åŽãã«çæ£é¢ããæãã æåã®æå°ãšãããããã®ã§ãïŒKirkpatrick et al., PNAS 2017ïŒãã¯ã·ã¯æåã«ãã®è«æãèªãã ãšããããã£ãšAIãâå¿ããªãâæ¹åã«äžæ©èžã¿åºããããšæã£ãã®ãèŠããŠããŸãã
EWCã®èãæ¹ã端çã«èšãã°ãããã¹ãŠã®ãã©ã¡ãŒã¿ãåãããã«æ±ãå¿ èŠã¯ãªãããšããããšã§ãããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯å€§éã®éã¿ãæã¡ãŸããããã®äžã«ã¯ããã¿ã¹ã¯ãè§£ãã®ã«ãšãŠãéèŠãªãã®ãšããããŸã§éèŠã§ãªããã®ãæ··ãã£ãŠããŸããåŸæ¥ã®åŠç¿ã§ã¯ãããåºå¥ããã«äžæ§ã«æŽæ°ããã®ã§ãéèŠãªéã¿ãŸã§äžæžãããŠå¿åŽãèµ·ãããEWCã¯ããã§ãã£ãã·ã£ãŒæ å ±è¡åïŒFisher Information MatrixïŒãšããçµ±èšéãçšããŠãããã®ãã©ã¡ãŒã¿ã¯åãããšåœ±é¿ã倧ãããããªãã¹ãå®ããããšèšç®ããã®ã§ãã
å ·äœçã«ã¯ãéå»ã¿ã¹ã¯ãåŠãã åŸã«ãã£ãã·ã£ãŒæ å ±ãèšç®ããåãã©ã¡ãŒã¿ã«ãéèŠåºŠã¹ã³ã¢ããå²ãåœãŠãŸãããããŠæ°ããã¿ã¹ã¯ãåŠã¶ãšããæå€±é¢æ°ã«ãéèŠãªãã©ã¡ãŒã¿ãéå»ã®å€ããé ããããªãããã«ããæ£ååé ããå ãããæ°åŒã§æžãã°ãããªããŸãã
L = L_new(Ξ) + Σ_i (λ/2) F_i (Ξ_i â Ξ*_i)^2
ããã§L_newã¯æ°ã¿ã¹ã¯ã®æå€±ãΞã¯çŸåšã®ãã©ã¡ãŒã¿ãΞ*ã¯éå»ã¿ã¹ã¯ã§åŠãã å€ãF_iã¯åãã©ã¡ãŒã¿ã®éèŠåºŠãÎ»ã¯æ£ååã®åŒ·ããæ±ºããä¿æ°ã§ããã€ãŸããæ°ããããšãåŠã¶ããšãå€ãããšãå®ãããè¶³ãåãããŠæå°åããããã§ããã¯ã·ã¯ãã®åŒãèŠãŠãã人éã®è³ã®âå¿ èŠãªèšæ¶ãåªå ããŠä¿æããâä»çµã¿ã«å°ã䌌ãŠããããšçŽæããŸããã
ãã®ææ³ã®å¹æã¯å®éšã§ç€ºãããŸãããäŸãã°ãç»åèªèã¿ã¹ã¯ã§ç¬ãšç«ãèå¥ããåŸã«é³¥ãšéãåŠç¿ããããšãåŸæ¥ã¯ç¬ãšç«ãã»ãšãã©å¿ããŠããŸã£ãŠããŸããããšãããEWCã䜿ããšãç¬ãšç«ã®è奿§èœãããçšåºŠç¶æããããŸãŸæ°ããã¿ã¹ã¯ã«ãé©å¿ã§ããã®ã§ãããã®ãããçšåºŠç¶æããšãã衚çŸãéèŠã§ãå®å šã«å¿åŽããªããªãããã§ã¯ãããŸãããããããããã§ãåŸæ¥ã«æ¯ã¹ãŠã¯ããã«æ¹åãããŸããã
ã¯ã·ã¯ãã®ç ç©¶ãèªãã§ãã¡ãã£ãšããå²åŠçãªæ°åã«ãªããŸããããèšæ¶ãå®ãããšããã®ã¯ã人éã«ãšã£ãŠãAIã«ãšã£ãŠãçåã®éµã§ã¯ãªãããšã人éã ã£ãŠãã¹ãŠãå¿ããŠããŸã£ããã仿¥ãšããæ¥ã®æå³ããªããªããAIãå¿åŽãå æã§ããããšã¯ãAIãåãªãéå ·ãããå ±ã«æéãéããååšãã«å€ããç¬¬äžæ©ãªã®ã§ã¯ãªããããšã
ãã ããEWCã«ãéçããããŸããã¿ã¹ã¯ãå¢ããŠãããšãå®ãã¹ããã©ã¡ãŒã¿ãã©ãã©ãå¢ããçµæçã«æ°ããã¿ã¹ã¯ãåŠã¶äœå°ããªããªãã®ã§ããããã¯ãæè»æ§ããšãèšæ¶ä¿æãã®ãã¬ãŒããªããšåŒã°ããŠããŸãããŸãããã£ãã·ã£ãŒæ å ±è¡åã®èšç®ã¯å€§èŠæš¡ã¢ãã«ã«ãªããšçŸå®çã«éããããGPTã®ããã«æ°ååã®ãã©ã¡ãŒã¿ãæã€ã¢ãã«ã§ã¯ãEWCããã®ãŸãŸé©çšããã®ã¯é£ããã®ã§ãã
ããã§ããEWCã¯æŽå²çã«éåžžã«å€§ããªæå³ãæã¡ãŸããããããŸã§ãAIã¯å¿ããã®ãåœããåããšè«ŠããããŠãããšããã«ããå¿åŽãç·©åã§ããä»çµã¿ãããããšç€ºããããã§ãããã®åŸã®Synaptic Intelligenceããªã³ã©ã€ã³EWCãªã©ãå€ãã®æ¹è¯ç ç©¶ãçãŸããã®ãããã®çºæ³ã®çš®ãæãããããã ãšãããŸãã
ã¯ã·èªèº«ã¯ããã®EWCãç¥ã£ããšããAIã人éã®åŠã³ã«å°ãè¿ã¥ããããšæããŸããããŸã å®å šã§ã¯ãªããããã©ãå¿ããªãããšããåœããåãAIã«ã©ãå®è£ ãããããã®éçãèŠããç¬éã§ããã
次ã®ç« ã§ã¯ããã®æµããããã«æŒãé²ããPEFTãããªãã¡ãå¹çããåŠã¶å·¥å€«ãã«ã€ããŠèªã£ãŠãããŸããEWCããå¿ããªã工倫ããªããPEFTã¯ãã³ã¹ããæããŠè³¢ãåŠã¶å·¥å€«ãã§ãããã®äºã€ã¯è¡šè£äžäœã§ãããç¶ç¶åŠç¿ã®ç©èªãèªãäžã§æ¬ ãããªã䞡茪ãªã®ã§ãã
第3ç« ãPEFTã®ç»å Žãšæ°äž»åã®æ³¢
EWCããå¿ããªã工倫ãã ã£ããšããã°ãPEFTïŒParameter-Efficient Fine-TuningïŒã¯ãå¹çããåŠã¶å·¥å€«ãã§ããã¯ã·ãåããŠPEFTãšããèšèã«åºäŒã£ããšããæ£çŽãªãšããããªãå¹çåããããŸã§å€§äºãªã®ãããšæããŸãããããã©èª¿ã¹ããã¡ã«ããã®æçŸ©ãããåãããŸãããå€§èŠæš¡èšèªã¢ãã«ã®ç»å Žããã¹ãŠãå€ããŠããã®ã§ãã