ã°ãªãŽãªãŒã»ãµãããïŒã€ã³ãã³ãïŒ
ç§ã®ååã¯ã°ãªãŽãªãŒã»ãµãããã§ããç§ã¯ã€ã³ãã³ã瀟ã®ãµãŒãã¹ã¹ããŒã·ã§ã³ã§ãã ç§ã¯é·ãéãã¥ãŒã©ã«ãããã¯ãŒã¯ã«æºãã£ãŠãããç¹ã«æ©æ¢°åŠç¿ã¯éè·¯æšèãšæ°åã®ãã¥ãŒã©ã«ãããã¯ãŒã¯èªèè£ çœ®ã®æ§ç¯ã«æºãã£ãŠããŸããã ç§ã¯ãç»åã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®æ§åŒåã«é¢ãããããžã§ã¯ãã«åå ããå€ãã®äŒæ¥ãæ¯æŽããŠããŸãã
ãã€ã³ãã«ãŸã£ããã«è¡ããŸãããã ç§ã®ç®æšã¯ãåºæ¬çãªçšèªãšããã®åéã§äœãèµ·ããŠããã®ããããªãã¯ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®çµã¿ç«ãŠå ãããã³ãã®äœ¿çšæ¹æ³ãç解ããããšã§ãã
ã¬ããŒãã®æŠèŠã¯æ¬¡ã®ãšããã§ãã æåã«ã ãã¥ãŒãã³ãäœã§ãããã«ã€ããŠã®ç°¡åãªçŽ¹ä»ã ãã¥ãŒã©ã«ãããã¯ãŒã¯ ã ãã£ãŒããã¥ãŒã©ã«ãããã¯ãŒã¯ã䜿çšããŠãåãèšèªã§éä¿¡ããŸãã
ããã«ããã®åéã§èµ·ãã£ãŠããéèŠãªåŸåã«ã€ããŠã話ããŸãã ãããããã¥ãŒã©ã«ãããã¯ãŒã¯ã®ã¢ãŒããã¯ãã£ãæãäžãããããã®3ã€ã®äž»ãªã¯ã©ã¹ãèããŠãã ããã ãããæãéèŠãªéšåã«ãªããŸãã
ãã®åŸã2ã€ã®æ¯èŒçé«åºŠãªãããã¯ãæ€èšãããã¥ãŒã©ã«ãããã¯ãŒã¯ãæäœããããã®ãã¬ãŒã ã¯ãŒã¯ãšã©ã€ãã©ãªã®ç°¡åãªæŠèŠã§çµãããŸãã
äŒè°ã§ã¯ãNTechLabã®Natalia Efremovaãå®éã®äºäŸã«ã€ããŠè©±ããŸããã ãã¥ãŒã©ã«ãããã¯ãŒã¯ãå éšã«ã©ã®ããã«é 眮ãããŠããããå éšã§ã©ã®ãããªããªãã¯ãäœãããŠãããã説æããŸãã
ãŸãšã
èŠçŽïŒãã¥ãŒãã³ããã¥ãŒã©ã«ãããã¯ãŒã¯ããã£ãŒããã¥ãŒã©ã«ãããã¯ãŒã¯
ç°¡åãªãªãã€ã³ããŒ
人工ãã¥ãŒãã³ã¯ãçäœãã¥ãŒãã³ã«éåžžã«äŒŒãŠããŸãã
人工ãã¥ãŒãã³ãšã¯äœã§ããïŒ ããã¯å®éã«ã¯åçŽãªé¢æ°ã§ãã 圌女ã«ã¯ã€ã³ãããããããŸãã åå ¥åã«ç¹å®ã®éã¿ãæããŠããããã¹ãŠãåèšããéç·åœ¢é¢æ°ãå®è¡ããçµæãåºåããŸã-ãã¹ãŠãããã¯1ã€ã®ãã¥ãŒãã³ã§ãã
ããžã¹ãã£ãã¯ååž°ïŒSIGMOIDã®éç·åœ¢é¢æ°ãæå³ããïŒã«ç²ŸéããŠããå Žåã1ã€ã®ãã¥ãŒãã³ã¯ããžã¹ãã£ãã¯ååž°ã®å®å šãªé¡äŒŒç©ã§ãããåçŽãªç·åœ¢åé¡åšã§ãã
å®éãå³ã«ç€ºãããŠããåæ²ç·æ£æ¥ïŒTANHïŒãSIGMOIDãRELUãªã©ãããŸããŸãªã¢ã¯ãã£ããŒã·ã§ã³é¢æ°ããããŸãã
çŸå®ã«ã¯ããã¹ãŠãã¯ããã«è€éã§ãã ãã®ãããã¯ã«ã€ããŠã¯è§ŠããŸããã
äžçš®ã®çç©åŠçãã¥ãŒãã³ãšããŠã人工ãã¥ãŒãã³ã®éåžžã«åºæ¬çãªèããè¿°ã¹ãŸããã
人工ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ãåé¡åé¡ãªã©ã®ç¹å®ã®åé¡ã解決ããããã«ããã¥ãŒãã³ããããã¯ãŒã¯ã«åéããæ¹æ³ã§ãã ãã¥ãŒãã³ã¯å±€ç¶ã«éãŸããŸãã å ¥åä¿¡å·ãäŸçµŠãããå ¥åå±€ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®çµæãååŸãããåºåå±€ãããã³ãããã®éã«é ãããå±€ããããŸãã 1ã2ã3ãå€ããããŸãã é ãå±€ãè€æ°ããå Žåããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯æ·±ããšèŠãªããã1ã®å Žåã¯æµ ããšèŠãªãããŸãã
å€çš®å€æ§ãªã¢ãŒããã¯ãã£ãããããã®äž»ãªãã®ãæ€èšããŸãã ãããããããã®å€ããããããšã«çæããŠãã ããã èå³ãããå Žåã¯ã ãªã³ã¯ããã©ã£ãŠãã ãã -ã芧ãã ããã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã«ã€ããŠè°è«ããããã«ç¥ã£ãŠããå¿ èŠããããã1ã€ã®æçšãªããšã 1ã€ã®ãã¥ãŒãã³ãã©ã®ããã«æ©èœãããã«ã€ããŠã¯æ¢ã«èª¬æããŸãããåå ¥åãéã¿ãä¿æ°ãåèšãéç·åœ¢æ§ãä¹ç®ããæ¹æ³ã§ãã ããã¯ãããšãã°ããã¥ãŒãã³ã®çç£ã¢ãŒããã€ãŸããæšè«ããã§ã«èšç·Žããã圢ã§ã©ã®ããã«æ©èœãããã§ãã
å®å šã«ç°ãªãã¿ã¹ã¯ããããŸã-ãã¥ãŒãã³ãèšç·Žããããšã ãã¬ãŒãã³ã°ã¯ããããã®é©åãªéã¿ãèŠã€ããããšã§ãã ãã¬ãŒãã³ã°ã¯ããã¥ãŒãã³ã®åºåã§çããã©ãããã¹ããããããŠäœãå€æããããç¥ããšããã®éãããšã©ãŒãèªèãããšããåçŽãªèãã«åºã¥ããŠããŸãã ãã®ãšã©ãŒã¯ããã¥ãŒãã³ã®ãã¹ãŠã®å ¥åã«éãè¿ãããã©ã®å ¥åããã®ãšã©ãŒã«å€§ããªåœ±é¿ãäžããããç解ã§ããŸãããããã£ãŠããã®å ¥åã®éã¿ã調æŽããŠããšã©ãŒãæžãããŸãã
ããããããã¯ãããã²ãŒã·ã§ã³ã¢ã«ãŽãªãºã ã§ããããã¯ãããã²ãŒã·ã§ã³ã®äž»ãªã¢ã€ãã¢ã§ãã ãã®ããã»ã¹ã¯ããããã¯ãŒã¯å šäœã§ããã¥ãŒãã³ããšã«é§åããŠãéã¿ãå€æŽããæ¹æ³ãèŠã€ããããšãã§ããŸãã ãã®ããã«ã¯ãããªããã£ããåãå¿ èŠããããŸãããååãšããŠãæè¿ã§ã¯ããã¯å¿ èŠãããŸããã ãã¥ãŒã©ã«ãããã¯ãŒã¯ãæäœããããã®ãã¹ãŠã®ããã±ãŒãžã¯ãèªåçã«åºå¥ãããŸãã 2幎åã«ããªãããŒãªã¬ã€ã€ãŒã®è€éãªæŽŸçç©ãæåã§èšè¿°ããå¿ èŠããã£ãå Žåãããã±ãŒãžã¯ãããèªåã§è¡ããŸãã
èŠçŽïŒéèŠãªãã¬ã³ã
ã¢ãã«ã®å質ãšè€éãã§äœãèµ·ãã£ãŠããã®ã
第äžã«ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®ç²ŸåºŠãåäžããéåžžã«åäžããŠããŸãã ãã¥ãŒã©ã«ãããã¯ãŒã¯ãããé åã«æ¥ãŠãå€å žçãªã¢ã«ãŽãªãºã ã«å®å šã«åã£ãŠä»£ããå Žåããã§ã«ããã€ãã®äŸããããŸãã ããã¯ãã§ã«ç»ååŠçãšé³å£°èªèã§çºçããŠãããããŸããŸãªåéã§çºçããŸãã ã€ãŸãããšã©ãŒãå€§å¹ ã«æžãããã¥ãŒã©ã«ãããã¯ãŒã¯ã衚瀺ãããŸãã
å³ã§ã¯ããã£ãŒãã©ãŒãã³ã°ã玫è²ã§åŒ·èª¿è¡šç€ºãããéè²ãå€å žçãªã³ã³ãã¥ãŒã¿ãŒããžã§ã³ã¢ã«ãŽãªãºã ã§ãã ãã£ãŒãã©ãŒãã³ã°ã衚瀺ããããšã©ãŒãæžå°ããããã«æžå°ãç¶ããŠããããšãããããŸãã ãã®ããããã£ãŒãã©ãŒãã³ã°ã¯ãæ¡ä»¶ä»ãã§ãã¹ãŠã®å€å žçãªã¢ã«ãŽãªãºã ãå®å šã«çœ®ãæããŸãã
ãã1ã€ã®éèŠãªãã€ã«ã¹ããŒã³ã¯ãå質ã®é¢ã§äººãè¿œãè¶ãå§ããããšã§ãã ImageNetã§ã¯ãããã2015幎ã«åããŠã§ããã ãããå®éã«ã¯ã人éãããå質ã«åªãããã¥ãŒã©ã«ãããã¯ãŒã¯ã·ã¹ãã ã以åã«ç»å ŽããŸããã æåã«ææžåãããåå¥ã®ã±ãŒã¹ã¯2011幎ã§ããã€ãã®éè·¯æšèãèªèãã人ããã2ååªããã·ã¹ãã ãæ§ç¯ãããŸããã
2çªç®ã®éèŠãªåŸå-ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®è€éããå¢å€§ããŠããŸãã æ·±ãã«é¢ããŠã¯ãæ·±ãã倧ãããªããŸãã ImageNetã§2012幎ã®åè ïŒAlexNetãããã¯ãŒã¯ïŒã10å±€æªæºã ã£ãå Žåã2014幎ã«ã¯2015幎ã«ã¯ãã§ã«20ãè¶ ãã150æªæºã§ãããä»å¹Žã¯æ¢ã«200ãè¶ ããŠããããã§ããããã«ãããŸãã
http://cs.unc.edu/~wliu/papers/GoogLeNet.pdf
æ·±ãã®å¢å€§ã«å ããŠãã¢ãŒããã¯ãã£èªäœãè€éåããŠããŸãã ã¬ã€ã€ãŒã1ã€ãã€çµåããã®ã§ã¯ãªããåå²ããããã¯ãæ§é ãçŸãå§ããŸãã äžè¬ã«ãã¢ãŒããã¯ãã£ã®è€éããå¢å€§ããŠããŸãã
https://culurciello.github.io/tech/2016/06/04/nets.html
ããã¯ãããŸããŸãªãã¥ãŒã©ã«ãããã¯ãŒã¯ã®ç²ŸåºŠã®ã°ã©ãã§ãã ããã¯ãå®è¡ã«ãããæéããã®ãããã¯ãŒã¯ã®èª€èšç®ãã€ãŸãããçš®ã®èšç®è² è·ã瀺ããŸãã åã®ãµã€ãºã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ã«ãã£ãŠèšè¿°ããããã©ã¡ãŒã¿ãŒã®æ°ã§ãã 2012幎以éã®ãããã¯ãŒã¯ã®åè ã§ããåŸæ¥ã®ãããã¯ãŒã¯AlexNetãæ¯èŒããã®ã¯èå³æ·±ãããšã§ãã 粟床ã¯åäžããŸãããéåžžå«ãŸãããã©ã¡ãŒã¿ãŒã¯å°ãªããªããŸãã ããã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ãéåžžã«æŽç·ŽãããŠãããšããéèŠãªåŸåã§ããããŸãã ã€ãŸããã¢ãŒããã¯ãã£ãå€æŽãããŠãããããã¬ã€ã€ãŒã®æ°ã150ã§ãã£ãŠãããã©ã¡ãŒã¿ãŒã®åèšæ°ã¯2012幎ã®6-7ã¬ã€ã€ãŒãããã¯ãŒã¯ãããå°ãªããªããŸãã ã¢ãŒããã¯ãã£ã¯éåžžã«èå³æ·±ãæ¹æ³ã§äœããã®åœ¢ã§è€éã«ãªã£ãŠããŸãã
å¥ã®åŸåã¯ãããŒã¿éã®å¢å ã§ãã 1998幎ã«ç³ã¿èŸŒã¿ãã¬ãŒãã³ã°çš
ææžãã®ãã§ãã¯ãèªèãããã¥ãŒã©ã«ãããã¯ãŒã¯ã10 7ãã¯ã»ã«ã2012幎ã«äœ¿çšãããŸããïŒIMAGENETïŒ-10 14 ã
14幎ã§7泚æ-ããã¯å€§ããªéãã§ããã倧ããªå€åã§ãïŒ
åæã«ãããã»ããµã§ã®ééæ°ãå¢ããèšç®èœåãåäžããŠããŸã-ã ãŒã¢ã®æ³åãæå¹ã§ãã ãã®14幎éã§ãããã»ããµã¯æ¡ä»¶ä»ãã§1000åé«éã«ãªããŸããã ããã¯ãçŸåšãã£ãŒãã©ãŒãã³ã°ã®åéãæ¯é ããŠããGPUã«ãã£ãŠç€ºãããŠããŸãã ã»ãšãã©ãã¹ãŠãã°ã©ãã£ãã¯ã¢ã¯ã»ã©ã¬ãŒã¿ã§ã«ãŠã³ããããŸãã
NVIDIAã¯ãã²ãŒã äŒç€Ÿãã人工ç¥èœã®äŒç€Ÿã«å€èº«ããŸããã ãã®åºå±è ã¯ãIntelã®åºå±è ã倧ããåŒãé¢ããŸããã
ããã¯ãããããšã³ãã®ã°ã©ãã£ãã¯ã«ãŒãã4.5 TFLOPSã ã£ã2013幎ã®åçã§ãã çŸåšãæ°ããTITAN Xã¯ãã§ã«11 TFLOPSã§ãã äžè¬çã«ãåºå±è ã¯ç¶ç¶ããŸãïŒ
å®éãè¿ãå°æ¥ã«FPGAãç»å ŽããGPUãéšåçã«çœ®ãæããããããšãäºæ³ãããŸãããŸãããã¥ãŒãã¢ãŒãã£ãã¯ããã»ããµã§ããæéãããŠã°ç»å Žããå¯èœæ§ããããŸãã ããã远跡ããŠãã ãã-èå³æ·±ãããšããããããããŸãã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®ã¢ãŒããã¯ãã£ã çŽæ¥ååžãã¥ãŒã©ã«ãããã¯ãŒã¯
å®å šã«æ¥ç¶ããããã£ãŒããã©ã¯ãŒããã¥ãŒã©ã«ãããã¯ãŒã¯ãFNN
æåã®å€å žçãªã¢ãŒããã¯ãã£ã¯ãçŽæ¥ãªã³ã¯ã®å®å šã«æ¥ç¶ããããã¥ãŒã©ã«ãããã¯ãŒã¯ããŸãã¯å®å šã«æ¥ç¶ããããã£ãŒããã©ã¯ãŒããã¥ãŒã©ã«ãããã¯ãŒã¯ïŒFNNïŒã§ãã
å€å±€ããŒã»ãããã³ã¯éåžžããã¥ãŒã©ã«ãããã¯ãŒã¯ã®å€å žã§ãã ããªããèŠããã¥ãŒã©ã«ãããã¯ãŒã¯ã®ç»åãããã¯ããã§ã-å€å±€å®å šæ¥ç¶ãããã¯ãŒã¯ã å®å šã«æ¥ç¶-ããã¯ãåãã¥ãŒãã³ãåã®å±€ã®ãã¹ãŠã®ãã¥ãŒãã³ã«æ¥ç¶ãããŠããããšãæå³ããŸãã è¯å¥œãªãããã¯ãŒã¯ãæ©èœããåé¡ã«é©ããŠããŸããå€ãã®åé¡åé¡ãæ£åžžã«è§£æ±ºãããŸãã
ãã ãã圌女ã«ã¯2ã€ã®åé¡ããããŸãã
- å€ãã®ãªãã·ã§ã³
ããšãã°ã100 * 100 psã®ç»åãåŠçããå¿ èŠããã3ã€ã®é ãå±€ã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ã䜿çšããå Žåãããã¯å ¥åã«10,000 psãããã3ã€ã®å±€ã§èµ·åããããšãæå³ããŸãã äžè¬ã«ããã¹ãŠã®ãã©ã¡ãŒã¿ãŒã«æ£çŽã«èšããšããã®ãããªãããã¯ãŒã¯ã«ã¯çŽ100äžã®ãã©ã¡ãŒã¿ãŒããããŸãã ããã¯å®éã«ã¯ãããããããŸãã 100äžåã®ãã©ã¡ãŒã¿ãŒã䜿çšããŠãã¥ãŒã©ã«ãããã¯ãŒã¯ããã¬ãŒãã³ã°ããã«ã¯ãåžžã«ååšãããšã¯éããªãå€ãã®ãã¬ãŒãã³ã°ãµã³ãã«ãå¿ èŠã§ãã å®éãä»ã§ã¯äŸããããŸãããäŸã¯ãããŸããã§ããããããã£ãŠãç¹ã«ããããã¯ãŒã¯ã¯é©åã«ãã¬ãŒãã³ã°ã§ããŸããã§ããã
ããã«ãå€ãã®ãã©ã¡ãŒã¿ãæã€ãããã¯ãŒã¯ã«ã¯ãåãã¬ãŒãã³ã°ããåŸåããããŸãã çŸå®ã«ã¯ååšããªããã®ãã€ãŸããã€ãºããŒã¿ã»ããã«éã蟌ããããå¯èœæ§ããããŸãã ããšãæçµçã«ãããã¯ãŒã¯ãäŸãèŠããŠããŠãããããèŠããªãã£ãå Žåã¯ããããã¯ãŒã¯ãéåžžã©ãã䜿çšããããšã¯ã§ããŸããã
ããã«ãå¥ã®åé¡ããããŸãïŒ
- ãã§ãŒãžã³ã°ã°ã©ããŒã·ã§ã³
åºåããã®ãšã©ãŒãå ¥åã«éä¿¡ããããã¹ãŠã®éã¿ã«åé ããããããã¯ãŒã¯ãä»ããŠããã«éä¿¡ããããšãã®ãããã¯ãããã²ãŒã·ã§ã³ã«é¢ãã話ãèŠããŠããŸããïŒ ããã«ããããã®å°é¢æ°ãã€ãŸãåŸé ïŒèª€å·®ã®å°é¢æ°ïŒã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ãä»ããŠæ»ãããŸãã ãã¥ãŒã©ã«ãããã¯ãŒã¯ã«å€ãã®å±€ãããå Žåãéåžžã«éåžžã«å°ããªéšåãæåŸã®ãã®åŸé ããæ®ãããšããããŸãã ãã®å Žåããã®åŸé ã¯å®éã«ã¯ãæ»ãã§ããããããå ¥åã®éã¿ãå€æŽããããšã¯ã»ãšãã©äžå¯èœã§ãã
ãããåé¡ã§ãããã®çç±ã¯ããã£ãŒããã¥ãŒã©ã«ãããã¯ãŒã¯ãèšç·Žããã®ãé£ããããã§ãã ç¹ã«ãªã«ã¬ã³ããããã¯ãŒã¯ã«ã€ããŠã¯ããã®ãããã¯ã«æ»ããŸãã
FNNãããã¯ãŒã¯ã«ã¯ããŸããŸãªããªãšãŒã·ã§ã³ããããŸãã ããšãã°ãAuto Encoderã®éåžžã«èå³æ·±ãããªãšãŒã·ã§ã³ã§ãã ããã¯ãããããããã«ããã¯ãçãäžã«ããçŽæ¥é ä¿¡ãããã¯ãŒã¯ã§ãã ããã¯éåžžã«å°ããå±€ã§ãããšãã°10åã®ãã¥ãŒãã³ã®ã¿ã§ãã
ãã®ãããªãã¥ãŒã©ã«ãããã¯ãŒã¯ã®å©ç¹ã¯äœã§ããïŒ
ãã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®ç®çã¯ãäœããã®å ¥åãååŸãããããä»ããŠãããé§åããåºåã§åãå ¥åãçæããããšã§ããã€ãŸããäžèŽããŸãã ãã€ã³ãã¯äœã§ããïŒ å ¥åãåãåããããèªäœãé§åãããŸã£ããåãåºåãçæãããããªãããã¯ãŒã¯ããã¬ãŒãã³ã°ã§ããå Žåãããã¯ããã®10åã®ãã¥ãŒãã³ããã®å ¥åãèšè¿°ããã®ã«ååã§ããããšãæå³ããŸãã ã€ãŸããã¹ããŒã¹ãå€§å¹ ã«åæžããããŒã¿éãåæžããå ¥åããŒã¿ã10åã®ãã¯ãã«ãšããæ°ããæ¡ä»¶ã§çµæžçã«ãšã³ã³ãŒãã§ããŸãã
ããã¯äŸ¿å©ã§åäœããŸãã ãã®ãããªãããã¯ãŒã¯ã¯ãããšãã°ãã¿ã¹ã¯ã®èŠæš¡ãçž®å°ãããã䜿çšã§ããèå³æ·±ãæ©èœãèŠã€ãããããã®ã«åœ¹ç«ã¡ãŸãã
èå³æ·±ãRBMã¢ãã«ããŸã ãããŸãã ç§ã¯FNNã®ããªãšãŒã·ã§ã³ã§ãããæžããŸããããå®éã¯ããã§ã¯ãããŸããã 第äžã«ãããã¯æ·±ããªãã第äºã«ããã£ãŒããã©ã¯ãŒãã§ã¯ãããŸããã ãã ããå€ãã®å ŽåãFNNãããã¯ãŒã¯ã«é¢é£ä»ããããŠããŸãã
ããã¯äœã§ãã
ããã¯æµ ãã¢ãã«ïŒã¹ã©ã€ãã®é ã«æãããŠããŸãïŒã§ãå ¥åããããäœããã®é ãã¬ã€ã€ãŒããããŸãã å ¥åä¿¡å·ãäžãããã®å ¥åãçæããããã«é ãå±€ãèšç·ŽããããšããŸãã
ããã¯çæã¢ãã«ã§ãã 圌女ãèšç·Žããå ŽåãåŸã§å ¥åä¿¡å·ã®ã¢ããã°ãçæã§ããŸããããããã«ç°ãªããŸãã ããã¯ç¢ºççã§ããã€ãŸããæ¯åå°ããã€ç°ãªããã®ãçæããŸãã ããšãã°ãææžãã®ãŠããããçæããããã«ãã®ãããªã¢ãã«ããã¬ãŒãã³ã°ããå Žåãäžå®éã®ãããã«ç°ãªããŠããããçæãããŸãã
RBMã®åªããŠããç¹ã¯ãRBMã䜿çšããŠæ·±ããããã¯ãŒã¯ããã¬ãŒãã³ã°ã§ããããšã§ãã ãã®ãããªçšèªããããŸã-Deep Belief NetworksïŒDBNïŒ-å®éãããã¯æ·±ããããã¯ãŒã¯ã®ãã¬ãŒãã³ã°ã®æ¹æ³ã§ãã ãã®åŸããããã®éã¿ã¯ä¿®æ£ãããŸãã 次ã«ã2çªç®ã®ã¬ã€ã€ãŒãååŸãããå¥ã®RBMãšèŠãªãããåãæ¹æ³ã§ãã¬ãŒãã³ã°ãããŸãã ãªã©ããããã¯ãŒã¯å šäœã§ã 次ã«ããããã®RBMãçµåããã1ã€ã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ã«çµåãããŸãã ãã£ãŒããã¥ãŒã©ã«ãããã¯ãŒã¯ã§ããããšãå€æããŸããã
ããããä»ã§ã¯å€§ããªå©ç¹ããããŸã-åã«ã©ã³ãã ïŒã©ã³ãã ïŒç¶æ ãããã¬ãŒãã³ã°ããå Žåãä»ã§ã¯ã©ã³ãã ã§ã¯ãããŸãã-ãããã¯ãŒã¯ã¯ãåã®ã¬ã€ã€ãŒããããŒã¿ã埩å ãŸãã¯çæããããã«ãã¬ãŒãã³ã°ãããŠããŸãã ã€ãŸãã圌女ã®éã¿ã¯åççã§ãããå®éã«ã¯ãããã¯ãã®ãããªãã¥ãŒã©ã«ãããã¯ãŒã¯ãå®éã«éåžžã«ããèšç·ŽãããŠãããšããäºå®ã«ã€ãªãããŸãã ãã®åŸãããã€ãã®äŸã§å°ãèšç·Žããããšãã§ãããã®ãããªãããã¯ãŒã¯ã®å質ã¯è¯å¥œã§ãã
ããã«ãè¿œå ã®å©ç¹ããããŸãã RBMã䜿çšããå Žåãåºæ¬çã«ã¯ãæåž«ãªãåŠç¿ãšåŒã°ããæªå²ãåœãŠããŒã¿ã§äœæ¥ããŸãã ããªãã¯åçããæã£ãŠããŸãããããªãã¯åœŒãã®ã¯ã©ã¹ãç¥ããŸããã FlickrãŸãã¯ä»ã®å ŽæããããŠã³ããŒãããæ°çŸäžãæ°ååã®åçãè¿œãåºãããããã®åçãèšè¿°ãããããã¯ãŒã¯èªäœã«äœããã®æ§é ããããŸãã
ããªãã¯ãŸã ãããäœã§ããããç¥ããŸãããããããã¯ããªããããããããªããæ®ã£ãŠããããŠå°æ°ã®ç°ãªãåçã§åèšç·Žããããšãã§ããåççãªéã¿ã§ãããããŠããã¯ãã§ã«è¯ãã§ãããã ããã¯ã2ã€ã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®çµã¿åããã䜿çšããã¯ãŒã«ãªæ¹æ³ã§ãã
ããã«ããã®ã¹ããŒãªãŒå šäœãå®éã«ã¯ã¬ãŽã«é¢ãããã®ã§ããããšãããããŸãã ã€ãŸããåå¥ã®ãããã¯ãŒã¯ïŒãªã«ã¬ã³ããã¥ãŒã©ã«ãããã¯ãŒã¯ããã®ä»ã®ãããã¯ãŒã¯ïŒãããããããã¯ãã¹ãŠçµåå¯èœãªãããã¯ã§ãã ãããã¯ç°ãªãã¿ã¹ã¯ã§ããŸãçµã¿åããããŸãã
ãããã¯ãå€å žçãªçŽæ¥ååžãã¥ãŒã©ã«ãããã¯ãŒã¯ã§ããã 次ã«ãç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ã«ç§»ããŸãã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¢ãŒããã¯ãã£ïŒç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯
ç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ãCNN
https://research.facebook.com/blog/learning-to-segment/
ç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ã3ã€ã®äž»ãªã¿ã¹ã¯ã解決ããŸãã
- åé¡ã ããªãã¯åçãæåºãããšããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯èšãã ãã§ã-ããªãã¯ç¬ã銬ãäœãã«ã€ããŠã®åçãæã£ãŠããŸãããããŠããã¯ã¯ã©ã¹ãäžããŸãã
- ãã¥ãŒã©ã«ãããã¯ãŒã¯ãç»åå ã«ç¬ãŸãã¯éŠ¬ããããšèšãã ãã§ãªããç»åå ã®å¢çããã¯ã¹ãæ€åºããå Žåã æ€åºã¯ããé«åºŠãªã¿ã¹ã¯ã§ãã
- ã»ã°ã¡ã³ããŒã·ã§ã³ã ç§ã®æèŠã§ã¯ãããã¯æãã¯ãŒã«ãªã¿ã¹ã¯ã§ãã å®éãããã¯ãã¯ã»ã«ããšã®åé¡ã§ãã ããã§ã¯ãç»åå ã®ãã¹ãŠã®ãã¯ã»ã«ã«ã€ããŠèª¬æããŠããŸãããã®ãã¯ã»ã«ã¯ç¬ããã®ãã¯ã»ã«ã¯éŠ¬ããããŠãã®ãã¯ã»ã«ã¯äœãä»ã®ãã®ãæããŸãã å®éãã»ã°ã¡ã³ããŒã·ã§ã³ã®åé¡ã解決ããæ¹æ³ãç¥ã£ãŠããã°ãæ®ãã®2ã€ã®ã¿ã¹ã¯ã¯ãã§ã«èªåçã«äžããããŠããŸãã
ããã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ãšã¯äœã§ããïŒ å®éãç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯éåžžã®ãã£ãŒããã©ã¯ãŒããããã¯ãŒã¯ã§ãããã»ãã®å°ãç¹å¥ãªãã®ã§ãã ã¬ãŽã¯ãã§ã«å§ãŸã£ãŠããŸãã
ç³ã¿èŸŒã¿ãããã¯ãŒã¯ã«ã¯äœããããŸããïŒ åœŒå¥³ãæã£ãŠããïŒ
- ç³ã¿èŸŒã¿å±€-ãããäœã§ããããããã«èª¬æããŸãã
- ç»åãµã€ãºãçž®å°ãããµããµã³ããªã³ã°ãŸãã¯ããŒãªã³ã°å±€ã
- éåžžã®å®å šã«æ¥ç¶ãããå±€ãåãå€å±€ããŒã»ãããã³ããããã¯æåã®2ã€ã®ããªãããŒãªå±€ã«åçŽã«äžããã¶ãäžãã£ãŠããŸãã
ããããã¹ãŠã®ã¬ã€ã€ãŒã«ã€ããŠããå°ã詳ãã説æããŸãã
- ç³ã¿èŸŒã¿å±€ã¯éåžžãå¹³é¢ãŸãã¯ããªã¥ãŒã ã®ã»ãããšããŠæç»ãããŸãã ãã®ãããªå³ã®åãã¬ãŒã³ãŸãã¯ãã®ããªã¥ãŒã ã®åã¹ã©ã€ã¹ã¯ãå®éã«ã¯ãç³ã¿èŸŒã¿æŒç®ãå®è£ ãã1ã€ã®ãã¥ãŒãã³ã§ãã ç¹°ãè¿ãã«ãªããŸãããåŸã§èª¬æããŸãã å®éãããã¯å ã®ç»åãä»ã®ç»åã«å€æãããããªãã¯ã¹ãã£ã«ã¿ãŒã§ãããããã¯äœåºŠãå®è¡ã§ããŸãã
- ãµããµã³ããªã³ã°ã¬ã€ã€ãŒ ïŒãµããµã³ããªã³ã°ãšåŒã³ãŸãïŒã¯åçŽã«ç»åãµã€ãºãçž®å°ããŸãïŒ200 x 200 psã§ãããããµããµã³ããªã³ã°åŸã¯100 * 100 psã«ãªããŸããã å®éãå¹³ååã¯å°ãè€éã§ãã
- éåžžãããŒã»ãããã³ã¯åé¡ã®ããã«å®å šã«æ¥ç¶ãããã¬ã€ã€ãŒã䜿çšããŸãã ãããã«ã€ããŠç¹å¥ãªããšã¯äœããããŸããã
http://intellabs.github.io/RiverTrail/tutorial/
ç³ã¿èŸŒã¿æŒç®ãšã¯äœã§ããïŒ ããã¯çãæããããŸãããå®éã«ã¯éåžžã«ç°¡åãªããšã§ãã Photoshopã§äœæ¥ããGaussian BlurãEmbossãSharpenããã®ä»å€æ°ã®ãã£ã«ã¿ãŒã䜿çšããå Žåããããã¯ãã¹ãŠãããªãã¯ã¹ãã£ã«ã¿ãŒã§ãã ãããªãã¯ã¹ãã£ã«ã¿ãŒ-ããã¯å®éã«ã¯ç³ã¿èŸŒã¿æŒç®ã§ãã
ã©ã®ããã«å®è£ ãããŠããŸããïŒ ãã£ã«ã¿ãŒã³ã¢ãšåŒã°ãããããªãã¯ã¹ããããŸãïŒã«ãŒãã«ãã¯ãã£å ïŒã ãŒããã®å Žåããã¹ãŠã®ãŠãããã«ãªããŸãã ç»åããããŸãã ãã®ãããªãã¯ã¹ã¯ç»åã®äžéšã«éãããã察å¿ããèŠçŽ ãåçŽã«ä¹ç®ãããçµæãè¿œå ãããŠäžå¿ç¹ã«æžã蟌ãŸããŸãã
http://intellabs.github.io/RiverTrail/tutorial/
ãã®ãããããæ確ã«èŠããŸãã ç»åå ¥åãããããã£ã«ã¿ãŒããããŸãã ç»åå šäœã«ãã£ã«ã¿ãŒãå®è¡ãã察å¿ããèŠçŽ ãæ£çŽã«ä¹ç®ããäžå€®ã«è¿œå ãæžã蟌ã¿ãŸãã å®è¡ãå®è¡-æ°ããã€ã¡ãŒãžãäœæããŸããã ããã¯ç³ã¿èŸŒã¿æŒç®ã§ãã
ã€ãŸããå®éã«ã¯ãç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®ç³ã¿èŸŒã¿ã¯ã圌ãåŠç¿ããããªãããŒãªããžã¿ã«ãã£ã«ã¿ãŒïŒBlurãEmbossãªã©ïŒã§ãã
http://cs231n.github.io/convolutional-networks/
å®éãç³ã¿èŸŒã¿å±€ã¯ãã¹ãŠããªã¥ãŒã ã§æ©èœããŸãã ã€ãŸããéåžžã®RGBç»åãæ®åœ±ããŠãããã§ã«3ã€ã®ãã£ãã«ããããŸããããã¯ãå®éã«ã¯å¹³é¢ã§ã¯ãªããæ¡ä»¶ä»ãã§3ã€ã®ç«æ¹äœã§ãã
ãã®å Žåã®ç³ã¿èŸŒã¿ã¯ããããªãã¯ã¹ã§ã¯ãªãããã³ãœã«ãå®éã¯ç«æ¹äœã§è¡šãããŸãã
ãã£ã«ã¿ãŒããããã€ã¡ãŒãžå šäœãå®è¡ãããšãããã«3ã€ã®ã«ã©ãŒã¬ã€ã€ãŒãã¹ãŠã衚瀺ããããã®1ã€ã®ããªã¥ãŒã ã«å¯ŸããŠ1ã€ã®æ°ãããã€ã³ããçæãããŸãã ã€ã¡ãŒãžå šäœãå®è¡ããŸã-1ã€ã®ãã£ãã«ãæ°ããã€ã¡ãŒãžã®1ã€ã®å¹³é¢ãæ§ç¯ããŸãã 5ã€ã®ãã¥ãŒãã³ãããå Žåã5ã€ã®ãã¬ãŒã³ãæ§ç¯ããŠããŸãã
ãããç³ã¿èŸŒã¿å±€ã®ä»çµã¿ã§ãã ç³ã¿èŸŒã¿å±€ããã¬ãŒãã³ã°ããã¿ã¹ã¯ã¯ãéåžžã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ãšåãã¿ã¹ã¯ã§ããã€ãŸããéã¿ãèŠã€ããããšãã€ãŸãããã¥ãŒãã³ã®éã¿ãšå®å šã«çããéåžžã«ç³ã¿èŸŒã¿è¡åãå®éã«èŠã€ããããšã§ãã
ãã®ãããªãã¥ãŒãã³ã¯äœãããŸããïŒ åœŒãã¯å®éã«ãããã€ãã®æ©èœã圌ããèŠããã®å°ããªéšåã®ããã€ãã®ããŒã«ã«ãµã€ã³ãæ¢ãããšãåŠã³ãŸã-ããããã¹ãŠã§ãã ãã®ãããªãã£ã«ã¿ãŒãå®è¡ãããšãç»åå ã®ãããã®æšèã®äœçœ®ã®ããããäœæãããŸãã
次ã«ããã®ãããªãã¬ãŒã³ãå€æ°äœæããããããç»åãšããŠäœ¿çšããŠã次ã®å ¥åã«é©çšããŸãã
http://vaaaaaanquish.hatenablog.com/entry/2015/01/26/060622
ããŒãªã³ã°ã¯ããã«ç°¡åãªæäœã§ãã ããã¯åçŽã«å¹³ååããããæ倧å€ãåãããšã§ãã ãŸãã2 * 2ãªã©ã®ããã€ãã®å°ããªæ£æ¹åœ¢ã§ãæ©èœããŸãã ç»åã«ã¹ãŒããŒã€ã³ããŒãºããããšãã°ããã®2 * 2ããã¯ã¹ããæ倧èŠçŽ ãéžæããŠãåºåã«éä¿¡ããŸãã
ãããã£ãŠãç»åãçž®å°ããŸããããããªãããŒãªå¹³åã§ã¯ãªãããããã«é«åºŠãªãã®ã䜿çšããŠ-æ倧å€ãåããŸããã ããã«ãããããããªå€äœäžå€æ§ãåŸãããŸãã ã€ãŸããããªãã«ãšã£ãŠã¯åé¡ã§ã¯ãããŸããããã®äœçœ®ãŸãã¯å³åŽ2 psã«äœããã®å åãèŠã€ãããŸããã ãã®ããšã«ããããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ã€ã¡ãŒãžã·ããã«å¯Ÿããèæ§ããããã«é«ããªããŸãã
http://cs231n.github.io/convolutional-networks/
ãããããŒãªã³ã°å±€ã®ä»çµã¿ã§ãã ç³ã¿èŸŒã¿ãšããŠæ°ãã3ã€ã®ãã£ãã«ã10ããŸãã¯100ã®ãã£ãã«-ããã€ãã®ããªã¥ãŒã ã®ãã¥ãŒãããããŸãã å¹ ãšé«ããåçŽã«çž®å°ããŸãããæ®ãã®å¯žæ³ã«ã¯åœ±é¿ããŸããã ãã¹ãŠã¯åå§çãªãã®ã§ãã
ããã¿èŸŒã¿ãããã¯ãŒã¯ã¯äœã«é©ããŠããŸããïŒ
ãããã¯ãåŸæ¥ã®å®å šã«æ¥ç¶ããããããã¯ãŒã¯ããããã©ã¡ãŒã¿ãŒãå€§å¹ ã«å°ãªããšããç¹ã§åªããŠããŸãã å®å šã«æ¥ç¶ããããããã¯ãŒã¯ã®äŸãæãåºããŠãã ãããããã§ã¯ã100äžã®éã¿ãååŸããŸããã é¡äŒŒãããæ£ç¢ºã«é¡äŒŒãããã®ããšããšãé¡äŒŒãšåŒã¶ããšã¯ã§ããŸããããåãå ¥åãåãåºåãæã€å¯æ¥ãªç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ãåºåã«1ã€ã®å®å šã«æ¥ç¶ãããã¬ã€ã€ãŒãš100åã®ãã¥ãŒãã³ãããå¥ã®2ã€ã®ç³ã¿èŸŒã¿ã¬ã€ã€ãŒãæã¡ãŸãã³ã¢ãããã¯ãŒã¯ã§ã¯ããã®ãããªãã¥ãŒã©ã«ãããã¯ãŒã¯ã®ãã©ã¡ãŒã¿ãŒã®æ°ã1æ¡ä»¥äžæžå°ããŠããããšãããããŸãã
ãã©ã¡ãŒã¿ãããããå°ãªãå Žåã¯çŽ æŽãããããšã§ãããããã¯ãŒã¯ã®ãã¬ãŒãã³ã°ã¯ç°¡åã§ãã ç§ãã¡ã¯ãããèŠãŠãèšç·Žããã®ã¯æ¬åœã«ç°¡åã§ãã
ããã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯äœãããŸããïŒ
å®éãç»åã®ããã€ãã®éå±€çãªæ©èœãèªåçã«æããŸããæåã®åºæ¬çãªæ€åºåšãããŸããŸãªåŸãã®ç·ãåŸé ãªã©ã ãããã®ãã¡ã圌女ã¯ããã«è€éãªãªããžã§ã¯ããåéããããã«è€éãªãªããžã§ã¯ããåéããŸãã
ãã¥ãŒãã³ãåçŽãªããžã¹ãã£ãã¯ååž°ãåçŽãªåé¡åšãšèŠãªãå Žåããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯åãªãéå±€åé¡åšã§ãã æåã«åçŽãªæšèãéžã³åºããããããè€éãªæšèãçµã¿åãããããã«è€éãªæšèãããã«è€éã«ããæçµçã«ã¯ç¹å®ã®äººãç¹å®ã®æ©æ¢°ã象ãªã©ãéåžžã«è€éãªæšèãçµã¿åãããããšãã§ããŸãã
ç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®ææ°ã®ã¢ãŒããã¯ãã£ã¯ãã¯ããã«è€éã«ãªã£ãŠããŸãã ææ°ã®ImageNetã³ã³ãã¹ãã§åªåãããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ãåãªãç³ã¿èŸŒã¿å±€ã§ã¯ãªããããŒãªã³ã°å±€ã§ãã ãããã¯çŽæ¥å®æãããããã¯ã§ãã ãã®å³ã¯ããããã¯ãŒã¯InceptionïŒGoogleïŒããã³ResNetïŒMicrosoftïŒã®äŸã瀺ããŠããŸãã
å®éãå éšã¯åãåºæ¬ã³ã³ããŒãã³ãã§ãããåãç³ã¿èŸŒã¿ãšåŒãæãã§ãã ã¡ããã©ä»ããããã®å€ããããããããã¯äœããã®åœ¢ã§å·§ã¿ã«çµã¿åããããŠããã ãã§ãã ããã«ãç»åããŸã£ããå€æããã«ãåã«åºåã«è»¢éããçŽæ¥æ¥ç¶ããããŸãã ããã¯ãå¶ç¶ã«ããã°ã©ããŒã·ã§ã³ããã§ãŒãããªãããã«ããã®ã«åœ¹ç«ã¡ãŸãã ããã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®æåŸããæåã«åŸé ãæž¡ãè¿œå ã®æ¹æ³ã§ãã ãŸãããã®ãããªãããã¯ãŒã¯ã®ãã¬ãŒãã³ã°ã«ã圹ç«ã¡ãŸãã
ããã¯éåžžã«å€å žçãªç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ã§ããã ã¯ããåé¡ã«äœ¿çšã§ããããŸããŸãªã¿ã€ãã®ã¬ã€ã€ãŒããããŸãã ãããããã£ãšé¢çœãã¢ããªã±ãŒã·ã§ã³ããããŸãã
https://arxiv.org/abs/1411.4038
ããšãã°ãå®å šç³ã¿èŸŒã¿ãããã¯ãŒã¯ïŒFCNïŒãšåŒã°ããç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ããããŸãã 圌ãã¯ãã£ãã«è©±ãããŸããããããã¯ã¯ãŒã«ãªããšã§ãã ããªãã¯æåŸã®å€å±€ããŒã»ãããã³ãåãã¯ããããšãã§ããŸããããã¯å¿ èŠãããŸãã-ãããæšãŠãŠãã ããã ãããŠããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ãä»»æã®ãµã€ãºã®ç»åã«å¯ŸããŠéæ³ã®ããã«æ©èœããŸãã
ã€ãŸããäŸãã°ãåç«ãç¬ããã®ä»ã®ç»åã§1000ã¯ã©ã¹ãå®çŸ©ããããšãåŠãã åŸãæåŸã®ã¬ã€ã€ãŒãåãããããå¥ããããšãªããç³ã¿èŸŒã¿ã¬ã€ã€ãŒã«å€æããŸããã åé¡ã¯ãããŸãã-ããªãã¯äœéãæ°ããããšãã§ããŸãã 次ã«ããã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ããã¬ãŒãã³ã°ãããåããŠã£ã³ããŠ100 * 100 psã§åäœããããšãããããŸããããç»åå šäœã§ãã®ãŠã£ã³ããŠãä»ããŠå®è¡ã§ããåºåã§äžçš®ã®ããŒãããããäœæã§ããŸã-ãã®ç¹å®ã®ç»åç¹å®ã®ã¯ã©ã¹ã
ããšãã°ããã¹ãŠã®ã¯ã©ã¹ã«å¯ŸããŠãããã®ããŒããããã1000åäœæããããã䜿çšããŠç»åå ã®ãªããžã§ã¯ãã®äœçœ®ã決å®ã§ããŸãã
, , .
http://cvlab.postech.ac.kr/research/deconvnet/
â Deconvolution networks. , .
Deconvolution â . â , .
ããã¯äœã§ãã , Upsampling. - - , , 1 ps. , , - . .
, - 10*10 ps, Upsampling , - , Upsampling .
â , , , - . /, , , . ããã¯é¢çœãã§ãã
. â , - . , . â , .
, , -, .
https://arxiv.org/abs/1411.5928
, , , - , â , , , : - - . - , .
, . , . , , , - .
https://arxiv.org/abs/1508.06576
, , , . , . .
https://arxiv.org/abs/1508.06576
. , , , . , â , .
, , , , . , , , . () , , , . Backpropagation. , , , . .
, . - â . , , , . , - , -.
https://arxiv.org/abs/1603.03417
, 1 . , : -, -, , .
, â , , .
, , , .
.
:
Recurrent Neural Networks, RNN
. , FNN- , - . . , , .
Feed-Forward , . - ( ). , â . .
â . â . . , .
, Feed-Forward- â , . , .
Feed-Forward , , , . , , . . .
http://colah.github.io/posts/2015-08-Understanding-LSTMs/
? . Backpropagation, , , Backpropagation .
â Backpropagation through time. â - , , 10, 20 100, , Backpropagation.
ããããåé¡ããããŸãã â 10, 20, 100 â , , 100 . - . , LSTM GRU- .
https://deeplearning4j.org/lstm
, , gate, , , .. gate , . , , , , , 100 . , - , . , .
- , . LSTM, GRU.
http://kvitajakub.github.io/2016/04/14/rnn-diagrams
, , , , , , , . gate, « â », « â ».
. , , .
, , , , â . ! , , , â , , ?
. , - . , , 2 , .
Bidirectional â . , , : 2 â , . , , .
, , - , â .
, â , . â . 4 , -. , , 4 .
: , . . , , , 10, , .
(2015 ). â LeNet, . , . , .
â , . , , - .
. , , . .
https://arxiv.org/abs/1507.01526
Grid LSTM, . â , - - , . , gate, . , - ? , , !
â , . 1-2 , . , , .
, , , .
.
(Multimodal Learning)
, ,
â , 2 , , . 1 â , , . !
http://arxiv.org/abs/1411.4555
, â . , , , , , . , , . .
, HighLoad++ . , , YouTube .
. , , - - . , .
2- . .
https://www.cs.utexas.edu/~vsub/
, , . , â .
äŸïŒ
- , ;
- , ;
- , , .
!
, .
http://arxiv.org/abs/1411.2539
- , , , . , 2 : , . 2 .
, : â , â . .
ãšããã§ãããã¹ãã®ç»åãçæãããããã¯ãŒã¯ãæ§ç¯ããããšããè©Šã¿ã¯æ¢ã«ãããŸããããã¯é¢çœãã§ãããããåäœããŸãããŸã ããŸãè¯ããããŸããããå¯èœæ§ã¯èšå€§ã§ãã
ã·ãŒã±ã³ã¹åŠç¿ãšseq2seqãã©ãã€ã
å ¥åããã³/ãŸãã¯åºåã§ä»»æã®é·ãã®ã·ãŒã±ã³ã¹ã䜿çšããå Žå
2çªç®ã®èå³æ·±ããããã¯ã¯ãã·ãŒã±ã³ã¹åŠç¿ãŸãã¯seq2seqãã©ãã€ã ã§ãã 翻蚳ãããŸããã ã¢ã€ãã¢ã¯ãå€ãã®ã¿ã¹ã¯ãã·ãŒã±ã³ã¹ãæã€ããšã«åž°çãããšããããšã§ãã ã€ãŸããåé¡ããå¿ èŠãããç»åã ãã§ãªãã1ã€ã®çªå·ãäžããã ãã§ãªãã1ã€ã®ã·ãŒã±ã³ã¹ããããåºåã«ã¯å¥ã®ã·ãŒã±ã³ã¹ãå¿ èŠã§ãã
ããšãã°ã翻蚳ã¯ã·ãŒã±ã³ã¹2ã·ãŒã±ã³ã¹åŠç¿ã®å€å žçãªã¿ã¹ã¯ã§ãã圌ãã¯ããã¹ããè±èªã§èšå®ãããã©ã³ã¹èªã§ååŸããããšèããŠããŸãã
å®éããã®ãããªã¿ã¹ã¯ã¯æ°å€ããããŸãã ããã¯ãåçã®èª¬æã®å Žåã§ãã
http://karpathy.github.io/2015/05/21/rnn-effectiveness/
ç§ãã¡ã調ã¹ãéåžžã®ãã¥ãŒã©ã«ãããã¯ãŒã¯-äœããé転ãããããã¯ãŒã¯ãé転ããåºåã§åé€ãã-ã¯ããããããã®ã§ã¯ãããŸããã
One to manyãšãããªãã·ã§ã³ããããŸãã 圌ãã¯åçããããã¯ãŒã¯ã«éãããããŠåœŒå¥³ã¯ä»äºã«è¡ããä»äºããããã®åçã®èª¬æãäœæããŸããã ãã
å察æ¹åã«ã§ããŸãã ããšãã°ãããã¹ãã®åé¡ã ããã¯ãã¹ãŠã®ããŒã±ãã£ã³ã°æ åœè ã®ãæ°ã«å ¥ãã®ã¿ã¹ã¯ã§ã-ãã€ãŒããåé¡ããããã«-圌ãã¯ææ çãªè²ã®é¢ã§è¯å®çãŸãã¯åŠå®çã§ãã ææ¡ããªã«ã¬ã³ããã¥ãŒã©ã«ãããã¯ãŒã¯ã«å°ããæåŸã«1ã€ã®æ°åãåºããŸãã-ã¯ããè¯å®çãªè²ã®ãã€ãŒããããããåŠå®çãªè²ã®ãã€ãŒãããã¥ãŒãã©ã«ãªã©ã
翻蚳ã«ã€ããŠã®è©±ããããŸãã é·ãéã1ã€ã®èšèªã§ã·ãŒã±ã³ã¹ãé転ããŠããŸããã ãã®åŸããããã¯ãŒã¯ã¯æ©èœããå¥ã®èšèªã§ã·ãŒã±ã³ã¹ãçæãå§ããŸããã ããã¯äžè¬çã«æãäžè¬çãªèšå®ã§ãã
å ¥åãšåºåãåæãããå Žåãå¥ã®èå³æ·±ãèšå®ããããŸãã ããšãã°ãç»åã®ãã¹ãŠã®ãã¬ãŒã ã«æ³šéãä»ãããå Žåãäœãããããã©ããã
ãã®å³ã¯ãã·ãŒã±ã³ã¹2ã·ãŒã±ã³ã¹åŠç¿ã®ãã¹ãŠã®ãªãã·ã§ã³ã瀺ããŠããŸããããã¯éåžžã«åŒ·åãªãã©ãã€ã ã§ãã ãã¥ãŒã©ã«ãããã¯ãŒã¯å ã®ãã¹ãŠã埮åå¯èœã§ãããšããç¹ã§åŒ·åã§ã-ãããŠãè°è«ãããã¥ãŒã©ã«ãããã¯ãŒã¯ãå éšã§å®å šã«åŸ®åå¯èœã§ãããšããããšã¯ãããã°ããšã³ãããŒãšã³ãã§ãã¥ãŒã©ã«ãããã¯ãŒã¯ããã¬ãŒãã³ã°ã§ããããšãæå³ããŸããå éšã§äœãèµ·ãã£ãŠãããã¯ããªãã«ã¯ãŸã£ããé¢ä¿ãããŸããã ãã¥ãŒã©ã«ãããã¯ãŒã¯èªäœã察åŠããŸããè±èªã®å€æ°ã®äŸãå ¥åããããã©ã³ã¹èªã®å€æ°ã®äŸãåºåã§ãããåªããŠããŸãã圌女èªèº«ã翻蚳æ¹æ³ãåŠç¿ããŸãã ãããŠãããããªãã倧èŠæš¡ãªããŒã¿ããŒã¹ãšãããè¿œãæãããã®åªããèšç®èœåãæã£ãŠãããªããæ¬åœã«è¯ãå質ã§ã
ã»ãšãã©èª°ã話ããŠããªãããGoogle Speech RecognitionãBaiduããŸãã¯Microsoftãæ©èœããªããã1ã€ã®éåžžã«éèŠãªããšã¯ãCTCã§ãã
https://github.com/baidu-research/warp-ctc
CTCã¯éåžžã«æ±ãã«ããåºåå±€ã§ãã 圌ã¯äœãããŠããŸããïŒ ãã®ã·ãŒã±ã³ã¹å ã®ã¢ã©ã€ã¡ã³ããå®éã«éèŠã§ã¯ãªãå€ãã®ã¿ã¹ã¯ããããŸãã é³å£°èªèã¿ã¹ã¯ããããŸãã ããšãã°ãé³ãåããããã50ããªç§ã®çããã¬ãŒã ã«ã«ããããåºåã§ãã©ã®åèªã§ããããé³çŽ ã®ã·ãŒã±ã³ã¹ãçæããå¿ èŠããããŸãã æŠããŠãå ã®ã·ã°ãã«ã®ãã®é³çŽ ãŸãã¯ãã®é³çŽ ãã©ãã«ãã£ããã¯éèŠã§ã¯ãããŸããã åºå£ã§ç°¡åãªåèªãååŸããããã«ãçžäºéã®é åºã®ã¿ãéèŠã§ãã
æ£ç¢ºãªäœçœ®ã«é¢ãããã¹ãŠã®æ å ±ãæšãŠãããšãã§ãããšããäºå®ã¯ãå®éã«ã¯ãããããããŸãã ããšãã°ããã¹ãŠã®ãµãŠã³ããã¬ãŒã ã«æ£ç¢ºãªé³çŽ ããŒã¯ãä»ããå¿ èŠã¯ãããŸããããã®ãããªããŒã¯ã¢ãããååŸããã®ã¯éåžžã«é«äŸ¡ãªããã§ãã ãã¹ãŠãããŒã¯ã¢ããããç·ãæ€ããå¿ èŠããããŸãã
ãã¹ãŠãåãåºããŠæšãŠãããšãã§ããŸã-å ¥åããããåºå£ããããŸã-åºåã·ãŒã±ã³ã¹ã«é¢ããŠäœãèµ·ããã¹ãã-ã€ãŸãããã®ããªãããŒãªCTCã¬ã€ã€ãŒããããããèªäœãäœããã®æŽåãè¡ãããããåã³çµäºããŸã-ãã®ãããªããªãããŒãªãããã¯ãŒã¯ãèšç·Žããããã«ãããªãã¯äœãããŒã¯ããªãã£ãã
ããã¯åŒ·åãªããšã§ãããææ°ã®ãã¹ãŠã®ããã±ãŒãžã«å®è£ ãããŠããããã§ããããŸããã ããããããšãã°ã1幎åãBaiduã¯CTCã¬ã€ã€ãŒã®å®è£ ãã¬ã€ã¢ãŠãããŸãã-ããã¯çŽ æŽãããããšã§ãã
ããŸããŸãªã¢ãŒããã¯ãã£ã«é¢ããããã€ãã®èšèã
https://github.com/farizrahman4u/seq2seq
å€å žçãªãšã³ã³ãŒããŒãã³ãŒããŒã¢ãŒããã¯ãã£ãŒããããŸãã ç§ã話ãã翻蚳ã®äŸã¯ãã»ãŒå®å šã«ãã®ã¢ãŒããã¯ãã£ã«éå®ãããŠããŸãã
å ¥åãã¥ãŒã©ã«ãããã¯ãŒã¯ã1ã€ãããããã«åèªãé ä¿¡ãããŸãã ãã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®åºåã¯ãæã®çµããã®èšå·ãæå®ããããŸã§ç¡èŠãããŸãã ãã®åŸã2çªç®ã®ãããã¯ãŒã¯ãã±ãŒã¹ã«å«ãŸããæåã®ãããã¯ãŒã¯ã®ç¶æ ãèªã¿åããããããåºåã¯ãŒãã®çæãéå§ããŸãã åã®ã¹ãããã§ã®çµæã¯å ¥åã«éãããŸãã
åäœããŸãã å€ãã®ç¿»èš³ã·ã¹ãã ã¯ãã®ããã«æ©èœããŸãã
ãããããã®ã¢ãŒããã¯ãã£ã«ã¯1ã€ã®åé¡ããããŸã-ããã«ããã¯ãã§ãã éä¿¡ãããç¶æ ãã¯ãã«ïŒé衚瀺局ã®ãµã€ãºïŒã¯å¶éãããåºå®ãããŠããŸãã ã€ãŸããçãæã§ããéåžžã«é·ãæã§ãåãã§ããããšãããããŸããããã¯ããŸãè¯ããããŸããã é·ãæããã®ããªã¥ãŒã ã«åãŸããªãããšãå€æããå ŽåããããŸãã
https://research.googleblog.com/2016/09/a-neural-network-for-machine.html
圌ããèšãããã«ã建ç¯ã¯æ³šæãæã£ãŠç»å Žããã
泚æã¯ãã®ãããªããªãããŒãªãã®ã§ãããå®éã«ã¯å®éã«ã¯éåžžã«åçŽã§ãã ããã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®åºåãã³ãŒããŒã以åã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®åºåå€ã§ã¯ãªãããã¹ãŠã®äžéç¶æ ã§ãããçš®ã®ãã©ã³ã¹ã§èŠããšããèãæ¹ã§ãã éã¿ã¯ããã³ãŒããŒãæ©èœããæçµçãªå€§éã«ãããã®åç¶æ ãåã蟌ãããã«å¿ èŠãªéã®ä¿æ°ã§ãã
ã€ãŸãã泚æã¯å®éã«ã¯ããšã³ã³ãŒãã®ä»¥åã®ãã¹ãŠã®ç¶æ ã®åçŽãªç·åœ¢çµåã§ããããããèšç·ŽãããŠããŸãã
äºå®ã«æ³šç®ãããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯éåžžã«ããŸãæ©èœããŸãã 翻蚳ããã®ä»ã®è€éãªã¿ã¹ã¯ã§ã¯ããããã¯æ³šæãæããã«ãã¥ãŒã©ã«ãããã¯ãŒã¯ãããå質ãéåžžã«åªããŠããŸãã
http://kelvinxu.github.io/projects/capgen.html
ãã®ãããªãããã¯ãŒã¯ã®è¿œå ããŒãã¹ã ãã®å³ã¯ã2ã€ã®ç°ãªããã¥ãŒã©ã«ãããã¯ãŒã¯ã®çµã¿åããã瀺ããŠããŸããç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ããããã€ãã®å åãåŸãåŸãååž°ãã¥ãŒã©ã«ãããã¯ãŒã¯ãããã¹ããçæããŸãã 泚æã®æŠå¿µãå®è£ ããããã€ãã®åçãèšå®ãããšãéã¿ã倧ããç¹å®ã®åèªã®çæãèŠãããšãã§ããŸãã ããã¯ãç¹å®ã®ç¬éã«ç»åã®ã©ã®ãã¯ã»ã«ããã®ç¹å®ã®åèªãçæãã圹å²ãæãããããå®éã«ç€ºããŸãã ã€ãŸãããã¥ãŒã©ã«ãããã¯ãŒã¯ã泚æãæã£ãŠããããã«èŠãããã®ã§ãã
ãšããã§ã泚æã®æŠå¿µã¯ãã¹ãŠã®ã©ã€ãã©ãªã«å®è£ ãããã«ã¯ã»ã©é ããã€ãŸããæ¢æã®ããã¯ã¹ãœãªã¥ãŒã·ã§ã³ã¯ãããŸããã 誰ããèªåã®ä»äºã®äžéšãšããŠå ¬éããæ¢è£œã®ã³ãŒããèŠã€ããããšãã§ããŸãã
http://kelvinxu.github.io/projects/capgen.html
CNN + RNN with attention =çŸããåçã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ãSTOPãµã€ã³ã«é¢ããããã¹ããçæãããšããããã¯æ¬åœã«ãã®ãµã€ã³ã®ããã«èŠããŸã-ãã®ééãç¹å®ã®STOPãµã€ã³ã®çæãžã®è²¢ç®ã¯éåžžã«é«ããä»ã®ãã¹ãŠã®ãã¯ã»ã«ã¯ã»ãšãã©åœ¹å²ãæãããŸããã
ããã¯èå³æ·±ãæŠå¿µã§ãã泚æããŠãã ããã ãŸããå€ã䜿çšãããŸãã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ãæäœããããã®ãã¬ãŒã ã¯ãŒã¯ãšã©ã€ãã©ãª
éåžžã«çãã¬ãã¥ãŒ
å®éãäœæéã話ãããšãã§ããŸãã ç§ã¯ããªãã«äŒããç®çã¯ãããŸãã-ã¯ãããã®ã©ã€ãã©ãªãŸãã¯ããã䜿çšããŠãã ãã-ããã¯ããã§ã¯ãªãããã§ãã èšå€§ãªæ°ã®ã©ã€ãã©ãªããããŸãã å€ããå°ãªããé¢é£ã®ç°ãªãã©ã€ãã©ãªã®ãªã¹ããæäŸããŸãã
詳现ãªã¹ãïŒ http : //deeplearning.net/software_links/
ãŠãããŒãµã«ã©ã€ãã©ãªãšãµãŒãã¹ïŒ
- Torch7 ïŒ http://torch.ch/ ïŒ[Lua]
- TensorFlow ïŒ https://www.tensorflow.org/ ïŒ[PythonãC ++]
- Theano ïŒ http://deeplearning.net/software/theano/ ïŒ[Python]
- ã±ã©ã¹ ïŒ http://keras.io/ ïŒ
- ã©ã¶ãã¢ ïŒ https://github.com/Lasagne/Lasagne ïŒ
- ãããã¯ ïŒ https://github.com/mila-udem/blocks ïŒ
- pylearn2 ïŒ https://github.com/lisa-lab/pylearn2 ïŒ
- Microsoft Cognitive ToolkitïŒCNTKïŒ ïŒ http://www.cntk.ai/ ïŒ[PythonãC ++ãCïŒãBrainScript]
- ããªã³ ïŒ http://neon.nervanasys.com/ïŒ[Python ]
- Deeplearning4j ïŒ http://deeplearning4j.org/ ïŒ[Java]
- MXNet ïŒ http://mxnet.io/ ïŒ[C ++ãPythonãRãScalaãJuliaãMatlabãJavascript]
- ...
ãŸããæ®éçãªã©ã€ãã©ãªãããããã®å€ãã«ã€ããŠèããããšããããŸãã
ããšãã°ãTensorFlowïŒGoogleïŒã¯ãæè¿ã§ã¯ãããŸããããããæã人æ°ã®ãããã®ã®1ã€ã§ãã Pythonããã³C ++ã§äœ¿çšã§ããŸãã
çŸåšãFacebookãç©æ¥µçã«ãµããŒãããŠããTorchã©ã€ãã©ãªããããŸãã ããã¯Luaèšèªã§ãã ããããæãããªãã§ãã ãããããã¯å®éã«ã¯ãŒã«ãªèšèªã§ãã ãã®ã©ã€ãã©ãªã«ã¯å®è£ ããããã®ããããããããLuaã³ãŒãã®åœ¢ã§æ°ããç 究ããããããããŸãã ããã¯çŽ æŽãããã
Theanoã©ã€ãã©ãªããããŸã-TensorFlowã¯çŸåšããããããã«çœ®ãæããŠããŸãããTheanoã®åšãã«ã¯å€ãã®ç°ãªãã¯ãŒã«ãªé«ã¬ãã«ã©ãããŒãæ§ç¯ãããŠããŸã-ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯æ°è¡ã§èšè¿°ã§ããŸãã ããã¯æ¬åœã«çŽ æŽãããã§ãïŒ
Kerasãªã©ã®ãããã®ã©ãããŒã®äžéšã¯ã圌ããèšãããã«ãããã¯ãšã³ãã®ããã«TensorFlowã§åäœã§ããŸãã ã€ãŸããTensorFlowã¯ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®èŠ³ç¹ã§ã¯ããªãäœã¬ãã«ã®ã³ãŒãã§ãããKerasã¯é«ã¬ãã«ã®ã³ãŒãã§ããããåäžè¡ã®ã¬ã€ã€ãŒã䟿å©ã§ãã
Microsoftã¯äœããå ¬ââéããŸãããNeepãDeeplearning4j-ãŸããªã±ãŒã¹-Deeplearningã®Javaã©ã€ãã©ãªããããŸãã Javaã«ã¯ã»ãšãã©ãããŸããã å€ãã®Pythonããã³C ++ã ä»ã®èšèªã§ã¯å°ãªãã
ããã«ããããªåŠççšã®ç¹å¥ãªããŒã«ããããŸãã
ç»åãšãããªã®åŠçïŒ
- OpenCV ïŒ http://opencv.org/ ïŒ[CãC ++ãPython]
- Caffe ïŒ http://caffe.berkeleyvision.org/ ïŒ[C ++ãPythonãMatlab]
- Torch7 ïŒ http://torch.ch/ ïŒ[Lua]
- clarifai ïŒ https://www.clarifai.com/ ïŒ
- Google Vision API ïŒ https://cloud.google.com/vision/ ïŒ
- ...
ããã«OpenCVãå«ããŸããã ãã¡ãããããã¯Deeplearningã©ã€ãã©ãªã§ã¯ãããŸããããä»ã®ã©ã€ãã©ãªãšããŸãçµ±åã§ããŸãã
Caffeã¯çŽ æŽãããã©ã€ãã©ãªã§ãããæ¬çªç°å¢ã§äœ¿çšããŸããã ããã¯ãã©ã¹ã®ã©ã€ãã©ãªã§ãããéåžžã«é«éã§ããå®éãããé«éã§ã¯ãããŸããã çŸåšããã¥ãŒã©ã«ãããã¯ãŒã¯ãç¿åŸããŠãã人ã¯ãäœããã®çç±ã§TensorFlowã®ã¿ãèããŠããŸãããããã§ããŸã ã¯ãŒã«ã§ãã ããããCaffeãå«ãä»ã®åªãããœãªã¥ãŒã·ã§ã³ãããããããããšã«æ³šæããŠãã ãã-éåžžã«ã¯ãŒã«ãªããšã§ãã
ããã«ãWEBã§äœ¿çšã§ããããŸããŸãªAPIãããã€ããããŸãã
é³å£°èªèã å®éã«æªåããŠããŸãã
é³å£°èªèïŒ
- Microsoft Cognitive ToolkitïŒCNTKïŒ ïŒ http://www.cntk.ai/ ïŒ[PythonãC ++ãCïŒãBrainScript]
- KALDI ïŒ http://kaldi-asr.org/ ïŒ[C ++]
- Google Speech API ïŒ https://cloud.google.com/ ïŒ
- Yandex SpeechKit ïŒ https://tech.yandex.ru/speechkit/ ïŒ
- Baidu Speech API ïŒ http://www.baidu.com/ ïŒ
- wit.ai ïŒ https://wit.ai/ ïŒ
- ...
é³å£°èªèçšã®ã¯ãŒã«ãªKALDIã©ã€ãã©ãªã1ã€ãããŸãããããã¯ãã©ã¹ã§ãã ããããäžè¬çã«ã倧äŒæ¥ã§ã¯é³å£°èªèã¯å€ããå°ãªããééãããŠããŸãããªããªãã誰ãé³å£°ãšé³å£°ã«é¢ãã倧ããªããŒã¿ã»ãããæã£ãŠããªãããã§ãã ããããYandexãGoogleãBaiduã«ã¯å€æ°ã®APIããããMicrosoftã«ãAPIãããããã§ãã
ã¯ãŒããïŒ
- Torch7 ïŒ http://torch.ch/ ïŒ[Lua]
- Theano / Keras / ... [Python]
- TensorFlow ïŒ https://www.tensorflow.org/ ïŒ[C ++ãPython]
- Google Translate API ïŒ https://cloud.google.com/translate/ ïŒ
- ...
ããã¹ãã«ã€ããŠã¯ãç¹ã«ç¹å¥ãªãã®ã¯ãããŸãããããã¹ãŠã®ãŠãããŒãµã«ã©ã€ãã©ãªã¯åªããŠããŸãã KerasïŒãŸãã¯ä»ã®ãããã¯åé¡ã§ã¯ãªãïŒãåããæ°è¡ã§äœããæžã-ããã¹ããæ±ãæºåãã§ããŠãããã¥ãŒã©ã«ãããã¯ãŒã¯ããããŸãã ãŸãã¯ä»ã®ã©ã€ãã©ãª-ããã¯éèŠã§ã¯ãããŸããã
ããã ãã§ããããããšãã ã©ã®ãŠãããŒãµã«ã©ã€ãã©ãªãŒãæ¡çšãã¹ãããšãã質åã«å¯Ÿããæ®éçãªçãã¯ãããŸããã ã¿ã¹ã¯ãèŠãŠãã ããã å€ãã®åŸ®åŠãªç¹ããããŸã-ãããŠãããªããæã£ãŠãããã¯ãããžãŒã®çš®é¡ããã®äžã«åã蟌ãŸããŠãããã®ãæ¢æã®ã³ãŒãã¯ãã§ã«èªç¶ã®ãã®ã§ã-http://github.com/ã§äœ¿çšã§ããã³ãŒãã¯æ¬åœã«ãããããããŸã ã ããã¯ãæ éã«åãçµãå¿ èŠããããšã³ãžãã¢ãªã³ã°ã¿ã¹ã¯ã§ãã åäžã®æ®éçãªçãã¯ãªããããããŸããã
ãããã質å
質å ïŒåå¿è åãã«ããçš®ã®æç®ãã¢ããã€ã¹ããŠããããŸããïŒãã¥ãŒã©ã«ãããã¯ãŒã¯ã®ããã°ã©ãã³ã°æ¹æ³ãèªãã ããèŠãããç解ãããããã«ã¯ã©ãããã°ããã§ããããïŒ
åç ïŒããã§ã®å€ãã¯ãçŸåšã®ã¬ãã«ãç解ãæ·±ãããçšåºŠã«äŸåããŸãã å®éãèšå€§ãªæ°ã®ããã°ããããŸãã æåã«ããã·ã¢èªãå¿ããŠãã ãã-å®éã«ã¯äœããããŸããã Habréã«ã¯ããã€ãã®ç¿»èš³ããããŸãããããã¯èªç¶çã«ååšããé åã®èæ¯ã«å¯ŸããŠæ·±å»ã§ã¯ãããŸããã
è±èªã§ã¯ãããŸããŸãªäŸãç解ã§ããã¯ãŒã«ãªããã°ãèšå€§ã«ãããŸãã ãããããããŸããã°ãŒã°ã«ã§æ€çŽ¢ããŠãç¹å®ã®ãããã¯ã§äœããèŠã€ããŠãã ããã ããŸããŸãªãã¥ãŒããªã¢ã«ããããŸããããããè±èªã§ãå€å°ãªããšãå°ãããªã£ãŠããŸãã 800ããŒãžã®Deeplearningããã¯ããããŸããããã¯çŸåšAMT-Pressã®çŽã§å ¥æã§ããé·ãéPDFã§å ¥æã§ããŸãã
äžè¬ã«ãæç®ããããŸãã Courseraãªã©ããªã³ã©ã€ã³ã§ããã€ãã®ã³ãŒã¹ããããã³ãŒã¹ããªãã©ã€ã³ã§èµ·åããããšããŠããŸãã ç¹ã«ãç§ã¯ãããã®ã³ãŒã¹ã®ããããã«ããã«åå ããŸãã
å®éã«ã¯ãããªãå€ãã®ç°ãªããªãã·ã§ã³ããããŸãã ã€ã³ã¿ãŒããããèŠãŠãã ãã-ãã ããå€ãã®æ©äŒããããŸãã 圌ãã®å€ãã¯ããŸã ã«æ§ã ãªå€åœæåŠãèªãã§ããŸãããããã¯æ¬åœã«è¯ããŠå æ¬çãªãã®ã§ãã
ãããåæã«ãGitHubã®ã³ãŒããåªããŠããŸãã å€ãã®ã³ãŒããå ¬éãããŠãããå°ãªããšã確èªã§ããŸãã å€ãã®å Žåãèªãããšãã§ããŸãããããã»ã©æããã®ã§ã¯ãããŸããã ãããŠããã®ã³ãŒãã«ã¯ããããã©ã®ããã«æ©èœãããã«ã€ããŠã®ããããããã³ã¡ã³ãããã°ãã°ãããŸãã ã€ã³ã¿ãŒãããã«è¡ãã ãã§ã-ããã«ã¯ãã¹ãŠããããããããŸãã
質å ïŒäžè¬ã«ããã¥ãŒã©ã«ãããã¯ãŒã¯ããã¬ãŒãã³ã°ããæ¹æ³ã¯äœã§ããïŒ ã€ã³ã¿ãŒãããããããããã®åçãç°¡åã«ã°ãŒã°ã«æ€çŽ¢ããããšã¯å¯èœã§ããããããšãä»ã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ãèšç·Žãããã¥ãŒã©ã«ãããã¯ãŒã¯ãå©çšã§ããŸããïŒ
åç ïŒã¯ããããã¯çŽ æŽããã質åã§ãã ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®ãã¬ãŒãã³ã°ã§ã¯ãä»åŸæ°å¹Žã§å€ãã®é²æ©ããããšæããŸãã ããŸããŸãªã¢ãããŒãããããŸãã æåã«ãã¯ã-ããŒã¿ã»ãããå ¥åããŠãã¬ãŒãã³ã°ãããšã-ããã¯å€å žçãªã¢ãããŒãã§ãã 圌ã¯ã圌ããéããããšã¯ã§ããŸããã圌ã¯å€ãã®å Žååºæ¬çã§ãã
ãããä»ããšããã§ã転移åŠç¿ãšåŒã°ããå¥ã®ã¢ãããŒãããã°ãã°åºæ¬çãªã¢ãããŒãã«ãªããŸãã åãImageNetäžã§ãããŸããŸãªã¿ã¹ã¯ã§ãã¬ãŒãã³ã°ãããããã§ã«å ¬éãããŠããå€æ°ã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ããããŸãã 1000ã¯ã©ã¹ã®æ¢è£œã®ImageNetãã¥ãŒã©ã«ãããã¯ãŒã¯ã䜿çšããŠãäžéšã®ç¹å¥ãªã¯ã©ã¹ã§åãã¬ãŒãã³ã°ã§ããŸãã ããšãã°ãã¯ã©ã¹ã®ç»åã1000æãããªãããããã®ãããªããªã¥ãŒã ã§æåããåªãããã¥ãŒã©ã«ãããã¯ãŒã¯ããã¬ãŒãã³ã°ããããšã¯ã§ããªããããããã¯ããç°¡åã«ãªããŸãã ãã£ãŒããã¥ãŒã©ã«ãããã¯ãŒã¯ããã¬ãŒãã³ã°ããã«ã¯ã倧éã®ããŒã¿ãæ¬åœã«å¿ èŠã§ãã æ°åäžãæ°çŸäžã®ãªããžã§ã¯ãã§ãã ãã ããæ¢è£œã®ã°ãªãããããå Žåã¯ãããã䜿çšããŠå°ããã¬ãŒãã³ã°ãããã§ã«æ£åžžãªçµæãåŸãããŠããŸãã 転移åŠç¿ã¯è¯ãã¢ãããŒãã§ãã 圌ã¯åããŠããŸãã
ä»ã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ãæãããšãã®ãªãã·ã§ã³-ãã®ãããªãªãã·ã§ã³ããããŸãã ãããã¯çç£ãããç 究ã§ãã ããã¯éåžžã«ã¯ãŒã«ãªãããã¯ã§ããåŸãã®ã¯çŽ æŽãããããšã§ãã åªããå¶äœãœãªã¥ãŒã·ã§ã³ã¯ããããŸããããç§åŠçãªåŽé¢ã«èå³ããããªããããã§ããèªãã§ãã ãããããçš®ã®äžçã®ã¢ãã«ãå«ãæåž«ãšãã¥ãŒã©ã«ãããã¯ãŒã¯ãå¥ã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ãæããŠããããšããéåžžã«ã¯ãŒã«ãªèšäºããããŸãã ã
質å ïŒãã¥ãŒã©ã«ãããã¯ãŒã¯ã®æ¢åã®ã¢ãŒããã¯ãã£ãå€æŽããããç¬èªã®ã¢ãŒããã¯ãã£ãäœæãããã§ããããŒã«ã¯ãããŸããïŒ
åç ïŒèŠèŠçãªããŒã«ãå¿ èŠãªå Žåã¯ãã¯ãã§ã¯ãªããããããã芧ãã ããã TensorFlowã«ã¯äœããã¬ã³ããªã³ã°ãããã©ã°ã€ã³ãããã€ããããŸããã ããããå šäœãšããŠãããã¯å®éã«ã¯ããã»ã©å€§ããªåé¡ã§ã¯ãããŸããããªããªãããã¥ãŒã©ã«ãããã¯ãŒã¯ã¯å€ãã®å Žåãäœããã®æ§é ãããã¹ããã¡ã€ã«ããŸãã¯ã³ãŒãã®åœ¢ã§äžããããå€æŽããããšã¯ããã»ã©é£ãããªãã®ã§ãããã«ã¬ã€ã€ãŒãè¿œå ããŠããããããã°ã©ã ã§ããŸãã ããã¯å®éã«ã¯ããã°ã©ãã³ã°ã§ãããããŸãããå®éããã®ãããªç¹å¥ãªDSLã§ãã ããã€ãã®ã¬ã€ã€ãŒãè¿œå ããŸããã
ããããã¹ãŠã®äœåã§æãé£ããã®ã¯ãã¬ã€ã€ãŒéã§èŠ³å¯ãã次å ã§ãã ãããã®å€æ¬¡å é åã«ãã³ãœã«ãã©ã®ããã«é 眮ãããŠããããæ¬åœã«ç解ããŠããªãå Žåã次å ãšæ··åãããå¯èœæ§ããããŸãã ããã¯ãããããã¹ãŠã®äžã§æãé£ããéšåã§ãã
質å ïŒããªãã¯éåžžã«å€ãã®ç°ãªããªã«ã¬ã³ãã¢ãŒããã¯ãã£ã«ã€ããŠè©±ããŸãã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã ãããŠãããªãã¯æ£ç¢ºã«äœã䜿çšããã©ã®ã¿ã¹ã¯ã®ããã«
åç ïŒã»ãšãã©ã®ã¿ã¹ã¯ã§ã¯ãLSTMããã¯ã¹ããã®åçŽãªãã¥ãŒã©ã«ãããã¯ãŒã¯ãååã«æ©èœããååãªæ·±ããšãµã€ãºããããŸãã ããã¹ãåé¡ãã·ãŒã±ã³ã¹å ã®ãããããã®ã®åé¡ã«ã¯å€ãã®ã¿ã¹ã¯ããããŸãã LSTMãããã¯ãŒã¯ã®ããããã§éå§ããå Žåãããã¯ååãšããŠéåžžã®éå§ã§ãã ãã®å Žæã§åæ¹åæ§ã圹ç«ã€ããšãç解ããŠããå Žåãåæ¹åLSTMãå®è¡ããŠããŸãã
ããããçš®é¡ã®ã¯ãŒã«ãªãªãã·ã§ã³ãã泚æãæã£ãŠéå§ããããšã¯çŽ æŽãããããšã§ããããããããŒãããããã°ã©ã ããã®ã¯é£ãããããéå§ããã®ã¯å°é£ã§ãã çµå±ãäºçŽ°ãªããšã§ã¯ãããŸããã ãããŠãããªããåã£ãŠäœ¿ãã³ãŒãã®ãã®ãããªè¯ãéšåã¯ããŸããããŸããã ç§ã«ãšã£ãŠããããŸã§ã®ããŒã¹ã©ã€ã³ã¯LSTMãããã¯ãŒã¯ã§ã-åæ¹åãŸãã¯åæ¹åã§ãã ããã¹ããšç»åãåé¡ããããã«ãããã䜿çšããŸããïŒæ°åèªèïŒã
質å ïŒæå·åã¢ã«ãŽãªãºã ã解èªããããã«ãã¥ãŒã©ã«ãããã¯ãŒã¯ã䜿çšãããŠããããšãç¥ã£ãŠããŸããããšãã°ããã¬ãŒã³ããã¹ããå ¥åã§ãããæå·æãåºåã§ãã ãããŠãå察æ¹åã«ãæå·åãããããã¹ããéä¿¡ããããã¬ãŒãã³ã°ã¯åã«åºåã§ãªãŒãã³ãåãåããŸãã åé¡ã¯ããã®åéã§çŸåšã©ã®ãããªé²å±ããããå®éã«æ©èœããã®ãããããŠããã«ã©ã®ãããªã¢ãŒããã¯ãã£ã䜿çšã§ããã®ããšããããšã§ãã
åç ïŒããã«ã€ããŠã¯ããŸãèšããŸããã ç§ã¯ãã®åéã§ã¯æèœã§ã¯ãããŸãããããã°æèœã§ãã ç§ã¯ãæå·åã®ãã®ãããªã»ãã¥ãªãã£ãšã®ã€ã³ã¿ãŒãã§ãŒã¹ã§åããŠããŸããã 1ã€ã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ãã³ãŒããçæãããã1ã€ã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ãã¯ã©ââããã³ã°ãããšããGoogleã®æ°é®®ãªäœåãèŠãŸããã ããããããã«ããããããããããã®äŸã¯åªããæå·ã¢ã«ãŽãªãºã ãšã¯ã»ã©é ãããã§ãã ç§ã¯ããããã·ãªãŒãºã®ã¬ãã«ã§ã®ç 究ã§ããããã«æããŸãã æ·±å»ãªæå·ãç Žãã¯ãŒã«ãªä»äºã«ã€ããŠèããããšããªãã
ãã®ã¬ããŒãã¯ãé«è² è·ã·ã¹ãã HighLoad ++ã®éçºè ã®äŒè°ã§è¡ãããæé«ã®ã¹ããŒãã®1ã€ã®è»¢åã§ã ã HighLoad ++ 2017ã«ã³ãã¡ã¬ã³ã¹ã®1ãææªæºãæ®ããŸãã
ç§ãã¡ã¯ãã§ã«äŒè°ããã°ã©ã ãæºåããŠãããçŸåšã¹ã±ãžã¥ãŒã«ãç©æ¥µçã«åœ¢æãããŠããŸãã
ä»å¹Žã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®ãããã¯ãåŒãç¶ã調æ»ããŸãã
- å±æ§ã®å®çŸ©ãšè¡£æã®UGCåçã®èŠèŠçæ€çŽ¢ / Dmitry Soloviev
- 深局åŠç¿ã䜿çšããè¡æç»åäžã®é²ãšåœ±ã®èªè / Anatoly Filin
- ãã¥ãŒã©ã«ãããã¯ãŒã¯ïŒTensorRTïŒãã¢ïŒ / Dmitry Korobchenko ã䜿çšããGPUã§ã®é«éæšè«
- ãªãŒããšã³ã³ãŒããŒã䜿çšããæç³»åã®ç°åžžã®æ€åº / Pavel Filonov
- é¡èªèïŒã¹ã¯ã©ããããããããŸã§ / Eduard Tyantov
ãŸãããããã®è³æã®äžéšã¯ãé«è² è·ã·ã¹ãã HighLoadã®éçºã«é¢ãããªã³ã©ã€ã³ãã¬ãŒãã³ã°ã³ãŒã¹ã§äœ¿çšãããŸãã¬ã€ãã¯ãç¹å¥ã«éžæãããæåãèšäºãè³æããããªã®ãã§ãŒã³ã§ãã ç§ãã¡ã®æç§æžã«ã¯ãã§ã«30以äžã®ãŠããŒã¯ãªè³æããããŸãã æ¥ç¶ããŠãã ããïŒ