ç¥çµççåŠããã³ã³ãã¥ãŒã¿ãŒããžã§ã³ãžã®ç§»è¡
ã¹ããŒãªãŒã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ïŒäººå·¥çãªãã®ã ãã§ãªãïŒã®åéã®å é§è ãšãã®è²¢ç®ããå§ããŠãã ããã ããã«ããã¯ããããã¥ãŒãã³ã®æ£åŒãªã¢ãã«ã ããã®åŠç¿çè« ã ããŒãŒã³ãã©ããã®ããŒã»ãããã³ ã ããŒã«ããããªã¿ã®å®éšãªã©ãç¬ç«ããä»äºã®ããã ã ããã1981幎ã®ããŒãã«è³åè³è ã§ãããã€ãããã»ããŒãã«ãšããŒã¹ãã³ã»ãŽã£ãŒãŒã«ã«ããã«è¡ãããšãææ¡ããŸãã 圌ãã¯1959幎ã«è¡ãããä»äºã«å¯ŸããŠè³ãåè³ããŸããïŒåæã«ãRosenblattã¯å®éšãè¡ããŸããïŒã æ£åŒã«ã¯ããã®è³ã¯ãç¥çµæ§é ã«ãããæ å ±åŠçã®åçãšè³æŽ»åã®ã¡ã«ããºã ã«é¢ããç 究ãã«å¯ŸããŠæäžãããŸããã ææãªèªè ã¯ããã«æ³šæããå¿ èŠããããŸãã ç«ã®å®éšã«ã€ããŠã¯ä»¥äžã§èª¬æããŸãã å®éšã¢ãã«ã次ã®å³ã«ç€ºããŸããããŸããŸãªè§åºŠã§ãæãç»é¢ã§æãã现é·ãé·æ¹åœ¢ãç«ã«è¡šç€ºãããŸãã ãªã·ãã¹ã³ãŒãã®é»æ¥µã¯ãåºä¹³é¡ã®èŠèŠæ å ±åŠçã»ã³ã¿ãŒãããè³ã®åŸé éšã«æ¥ç¶ãããŠããŸãã å®éšäžãç§åŠè ã¯æ¬¡ã®å¹æã芳å¯ããŸããïŒææ°ã®ç³ã¿èŸŒã¿ãããã¯ãŒã¯ãšãªã«ã¬ã³ããããã¯ãŒã¯ãšã®é¡äŒŒç¹ãç°¡åã«èŠã€ããããšãã§ããŸãïŒã
- èŠèŠç®è³ªã®ç¹å®ã®é åã¯ãç·ã網èã®ç¹å®ã®éšåã«æ圱ãããå Žåã«ã®ã¿ã¢ã¯ãã£ãã«ãªããŸãã
- é·æ¹åœ¢ã®è§åºŠãå€ãããšãé åã®ãã¥ãŒãã³ã®æŽ»åã¬ãã«ãå€ãããŸãã
- äžéšã®é åã¯ããªããžã§ã¯ããç¹å®ã®æ¹åã«ç§»åãããšãã«ã®ã¿ã¢ã¯ãã£ãã«ãªããŸãã
ãã®èª¿æ»ã®çµæã®1ã€ã¯ã次ã®ç¹æ§ãæã€èŠèŠã·ã¹ãã ãŸãã¯å°åœ¢å³ã®ã¢ãã«ã§ããã
- é£æ¥ãã¥ãŒãã³ã¯ã網èã®é£æ¥é åããã®ä¿¡å·ãåŠçããŸãã
- ãã¥ãŒãã³ã¯éå±€æ§é ïŒäžã®ç»åïŒã圢æãã次ã®ã¬ãã«ããšã«ãŸããŸãé«ã¬ãã«ã®å åã匷調衚瀺ãããŸãïŒä»æ¥ã ãããã®å åãå¹æçã«æäœããæ¹æ³ã¯æ¢ã«ããã£ãŠããŸã ïŒã
- ãã¥ãŒãã³ã¯ããããåã«ç·šæãããŸã-æ å ±ãã¬ãã«ããã¬ãã«ã«å€æããã³éä¿¡ããèšç®ãŠãããã
ããŒãã«ãšãŽã£ãŒãŒã«ã®ã¢ã€ãã¢ãããã°ã©ã ã³ãŒãã«å€æããããšããæåã®è©Šã¿ã¯ã1975幎ãã1980幎ã®éã«ã³ã°ãããã³ãšããªã³ã°ãããã³ã® 2ã€ã®ã¢ãã«ãææ¡ããçŠå³¶éŠåœŠã§ããã ãããã®ã¢ãã«ã¯çç©åŠçã¢ãã«ãã»ãŒç¹°ãè¿ããŠãããä»æ¥ã§ã¯åçŽãªã»ã«ã®ç³ã¿èŸŒã¿ ãããã³è€éãªã»ã«ã®ããŒã«ãšåŒã°ããŠããŸãããããã¯äŸç¶ãšããŠçŸä»£ã®ç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®äž»èŠãªæ§æèŠçŽ ã§ãã ã¢ãã«ã¯ããšã©ãŒã®éäŒæã¢ã«ãŽãªãºã ã«ãã£ãŠã§ã¯ãªããæåž«ãªãã®ã¢ãŒãã§ã®å ã®ãã¥ãŒãªã¹ãã£ãã¯ã¢ã«ãŽãªãºã ã«ãã£ãŠãã¬ãŒãã³ã°ãããŸããã ãã®äœæ¥ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®ã³ã³ãã¥ãŒã¿ãŒããžã§ã³ã®å§ãŸãã ã£ããšæšæž¬ã§ããŸãã
ææžèªèã«é©çšãããåŸé ããŒã¹ã®åŠç¿ ïŒ1998ïŒ
æ°å¹ŽåŸã1998幎ãæ¥ãŸããã å¬ãéããŸããã ãšã©ãŒéäŒæã¢ã«ãŽãªãºã ã«é¢ããèšäºã®èè ã®1人ã®ãã¹ãã¯ã§ããJan LeCunn㯠ãïŒãã¥ãŒã©ã«ãããã¯ãŒã¯ã®ä»ã®èå人ãšååããŠïŒäœåãå ¬éããŸããããã¯ãç³ã¿èŸŒã¿ãšããŒãªã³ã°ã®ã¢ã€ãã¢ãbackpropãšçµã¿åãããæçµçã«æåã®åäœããç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ãååŸããŸãã ã€ã³ããã¯ã¹èªèã®ããã«ç±³åœã®ã¡ãŒã«ã«åã蟌ãŸããŸããã ãã®ã¢ãŒããã¯ãã£ã¯ãæè¿ãŸã§ç³ã¿èŸŒã¿ãããã¯ãŒã¯ãæ§ç¯ããããã®æšæºãã³ãã¬ãŒãã§ãããç³ã¿èŸŒã¿ã¯ãæ°åããŒã«ããããã®åŸãå®å šã«æ¥ç¶ãããè€æ°ã®å±€ã§äº€äºã«è¡ãããŸãã ãã®ãããªãããã¯ãŒã¯ã¯6äžåã®ãã©ã¡ãŒã¿ãŒã§æ§æãããŠããŸãã äž»ãªæ§æèŠçŽ ã¯ãã·ãã1ã®5Ã5ã³ã³ããªã¥ãŒã·ã§ã³ãšã·ãã2ã®2Ã2ããŒã«ã§ãããã§ã«ãåç¥ã®ããã«ãã³ã³ããªã¥ãŒã·ã§ã³ã¯ç¹åŸŽæ€åºåšã®åœ¹å²ãæãããããŒãªã³ã°ïŒãŸãã¯ãµããµã³ããªã³ã°ïŒã䜿çšããŠæ¬¡å ãæžããã ç»åã«ããããã£ããããšããäºå®ã掻çšããŸããã¯ã»ã«ã®å±æçžé¢ -ååãšããŠãé£æ¥ãããã¯ã»ã«ã¯äºãã«å€§ããç°ãªããŸããã ãããã£ãŠãããã€ãã®é£æ¥ãããã®ããéèšãååŸãããå Žåãæ å ±ã®æ倱ã¯ãããã§ãã
ãã£ãŒãã³ã³ããªã¥ãŒã·ã§ãã«ãã¥ãŒã©ã«ãããã¯ãŒã¯ã䜿çšããImageNetåé¡ ïŒ2012ïŒ
ãã14幎ãçµã¡ãŸããã LeCunããã¹ãããã¯ã§ãã£ãåãç 究æã®Alex Krizhevskyã¯ãæåŸã®æåãåŠæ¹ã«è¿œå ããŸããã ãã£ãŒãã©ãŒãã³ã°=ã¢ãã«+åŠç¿çè«+ããã°ããŒã¿+ ããŒããŠã§ã¢ ã GPUã¯ããã¬ãŒãã³ã°ããããã©ã¡ãŒã¿ãŒã®æ°ãå€§å¹ ã«å¢ãããŸããã ãã®ã¢ãã«ã«ã¯ã3æ¡ä»¥äžã®6,000äžåã®ãã©ã¡ãŒã¿ãŒãå«ãŸããŠããã2ã€ã®ã°ã©ãã£ãã¯ã¢ã¯ã»ã©ã¬ãŒã¿ãŒã䜿çšããŠãã®ãããªã¢ãã«ããã¬ãŒãã³ã°ããŸããã
ãã®ä»ã®ç»åAlexNet
GPUããŒã¿äº€æ
ã¬ã€ã€ãŒãµã€ãº
ã¬ã€ã€ãŒãµã€ãº
ãããã¯ãŒã¯ããããžã®èŠ³ç¹ããèŠããšãããã¯ã»ãŒåãLeNetã§ããã1000åã®å€§ããã§ãã ããã«ããã€ãã®ç³ã¿èŸŒã¿å±€ãè¿œå ãããç³ã¿èŸŒã¿ã«ãŒãã«ã®ãµã€ãºã¯ãããã¯ãŒã¯ã®å ¥åããåºåã«åãã£ãŠæžå°ããŸãã ããã¯ãæåã¯ãã¯ã»ã«ãé«åºŠã«çžé¢ããŠãããå容äœé åãå®å šã«å€§ããåãããšãã§ãããšããäºå®ã«ãããã®ã§ããããŸã å°ãã®æ å ±ã倱ãããŠããŸãã 次ã«ãããŒãªã³ã°ã䜿çšããŠãç¡çžé¢é åã®å¯åºŠãå¢ãããŸã ã 次ã®ã¬ãã«ã§ã¯ããããã«å°ããå容äœé åãåãããšãè«ççã§ãã ãã®çµæãèè ã¯ãã®ãããªç³ã¿èŸŒã¿ã®ãã©ããããååŸããŸãã11Ã11-> 5Ã5-> 3Ã3 ...
åãã¬ãŒãã³ã°ãåé¿ããããã«ä»ã®ããªãã¯ã䜿çšãããŸãããããããã®äžéšã¯çŸåšããã£ãŒããããã¯ãŒã¯ã®æšæºã§ããDropOut ïŒRIPïŒãData Augmentationã ReLuã§ãã ãããã®ããªãã¯ã«çŠç¹ãåœãŠãã®ã§ã¯ãªããã¢ãã«ã®ããããžãŒã«çŠç¹ãåœãŠãŸãã 2012幎以éãéãã¥ãŒã©ã«ãããã¯ãŒã¯ã¢ãã«ãç»åãç²åŸã§ããªããªã£ãããšã ããè¿œå ããŸãã
倧èŠæš¡ç»åèªèçšã®éåžžã«æ·±ãç³ã¿èŸŒã¿ãããã¯ãŒã¯ ïŒ2014幎4æ12æ¥ïŒ
ä»å¹Ž2ã€ã®èå³æ·±ãèšäºãçºè¡šãããŸããããã®èšäºãšGoogle Inceptionã§ããããã«ã€ããŠã¯ä»¥äžã§èª¬æããŸãã ãªãã¯ã¹ãã©ãŒãç 究æã®ç 究ã¯ãLeCunãå®ããããããžãŒã®ãã¿ãŒã³ã«æºæ ããæåŸã®ç 究ã§ãã VGG-19ã¢ãã«ã¯1å4400äžã®ãã©ã¡ãŒã¿ãŒã§æ§æããã8400äžã®ãã©ã¡ãŒã¿ãŒã«å ããŠãã¢ãŒããã¯ãã£ã«å¥ã®ç°¡åãªã¢ã€ãã¢ãè¿œå ããŸãã ããšãã°ã5Ã5ã®ç³ã¿èŸŒã¿ããã®ãããã³ã° ã25åã®ãã©ã¡ãŒã¿ãŒãå«ãŸããŠããŸãã ç³ã¿èŸŒã¿ã3Ã3ã®2ã€ã®ã¬ã€ã€ãŒã®ã¹ã¿ãã¯ã«çœ®ãæãããšãåã衚瀺ã«ãªããŸããããã©ã¡ãŒã¿ãŒã®æ°ã¯3Ã3 + 3Ã3 = 18ãšãªããããã¯22ïŒ å°ãªããªããŸãã 11Ã11ã4ã€ã®3Ã3ç³ã¿èŸŒã¿ã«çœ®ãæãããšãããã¯ãã§ã«ãã©ã¡ãŒã¿ãŒããã70ïŒ å°ãªããªããŸãã
VGG- *ã¢ãã«
ãããã¯ãŒã¯å ãããã¯ãŒã¯ ïŒ2014幎3æ4æ¥ïŒ
ãããŠãããããã¿ã³ããªã³ãšèžãå§ããŸãã ãã®ãããªåçã§ããã®åºçç©ã®èè ã¯åœŒã®ããã°æçš¿ãäŸèšŒããããã§ã¯åœŒãascadedrosshannelParameteric poolingã«ã€ããŠè©±ããŸããã ãã®èšäºã®äž»ãªã¢ã€ãã¢ãæ€èšããŠãã ããã æããã«ãç³ã¿èŸŒã¿æŒç®ã¯ç»åãããã®ç·åœ¢å€æã§ãããç³ã¿èŸŒã¿å±€ã¯äžè¬åç·åœ¢ã¢ãã«ïŒGLMïŒã§ãã ç»åã¯ç·åœ¢ã«åé¢å¯èœã§ãããšæ³å®ãããŠããŸãã ãããããªãCNNãæ©èœããã®ã§ããããïŒ ãã¹ãŠãåçŽã§ããéå°ãªè¡šçŸïŒå€æ°ã®ãã£ã«ã¿ãŒïŒã䜿çšããŠãå±æ§ç©ºéå ã®1ã€ã®ç»åã®ãã¹ãŠã®ããªãšãŒã·ã§ã³ãèæ ®ããŸãã 1ã€ã®ã¬ã€ã€ãŒã§ãã£ã«ã¿ãŒãå€ãã»ã©ã次ã®ã¬ã€ã€ãŒã§èæ ®ããå¿ èŠãããããªãšãŒã·ã§ã³ãå¢ããŸãã ã©ããã GLMãããå¹ççãªãã®ã«çœ®ãæããŸãããã æãå¹æçãªã®ã¯åå±€ããŒã»ãããã³ã§ãã ããã«ãããæšèã®ã¹ããŒã¹ãããå¹æçã«åå²ããæšèã®æ°ãæžããããšãã§ããŸãã
ããªãã¯ãã§ã«ãã®ãããªæ±ºå®ã®åŠ¥åœæ§ãçãå§ããŠãããšç¢ºä¿¡ããŠããŸããç§ãã¡ã¯ãµã€ã³ã®æ°ãæ¬åœã«æžãããŸããããã©ã¡ãŒã¿ãŒã®æ°ãå€§å¹ ã«å¢ãããŸãã èè ã¯ããœé£ããã®åé¡ã解決ãããšè¿°ã¹ãŠããŸãã ããã¯æ¬¡ã®ããã«æ©èœããŸãïŒç³ã¿èŸŒã¿å±€ã®åºå£ã§ããµã€ãºWÃHÃDã®ç«æ¹äœãååŸããŸããåäœçœ®ïŒwãhïŒã«ã€ããŠãDãã¯ãã¹ãã£ã³ãã«ã®ãã¹ãŠã®å€ãååŸããç·åœ¢çµåããã©ã¡ããªãã¯ããŒãªã³ã°ãèšç®ããŸããåãããã«ãã£ãŠæ°ããããªã¥ãŒã ãäœæããæ°å-ã«ã¹ã±ãŒãã ãããã¯ã1Ã1ã®ããã¿èŸŒã¿ãšããã«ç¶ãéç·åœ¢æ§ã§ããããšãããããŸãã ãã®ããªãã¯ã¯ãçæ³çã«ã¯3Ã3ã³ã³ããªã¥ãŒã·ã§ã³ã«é¢ããVGGã®èšäºãšåãã§ãããæåã¯éåžžã®ã³ã³ããªã¥ãŒã·ã§ã³ãMLPã«çœ®ãæããããšãææ¡ããŸãããMLPã¯éåžžã«é«äŸ¡ãªã®ã§ããã®å®å šã«æ¥ç¶ãããåã¬ã€ã€ãŒãç³ã¿èŸŒã¿ã¬ã€ã€ãŒã«çœ®ãæããŸã-ã·ã³ã¬ããŒã«ã®ããªãã¯ã NINã¯æ·±ãç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ã§ãããç³ã¿èŸŒã¿ã®ä»£ããã«ã1Ã1ã³ã¢ã®å°ããªç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ãããããšãããããŸããã
ããã«èå³æ·±ã次ã®ã¢ã€ãã¢ã¯ãå®å šã«æ¥ç¶ãããã¬ã€ã€ãŒãå®å šã«æŸæ£ããããšã§ã-ã°ããŒãã«å¹³åããŒãªã³ã°ã ãã®ãããªãããã¯ãŒã¯ã¯ãç¹å®ã®ç»åãµã€ãºãå ¥åããå¿ èŠããªããç³ã¿èŸŒã¿/ããŒãªã³ã°ã®ã¿ã§æ§æããããããå®å šãªç³ã¿èŸŒã¿ãããã¯ãŒã¯ãšåŒã°ããŸãã Nåã®ã¯ã©ã¹ãžã®åé¡åé¡ãããããã¥ãŒãWÃHÃDãããã¥ãŒãWÃHÃNãäœæããããã«ãå®å šã«æ¥ç¶ãããã¬ã€ã€ãŒã®ä»£ããã«1Ã1ã®ç³ã¿èŸŒã¿ã䜿çšããããšä»®å®ããŸããã€ãŸãã1Ã1ã®ç³ã¿èŸŒã¿ã¯ãã¥ãŒãã®æ·±ããä»»æã«å€æŽã§ããŸããµã€ã³ã Nåã®ãã€ã¹ãWÃHã®å Žåããã€ã¹ã®å¹³åå€ãèšç®ããsoftmaxãèšç®ã§ããŸãã å°ãæ·±ãèãã¯ã以åã®ç³ã¿èŸŒã¿å±€ã¯åã«èšå·ãæœåºããã¡ã«ããºã ãšããŠæ©èœããåŒå¥åšã¯å®å šã«æ¥ç¶ãããå±€ã§ãã£ããšããèè ã®äž»åŒµã«ãããŸãïŒããããLeCunnãæå®ãããã¿ãŒã³ïŒã ãããã¯ãŒã¯ã¯ç³ã¿èŸŒã¿ã®ã¿ã§æ§æãããåé¡åé¡ã解決ãããããèè ã¯ç³ã¿èŸŒã¿ãåŒå¥åã§ãããšäž»åŒµããŠããŸãã LeNet圢åŒã®ç³ã¿èŸŒã¿ãããã¯ãŒã¯ã§ã¯ãèšç®ã®80ïŒ ãç³ã¿èŸŒã¿å±€ã«ãããã¡ã¢ãªæ¶è²»ã®80ïŒ ãå®å šã«æ¥ç¶ãããå±€ã«ããããšã«æ³šæããŠãã ããã
LeCunã¯åŸã«ãã§ã€ã¹ããã¯ã«æ¬¡ã®ããã«æžããŠããŸã ã
ç³ã¿èŸŒã¿ãããã§ã¯ããå®å šã«æ¥ç¶ãããã¬ã€ã€ãŒããªã©ã¯ãããŸããã 1Ã1ã®ç³ã¿èŸŒã¿ã«ãŒãã«ãšå®å šãªæ¥ç¶ããŒãã«ãæã€ç³ã¿èŸŒã¿å±€ã®ã¿ããããŸãã
éåžžã«å€ãã®æ°ããã¢ã€ãã¢ãæãã€ããèè ãç»åã«åå ããªãã£ãããšã¯æ®å¿µã§ãã ããããGoogleã¯ãããã®ã¢ã€ãã¢ãããã«å€§éšããã次ã®åºçç©ã§äœ¿çšããŸããã ãã®çµæãGoogleã¯2014幎ã®åè³å°ããªãã¯ã¹ãã©ãŒãã®ããŒã ãšå ±æããŸããã
ç³ã¿èŸŒã¿ã§ããã«æ·±ã ïŒ2014幎9æ17æ¥ïŒ
Googleã¯ã²ãŒã ã«åå ¥ãã圌ãã¯ãããã¯ãŒã¯Inceptionãšåä»ããŸãããååã¯å¶ç¶ã«éžã°ãããã®ã§ã¯ãªããããããã¯ãŒã¯å éšã®ãããã¯ãŒã¯ããšæåãªââããŒã ã«é¢ããåäœã®ã¢ã€ãã¢ãåŒãç¶ãã§ããããã§ãã èè ãæžãããã®ã¯æ¬¡ã®ãšããã§ãã
ãã®è«æã§ã¯ãæåãªãæã ã¯ããæ·±ãè¡ãå¿ èŠãããããšãšãã«ããªã³ã[12]ã«ãããããã¯ãŒã¯è«æã®Networkãããã®ååãåŸããInceptionãšåŒã°ããã³ã³ãã¥ãŒã¿ãŒããžã§ã³ã®ããã®å¹ççãªãã£ãŒããã¥ãŒã©ã«ãããã¯ãŒã¯ã¢ãŒããã¯ãã£ã«çŠç¹ãåœãŠãŸãã€ã³ã¿ãŒãããããŒã [1]ã ç§ãã¡ã®å Žåããæ·±ãããšããèšèã¯2ã€ã®ç°ãªãæå³ã§äœ¿çšãããŸãããŸãããã€ã³ã»ãã·ã§ã³ã¢ãžã¥ãŒã«ããšãã圢ã§æ°ããã¬ãã«ã®çµç¹ãå°å ¥ãããšããæå³ãšããããã¯ãŒã¯ã®å¢å ãšããããçŽæ¥çãªæå³æ·±ãã äžè¬ã«ãã€ã³ã»ãã·ã§ã³ã¢ãã«ã¯[12]ã®è«ççé倧æã§ãããšèŠãªãããšãã§ããŸãããã¢ãã©ãã®çè«çç 究ããã€ã³ã¹ãã¬ãŒã·ã§ã³ãšã¬ã€ãã³ã¹ãåŸãŠããŸã[2]ã
å¹ ïŒã¬ã€ã€ãŒå ã®ãã¥ãŒãã³ã®æ°ïŒãšæ·±ãïŒã¬ã€ã€ãŒã®æ°ïŒã®ä»»æã®å¢å ã«ã¯ãå€ãã®æ¬ ç¹ããããŸãã 第äžã«ããã©ã¡ãŒã¿ã®æ°ãå¢ãããšåãã¬ãŒãã³ã°ã容æã«ãªããå±€ã®æ°ãå¢ãããšåŸé æžè¡°ã®åé¡ãå¢ããŸãã ãšããã§ãæåŸã®ã¹ããŒãã¡ã³ãã¯ResNetã«ãã£ãŠè§£æ±ºãããŸãããããã«ã€ããŠã¯åŸã§èª¬æããŸãã 第äºã«ãã¬ã€ã€ãŒå ã®ç³ã¿èŸŒã¿æ°ãå¢å ãããšããã®ã¬ã€ã€ãŒã§ã®èšç®ã2次çã«å¢å ããŸãã æ°ããã¢ãã«ã®ãã©ã¡ãŒã¿ãŒãéå¹ççã«äœ¿çšãããŠããå Žåãããšãã°ããããã®å€ãããŒãã«è¿ããªãå ŽåãåçŽã«èšç®èœåã浪費ããŠããŸãã ãããã®åé¡ã«ãé¢ãããããã®èšäºã®æåã®èè ã¯ãæ·±ããããã¯ãŒã¯ã§å®éšããããšãæãã§ããŸãããããã©ã¡ãŒã¿ãŒãå€§å¹ ã«æžãããŸããã ãããè¡ãããã«ã圌ã¯èšäºã ããã€ãã®æ·±ãè¡šçŸãåŠç¿ããããã®èšŒæå¯èœãªå¢ç ãã«ç®ãåããŸãããããã§ã¯ãããŒã¿ã®ç¢ºççååžããŸã°ãã§æ·±ããŠåºããã¥ãŒã©ã«ãããã¯ãŒã¯ãšããŠè¡šçŸã§ããå Žåããã¥ãŒãã³ã®çžé¢ãåæããããšã§ãã®ããŒã¿ã»ããã«æé©ãªãã¥ãŒã©ã«ãããã¯ãŒã¯ãæ§ç¯ã§ããããšã蚌æããŠããŸãåã®å±€ãšçžé¢ãããã¥ãŒãã³ã次ã®å±€ã®ãã¥ãŒãã³ã«ãªãã°ã«ãŒãã«çµåããŸãã ãããã£ãŠãåŸå±€ãã¥ãŒãã³ã¯ãå ã®ç»åã®åŒ·ã亀差ããé åããèŠãŸããã ãã¥ãŒãã³ã®å容äœé åãç°ãªãã¬ãã«ã§ã©ã®ããã«èŠãããããããŠããããæœåºããå åãæãåºãããŠãã ããã
VGG-19ãããã¯ãŒã¯ã®conv1_2å±€ã®å容äœé å
ã¯ããå容é åã¯éåžžã«å°ãããããã»ãšãã©äœã衚瀺ãããŸãããããã¯ãããã2çªç®ã®3Ã3ç³ã¿èŸŒã¿ã§ãããäžè¬çãªé¢ç©ã¯5Ã5ã§ãã
VGG-19ãããã¯ãŒã¯å容äœå±€conv3_3
VGG-19ãããã¯ãŒã¯ã®conv4_3å±€ã®å容äœé å
VGG-19ãããã¯ãŒã¯å容äœå±€conv5_3
VGG-19ãããã¯ãŒã¯pool5å容äœé å
åæå±€ïŒå ¥ãå£ã«è¿ãïŒã§ã¯ãçžé¢ãã¥ãŒãã³ã¯å±æé åã«éäžããŸãã ã€ãŸãããã¬ãŒãäžã®åã座æšïŒwãhïŒã«ããè€æ°ã®ãã¥ãŒãã³ãåãããšãåŠç¿ã§ããå Žåãæåã®å±€ã®åŸã®ãã³ãœã«ã§ã¯ããããã®æŽ»æ§åã¯ç¹å®ã®ç¹ïŒwãhïŒã®é åã®å°ããªé åã«é 眮ãããŸãããååžããŸããã£ã«ã¿ãŒãã³ã¯ã®æ¬¡å ã«æ²¿ã£ãŠé 眮ãããŸã-D.ãã®ãããªãã®ïŒ
ãã®ãããªçžé¢ããã£ããããŠããããã1ã€ã®å åã«å€ããæ¹æ³ã¯ïŒ åäœããã®1Ã1ç³ã¿èŸŒã¿ã®ã¢ã€ãã¢ãå©ãã«ãªããŸãã ãã®èããç¶ãããšãçžé¢ã¯ã©ã¹ã¿ãŒã®æ°ããããã«å°ãªããããšãã°3Ã3ã«ãªããšæ³å®ã§ããŸããããã¯5Ã5ãªã©ã«ãåœãŠã¯ãŸããŸãããGoogleã¯5Ã5ã§åæ¢ããããšã«ããŸããã
ããã¿èŸŒã¿1Ã1
ããã¿èŸŒã¿3Ã3
ããã¿èŸŒã¿5Ã5
åç³ã¿èŸŒã¿ã®ç¹åŸŽããããèšç®ããåŸãããããäœããã®æ¹æ³ã§éçŽããå¿ èŠããããŸãïŒããšãã°ãé£çµãã£ã«ã¿ãŒïŒã
æšèã®é£çµ
ã¡ãªã¿ã«ãåã®ã¬ã€ã€ãŒã®å ã®ç»åã倱ããªãããã«ãç³ã¿èŸŒã¿ã«å ããŠãããŒãªã³ã°ã®å åãè¿œå ãããŠããŸãïŒããŒãªã³ã°æäœã¯æ å ±ã®å€§ããªæ倱ã䌎ããªãããšãèŠããŠããŸãïŒã ãšã«ããã誰ããããŒãªã³ã°ã䜿çšããŸã-ãããæåŠããçç±ã¯ãããŸããã è€æ°ã®æäœã®ã°ã«ãŒãå šäœãInceptionãããã¯ãšåŒã³ãŸãã
ããã«ãçŸåšã®æå 端ã®ç³ã¿èŸŒã¿ãããã¯ãŒã¯ã®æåã«ã¯ããŒãªã³ã°æäœãäžå¯æ¬ ã§ããããããã®ãããªå段éã§ä»£æ¿ã®äžŠåããŒãªã³ã°ãã¹ãè¿œå ããããšãæçãªå¹æãããããšã瀺åããŠããŸãã
åºåãã³ãœã«ã®ãµã€ãºãåäžã«ããããã«ã1Ã1ã®ç³ã¿èŸŒã¿ã䜿çšããããšãææ¡ãããŠããŸããããã¯ããã«æäœã®çŽåŸã®å³ã§èŠãããšãã§ããŸãã ããã«ã1Ã1ã®ç³ã¿èŸŒã¿ã¯ããšãã«ã®ãŒéçŽçãªç³ã¿èŸŒã¿æäœã®åã«æ¬¡å ãåæžããããã«ã䜿çšãããŸãã ã芧ã®ãšãããGoogleã¯ãã®ãããªç®æšãéæããããã«1Ã1ã®ç³ã¿èŸŒã¿ã䜿çšããŠããŸãïŒæ¬¡äžä»£ãããã¯ãŒã¯ããããã®ææ³ã掻çšããŸãïŒã
- 寞æ³ã®å¢å ïŒæè¡åŸïŒ;
- 次å åæžïŒæè¡åïŒ;
- çžé¢å€ã®ã°ã«ãŒãåïŒãããã¯å ã®æåã®æäœïŒã
æçµçãªã¢ãã«ã¯æ¬¡ã®ãšããã§ãã
äžèšã«å ããŠãç°ãªãã¬ãã«ã®ããã€ãã®è¿œå åé¡åã«ãæ°ä»ãããšãã§ããŸãã å ã®ã¢ã€ãã¢ã¯ããã®ãããªåé¡åšãåæã®å±€ã«åŸé ããæŒãåºããããšãå¯èœã«ããããã«ãã£ãŠåŸé æžè¡°ã®åœ±é¿ãæžãããšãããã®ã§ããã Googleã¯ããããåŸã§æåŠããŸãã ãããã¯ãŒã¯èªäœããã³ãœã«ã®æ·±ããåŸã ã«å¢ããã空é次å ãæžãããŸãã èšäºã®çµããã«ãèè ã¯ãã®ãããªã¢ãã«ã®æå¹æ§ã®åé¡ãæªè§£æ±ºã®ãŸãŸã«ããŠããããŸããäžèšã®ååã䜿çšããŠãããã¯ãŒã¯ããããžãèªåçã«çæããå¯èœæ§ãæ¢ãããšã瀺åããŠããŸãã
ã³ã³ãã¥ãŒã¿ãŒããžã§ã³ã®ã€ã³ã»ãã·ã§ã³ã¢ãŒããã¯ãã£ã®åè ïŒ2015幎12æ11æ¥ïŒ
1幎åŸãã°ãŒã°ã«ã¯ãã®ç»åã«å¯ŸããŠååã«æºåãæŽããæè«ã®æ§é ãåèããããå®å šã«æ°ããã¢ãã«ã倱ããŸãããããã«ã€ããŠã¯æ¬¡ã®ããŒãã§èª¬æããŸãã æ°ããèšäºã§ã¯ãèè ã¯å®éã«ããŸããŸãªã¢ãŒããã¯ãã£ãæ€èšããã³ã³ãã¥ãŒã¿ãŒããžã§ã³çšã®æ·±ãç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ãæ§ç¯ããããã®4ã€ã®ååãéçºããŸããã
- è¡šçŸã®ããã«ããã¯ãåé¿ããŸããããŒã¿è¡šç€ºã®æ¬¡å ãå€§å¹ ã«æžããã¹ãã§ã¯ãããŸãããããã¯ããããã¯ãŒã¯ã®æåããåºååé¡åãŸã§ã¹ã ãŒãºã«è¡ãå¿ èŠããããŸãã
- é«æ¬¡å ã®è¡šçŸã¯ããŒã«ã«ã§åŠçãã次å ãå¢ãããŸãã次å ãã¹ã ãŒãºã«æžããã ãã§ã¯ååã§ã¯ãããŸãããåã®èšäºã§èª¬æããååã䜿çšããŠçžé¢ã»ã¯ã·ã§ã³ãåæããã³ã°ã«ãŒãåãã䟡å€ããããŸãã
- 空é調æŽã¯ãããã«å°ããªãã®ã«å解ããããšãã§ããŸãããããããå¿ èŠããããŸããããã«ããããªãœãŒã¹ãç¯çŽãããããã¯ãŒã¯ã®ãµã€ãºãå¢ããããšãã§ããŸãã
- ãããã¯ãŒã¯ã®æ·±ããšå¹ ã®ãã©ã³ã¹ãç¶æããå¿ èŠããããŸãããããã¯ãŒã¯ã®æ·±ããå¹ ãšã¯å¥ã«æ¥æ¿ã«å¢å ããããããã®éãè¡ã£ããããªãã§ãã ããã äž¡æ¹ã®å¯žæ³ãåäžã«å¢æžããå¿ èŠããããŸãã
倧ããªç³ã¿èŸŒã¿ã3Ã3ç³ã¿èŸŒã¿ã¹ã¿ãã¯ã«å解ã§ãããšãã2014 VGGã®ã¢ã€ãã¢ãèŠããŠããŸããïŒ ãããã£ãŠãã°ãŒã°ã«ã¯ããã«é²ãã§ããã¹ãŠã®ç³ã¿èŸŒã¿ãNÃ1ããã³1ÃNã§å æ°å解ããŸããã
ã€ã³ã»ãã·ã§ã³ãããã¯ã¢ãã«
æåã¯ãªãªãžãã«ã§ãã
VGGã®åçã«ãã£ãŠæåã«å æ°å解ãããŸããã
æ°ãããããã¯ã¢ãã«ã
VGGã®åçã«ãã£ãŠæåã«å æ°å解ãããŸããã
æ°ãããããã¯ã¢ãã«ã
ããã«ããã¯ã«é¢ããŠã¯ã次ã®ã¹ããŒã ãææ¡ãããŠããŸãã å ¥å次å ã ãååŸããã ããã®åŸãæåã«ç³ã¿èŸŒã¿ãé©çšããŸã 1ã®ã¹ãããã§ãã£ã«ã¿ãªã³ã°ããŠãããã±ãŒã¹ããŒãªã³ã°ã ãã®ãããªæäœã®å šäœçãªè€éã ã æåã«åŒã£åŒµãã次ã«ç³ã¿èŸŒã¿ãè¡ããšãè€éããäœäžããŸã ãããããã®åŸãæåã®ååã«éåããŸãã 䞊ååå²ã®æ°ãå¢ããã2ã®ã¹ãããã§ç³ã¿èŸŒã¿ãè¡ãããšãææ¡ãããŠããŸãããåæã«ãã£ãã«æ°ã2åã«ãããšã代衚衚çŸåã¯ãã¹ã ãŒã¶ãŒããæžå°ãããŸãã ãããŠããã³ãœã«ã®æ·±ããæäœããããã«ã1Ã1ã®ç³ã¿èŸŒã¿ã䜿çšãããŸãã
æ°ããã¢ãã«ã¯Inception V2ãšåŒã°ãã ãããæ£èŠåãè¿œå ãããšãInception V3ããããŸãã ã¡ãªã¿ã«ãè¿œå ã®åé¡åã¯åé€ãããŸããããããã¯ç¹ã«å質ãåäžãããããã§ã¯ãªãããæ£èŠååãšããŠæ©èœã§ããããšãå€æããããã§ãã ãããããã®æç¹ãŸã§ã«ãæ£èŠåããããã®ããèå³æ·±ãæ¹æ³ããã§ã«ãããŸããã
ç»åèªèã®ããã®æ·±å±€åŠç¿ ïŒ2015幎12æ10æ¥ïŒ
2015幎ã«Googleã倱ã£ãMicrosoft Researchã®äžåœéšéã®ä»äºãç 究ããæãæ¥ãŸããã åã«ãã£ãšå€ãã®ã¬ã€ã€ãŒããã¬ãŒã³ãããšããã®ãããªã¢ãã«ã®å質ã¯äžå®ã®éçãŸã§äžæãïŒVGG-19ãåç §ïŒããã®åŸäœäžãå§ããããšãé·ãéæ°ã¥ããŠããŸããã ãã®åé¡ã¯å£ååé¡ãšåŒã°ããããå€ãã®å±€ããã¬ãŒã³ããããšã«ãã£ãŠåŸããããããã¯ãŒã¯ã¯ãã¬ãŒã³ãªããŸãã¯ãã©ãããªãããã¯ãŒã¯ã§ãã èè ã¯ãæ°ããã¬ã€ã€ãŒãè¿œå ããããšã§ã¢ãã«ã®å質ãåäžããããããžãèŠã€ããããšãã§ããŸããã
ããã¯ãäžèŠãåçŽãªããªãã¯ã«ãã£ãŠéæãããŸãããäºæããªãæ°åŠççµæã«ã€ãªãããŸãã ã¢ãŒãã«ããšã³ã«ã¢ãŽããã®ãããã§ãç§ãã¡ãšåœŒãã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ãã»ãŒãã¹ãŠã®é¢æ°ãããšãã°ãããã€ãã®è€éãªé¢æ°ãè¿äŒŒã§ããããšãèªèããŠããŸãã ã ãããŠããã®ãããªãããã¯ãŒã¯ãæ®å·®é¢æ°ãç°¡åã«åŠç¿ã§ããããšã¯äºå®ã§ãïŒãã·ã¢èªã§ã¯æ®å·®é¢æ°ãšåŒã³ãŸãïŒã ã æããã«ãæåã®ç®çé¢æ°ã¯ ã ãããã¯ãŒã¯ãããšãã°VGG-19ã䜿çšããããã«20å±€ãè¿œå ããå Žåã深局ãããã¯ãŒã¯ãå°ãªããšããã®æµ å±€ãããã¯ãŒã¯ãããæªããªããªãããã«ããŸãã å£åã®åé¡ã¯ãè€éãªéç·åœ¢é¢æ° ããã€ãã®ã¬ã€ã€ãŒããã¬ãŒã³ããããšã§åŸããããã®ã¯ã以åã®ã¬ã€ã€ãŒã§å質ã®éçã«éããå Žåã«æçå€æãåŠç¿ããã¯ãã§ãã ããããããã¯äœããã®çç±ã§çºçããŸãããããããããªããã£ãã€ã¶ã¯åçŽã«éã¿ã®èª¿æŽã«å¯Ÿå¿ã§ããªããããè€éãªéç·åœ¢éå±€ã¢ãã«ãåäžã®å€æãè¡ããŸãã ããããã·ã§ãŒãã«ããæ¥ç¶ãè¿œå ããã®ãæäŒãã°ããªããã£ãã€ã¶ãŒããã¹ãŠã®éã¿ããŒãã«è¿ã¥ããããšã¯ãåäžã®å€æãäœæãããããç°¡åãããããŸããã 確ãã«ãããã¯ã¢ã€ãã¢ãåŸæŒãããããšãæãåºãããŸãã
ãã©ãããããã¯ãŒã¯ãæ®äœãããã¯ãŒã¯ã«å€æããäŸ
ãã¢ã/ã©ã¶ãã¢ã§ã®ResNet-52ã®æ§ç¯ã®è©³çŽ°ãããã§èª¬æããŸã ãããã§ã¯ãGoogleãéºèšããŠããããã«ã1Ã1ã®ç³ã¿èŸŒã¿ã次å ã®å¢æžã«äœ¿çšãããããšã®ã¿ã«èšåããŸãã ãããã¯ãŒã¯ã¯å®å šã«ç³ã¿èŸŒã¿ã§ãã
æ®å·®é¢æ°ããŒãã«è¿ããšãã仮説ã®æ£åœåã®1ã€ãšããŠãèè ã¯ãå±€ã«å¿ããŠæ®å·®é¢æ°ã®æŽ»æ§åã®ã°ã©ãã瀺ãããã©ãããããã¯ãŒã¯ã®å±€ã®æŽ»æ§åãæ¯èŒããŸãã
ãããŠãããããã¹ãŠããã®çµè«ã¯ããã次ã®ãšããã§ãã
ResNet'ovãšä»¥åã®ãã¹ãŠã®ãããã¯ãŒã¯ã®æ·±ããæ¯èŒãã
èè ã®å¥ã®èå³æ·±ãã¹ã©ã€ãã¯ã圌ãããããã¯ãŒã¯ã®æ·±ã/å質ã«çªç Žå£ãéããããšã瀺ããŠããŸãã , , , 19- VGG, 152 .
, . , 2016 , . .
ResNet:
Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning (23 Feb 2016)
, , Inception, ResNet'. , , , . , , . Inception V4 Inception ResNet.
Inception V4 , Inception-. , , 4 , , â , .
Inception V4
äžè¬çãªãããã¯ãŒã¯å³ã
ã¹ãã ãããã¯ã
ã€ã³ã»ãã·ã§ã³ãããã¯A
ã¹ãã ãããã¯ã
ã€ã³ã»ãã·ã§ã³ãããã¯A
Inception V4 Inception- , shortcut-, Inception ResNet.
, , , .
Identity Mappings in Deep Residual Networks (12 Apr 2016)
ResNet, , : . , . . , ; , . , 1001 . , 2014 19- «very deep convolutional neural networks».
. ããã
- â ;
- â shortcut connection, ;
- â residual network;
- â ;
- â .
, â , . K , . , . - k , , : . , - . , , vanishing gradients, , â1. , . , .
, - ? , ã . , . , . .
residual- , , , .
â ( JÃŒrgen Schmidhuber ). LSTM- , 1997 , , ResNet â , LSTM, 1. , â Highway Network . , , , (gate), , .
Bridging the Gaps Between Residual Learning, Recurrent Neural Networks and Visual Cortex (13 Apr 2016)
, ResNet RNN. . , , , , , , .
The dark secret of Deep Networks: trying to imitate Recurrent Shallow Networks?
A radical conjecture would be: the effectiveness of most of the deep feedforward neural networks, including but not limited to ResNet, can be attributed to their ability to approximate recurrent computations that are prevalent in most tasks with larger t than shallow feedforward networks. This may offer a new perspective on the theoretical pursuit of the long-standing question âwhy is deep better than shallowâ.
Residual Networks are Exponential Ensembles of Relatively Shallow Networks (20 May 2016)
, DropOut, , . , . . , ResNet â . . ResNet, , shortcut- flowing graph'.
:
, , . , , multiplicity, -, , â . , , (). , , , (). , , .
, . , , . , ResNet .
, ResNet . , , .
, VGG ResNet .
çµæã¯çŽ æŽãããã§ãã
次ã®å®éšã§ã¯ãèè ã¯ç°ãªãé·ãã®ãµããããããååŸããåŸé ã®çµ¶å¯Ÿå€ã®ååžãæ§ç¯ããŸãã ResNet-50ã®æå¹ãªãµããããé·ã¯5ã17å±€ã§ããããã¹ãŠã®ãã¹é·ã®45ïŒ ããå æããŠããŸããããã¢ãã«å šäœã®ãã¬ãŒãã³ã°ãžã®ã»ãŒãã¹ãŠã®å¯äžãå ããŠããŸãã
åŸé
ååžåæã
ãããã«
ResNetã¯é²åãç¶ããŠãããããŸããŸãªã°ã«ãŒããæ°ããããšã«ææŠããéå»ã®å®çžŸã®ãããã¿ãŒã³ãé©çšããŠããŸãã ããšãã°ãResNetã§ResNetãäœæããŸãã
- ç³ã¿èŸŒã¿æ®çã¡ã¢ãªãããã¯ãŒã¯ ïŒ2016幎7æ14æ¥ïŒ
- åºãæ®äœãããã¯ãŒã¯ ïŒ2016幎5æ23æ¥ïŒ
- 深床ããããã¢ãŠãïŒæ®çç³ã¿èŸŒã¿ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®å¹ççãªãã¬ãŒãã³ã°
- Resnetã®ResnetïŒæ®çã¢ãŒããã¯ãã£ã®äžè¬å ïŒ2016幎3æ25æ¥ïŒ
- æ®çãããã¯ãŒã¯ã®æ®çãããã¯ãŒã¯ïŒãã«ãã¬ãã«ã®æ®çãããã¯ãŒã¯ ïŒ2016幎8æ9æ¥ïŒ
- è€æ°ã®æ®çãããã¯ãŒã¯ ïŒ2016幎9æ24æ¥ïŒ
- åé¢å¯èœãªç³ã¿èŸŒã¿ã«ãããã£ãŒãã©ãŒãã³ã° ïŒ2016幎10æ7æ¥ïŒ
- æ·±ããã©ãããåæ®å·®ãããã¯ãŒã¯ ïŒ2016幎10æ10æ¥ïŒ
ãã®ã¬ãã¥ãŒã§èšåãããŠããªããã®ãèªãã å Žåã¯ãã³ã¡ã³ãã§æããŠãã ããã ãã®ãããªæ°ããèšäºã®ã¹ããªãŒã ã§ãã¹ãŠã远跡ããããšã¯ãåã«çŸå®çã§ã¯ãããŸããã ãã®æçš¿ã¯ããã°ã§ç®¡çãããŠããŸãããããªã¹ãã«ããã«2ã€ã®æ°ããåºçç©ãè¿œå ããããšãã§ããŸããã 1ã€ã¯ã°ãŒã°ã«ãæè«ã®æ°ããããŒãžã§ã³ã§ããXceptionã«ã€ããŠããã1ã€ã¯ãã©ãããåResNetã«ã€ããŠã§ãã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã«ã€ããŠå°ãç¥ã£ãŠãã人ã«ããããã¯ãŒã¯ã®çš®é¡ãå°ãããšã圌ã¯æ£åžžïŒå®å šã«æ¥ç¶ãããç³ã¿èŸŒã¿ïŒã§åçºããŠãããšèšããŸãã ãã¹ãŠã®ã¿ã€ãã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ãåãã¯ã©ã¹ã«å±ããçè«çåºç€ããŸããªãç»å ŽãããšæãããŸãã æ®éçãªè¿äŒŒå®çããããŸããããã¯ãä»»æã®ïŒã»ãŒïŒé¢æ°ããã¥ãŒã©ã«ãããã¯ãŒã¯ã§è¿äŒŒã§ãããšèšããŸãã RNNã®åæ§ã®å®çã¯ãåçã·ã¹ãã ã¯ãªã«ã¬ã³ããããã¯ãŒã¯ã«ãã£ãŠè¿äŒŒã§ãããšè¿°ã¹ãŠããŸãã ãããŠä»ãRNNã¯éåžžã®ãããã¯ãŒã¯ã§è¿äŒŒã§ããããšãããããŸããïŒãã ããããã蚌æããããã«æ®ã£ãŠããŸãïŒã äžè¬çã«ãç§ãã¡ã¯é¢çœãæéã«äœãã§ããŸãã
衚瀺ãããŸãã