¡]¥»¤å§@ªÌ¥ô¾©ó¤u¬ã°|§Þ³N²¾Âà»Pªk«ß¤¤¤ß¡^
²Ä¤T¦¸AI®ö¼é©Ò¿³°_ªº¾÷¾¹¾Ç²ß¦³³\¦hºØ¤âªk¡A³Ì¨üÆf¥Øªº´N¬O²`«×¾Ç²ß¡CµM¦Ó¡A²`«×¾Ç²ß¥u¬O¥ÑºÊ·þ¦¡¾Ç²ß¤UªºÃþ¯«¸gºô¸ô (Neural Network, NN) ©Òºt¤Æ¥X¨Óªº¡A¸g¹L³o´X¦~¤£Â_ªººt¤Æ¡A¤w±q¥u¯à³B²zºÊ·þ¦¡¾Ç²ß°ÝÃDÂX®i¨ì«DºÊ·þ¦¡©M±j¤Æ¾Ç²ß¡A¤H¤u´¼¼z¡B¾÷¾¹¾Ç²ß©M²`«×¾Ç²ß©¼¦¹¶¡¤§Ãö«Y¤Î¨äºt¶i¦p¹Ï1©Ò¥Ü¡C
¦]¦¹¡AÃþ¯«¸gºô¸ô(NN)¬O²`«×¾Ç²ßªº°ò¦¡A²¨¥¤§¡A¥¦¬O¤@ºØ¼Ò¥é¤H¸£¯«¸g²Õ´(¤SºÙ¯«¸g¤¸)¼Æ¾Ç¼Ò«¬¤Æªº¾÷¨î¡C¥Ø«eAI´¹¤ùªºµo®i«h¦h¥b¦b³B²z²`«×¾Ç²ßÃþ«¬ªº°ÝÃD¡A¦Ó«D³B²z¶Ç²Î¾÷¾¹¾Ç²ß¤Î¯u¥¿¤H¤u´¼¼zªº°ÝÃD¡C¬°¤F¤F¸Ñ²`«×¾Ç²ß´¹¤ù©Î¯«¸gºô¸ô(NN)´¹¤ùªº¹B§@¡A±o¥ý¤F¸Ñ²Õ¦¨¡u¯«¸gºô¸ô (NN)¡v¤§¯«¸g¤¸µ²ºc¤Î¨ä¤u§@ì²z¡C
¯«¸g¤¸ (neuron)¡A¤S¦W¯«¸g²ÓM (nerve cell)¡A¬O¯«¸g¨t²Îªºµ²ºc»P¥\¯à³æ¦ì¤§¤@¡A¨äµ²ºc¦p¹Ï2©Ò¥Ü¡C¤H¸£ªº¬¡Å鯫¸g²ÓM¥]¬A¬ðIJ (Synapses)¡B¾ð¬ð (Dendrites)¡B²ÓMÅé (Cell Body) ©M¶b¬ð (Axon)µ¥¥|Ó¥Dn¥\¯à²Õ¦¨¡A¨ä¬Û¤¬¹B§@Ãö«Y¤Î¥\¯à²z¦p¤U¡G
- ²ÓMÅé (cell body)¡G¥Ñ²ÓM½è°ò½è¡B²ÓM¾¹©M²ÓM®Ö²Õ¦¨¡A¬G²ÓM®Ö¦ì©ó¨ä¤¤¡C²ÓMÅé¬O¯«¸g¤¸³J¥Õ½è¦X¦¨ªº¥Dn³õ©Ò¡A¨ä¹B§@«Y¥Ñ¦¹µo¥X¬ð°_¡A¾ð¬ð (dendrites) ¦ù®i¨ì©P³òªº¯«¸g²Õ´¡A±q¬ðIJ¤¤¦¬¶°¯ß½Ä¡A¨Ã±N¯ß½Ä¶Ç¦^¯«¸g¤¸ªº¤ßŦ¡A¦¹§Y²ÓMÅé¹B§@¥\¯à¡C
- ¾ð¬ð (Dendrites)¡G¬O¤@¸s¥Ñ²ÓMÅéµo¥X¦h¤À¤äªº¾ðª¬¬ð°_¡A¥¦¥i¥H¬O¯«¸g¤¸ªº¿é¤J³q¹D¡A¨ä¥\¯à¬O±N¦Û¨ä¥L¯«¸g¤¸©Ò±µ¦¬ªº°Ê§@¹q¦ì (¹q«H¸¹) ¶Ç°e¦Ü²ÓM¥»Åé¡F¾ð¬ð¤]¥i¥H¬O¯«¸g¤¸ªº°T®§±µ¦¬¯¸¡A¦]»P¨ä¥L¯«¸g¤¸ªº¶b¬ð (Axon) ³s±µ¦Ó§â¶Ç¤Jªº°T®§°e¦Ü¯«¸g¤¸ªº¥»Åé¡C
- ¬ðIJ (Synapses)¡G°T¸¹Åv«¯«¸g¤¸©M¯«¸g¤¸¬Û±µ³B§YºÙ¬°¬ðIJ (synapse)¡A¦]¦¹¬O¯«¸g¤¸¤§¶¡³q«Hªº¯S²§©Ê±µÀY (junction)¡A¨ä¤À´²¦b¾ð¬ð¾ð®Úª¬ÅÖºûªºªí±¤W¡A¨Ì°T¸¹¶Ç»¼ªº¤è¦V«h¥i±N¯«¸g¤¸¤À¬°¡u¬ðIJ«e¯«¸g¤¸ (presynaptic neuron)¡v»P¡u¬ðIJ«á¯«¸g¤¸ (postsynaptic neuron)¡v¡C¯«¸g¤¸¤§¶¡ªº¬ðIJ¥i¥H¤À¬°¤Æ¾Ç¬ðIJ (chemical synapse)©M¹q¬ðIJ (electrical synapse)¨â¤jÃþ¡C
- ¶b¬ð (Axon)¡G§Y¯«¸g²ÓM¤§²ÓMÅéªø¥X¬ð°_¡C¤j¶q¶b¬ð²o³s¤@°_¡A¥H¨ä¥~«¬Ãþ¦ü¦ÓºÙ¬°¯«¸gÅÖºû¡C©ó¯«¸g¨t²Î¤¤¡A¶b¬ð¬O¥Dn¯«¸g«H¸¹¶Ç»¼ªº´ë¹D¡A¥\¯à¬°±N²ÓMÅ餧°Ê§@¹q¦ì (¹q«H¸¹©ÎºÙ¿é¥X¯ß½Ä) ¶Ç¾É¨ì¯«¸g²Õ´¡A³Ì²×¶Ç¾É¦Ü¦b¨ä¥L²ÓM¤§¾ð¬ð¤W¥½ºÝªº¬ðIJ (synapse)¡C
Ãþ¯«¸gºô¸ôªº¤u§@ì²z§Y¼ÒÀÀ¤H¸£ªº¯«¸g¬ðIJ¶Ç»¼µ²ºc¡A³z¹L¯ß½Ä¶Ç»¼°T®§¡C¯«¸g¤¸¤§¶¡ªº¨CÓ³s±µ¥i±N«H¸¹¶Ç»¼µ¹¥t¤@Ó¯«¸g¤¸¡A³B²z«H¸¹ªº¤è¦¡«Y³z¹L¾ð¬ð¦V³s±µ¨ì¥¦ªº¤U´å¯«¸g¤¸µo¥X«H¸¹¡A¬ðIJ¨ã¦³ÀH¾Ç²ß¶i¦æ¦ÓÅܤƪºÅv«¡A³o¥i¼W¥[©Î´î¤p¥¦µo°e¨ì¤U´åªº«H¸¹ªº±j«×¡C¦¹¥~¡A¥¦Ì¥i¨ã¦³ìHÈ (threshold)¡A¨Ï±o¥u¦³¦bÁ`«H¸¹§C©ó©Î°ª©ó¸ÓìHȪùÂe®É¤~µo°e«H¸¹¨ì¤U´å¡C¨CÓ¯«¸g¤¸³£»P¦UºØ¿é¤J¬Û³s¡A¤@¨Ç¿é¤J¦b¯«¸g¤¸¤¤²£¥Í¿Eµo¡A¦Ó¥t¤@¨Ç¿é¤J«h§í¨î¥¦¡A¦p¤H¤u¯«¸gºô¸ô¤¤ªº¥¿tÅv«¡C¦p¹Ï3©Ò¥Ü¡A¨CÓ¯«¸g¤¸¦³nÓ¿é¤JÈ¡A¨Ò¦pX0, X1, ¡K¡K., Xn-1, Xn¡A¦Ó¨CÓ¿é¤JȦ³¤@ÓÅv«(Weight)È¡A¨Ò¦pW0, W1, ¡K¡K..Wn-1, Wn¡A§â©Ò¦³(¿é¤JÈXi)¼¥H(Åv«ÈWi)¥[Á`°_¨Ó¡A¦³®É·|¥t¥~¥[¤W¤@Ó°¾¸mÈ(bi)¨Ó½Õ¾ã¡A±N±o¨ìªºÈ¦A¸g¹L¤@Ó¿Eµo¨ç¼Æ (Activation function) §Y¥i²£¥Í·sªº¿é¥XÈYi¡A¨Ã¥HYi = f (∑Xi•Wi+bi)ªº¦¡¤lªí¥Ü¡C¥H¹Ï3¬°¨Ò¡A¥Ñ¥ª¦Ü¥k¬°±À½× (Inference)¡A¥iÀò±o¿é¥Xµ²ªG¡A¹Lµ{¤¤»ÝnnÓÅv«ÈWi¥[¤W1Ó°¾¸mÈbi¡A¦Ó³o¨ÇȬO»Ýn¸g¥Ñ°V½m (Training) ¨ú±oªº¡C°V½m«e¶·¨ú±o³\¦h(¶V¦h¶V¦n)¤wª¾µª®×©ÎºÙ¬°¼ÐÅÒ (Label) ªº°V½m¸ê®Æ¡A¦pªG²Ä¤@²Õ¿é¤J¸ê®Æ¤wª¾µª®×¬°(1.0)¡A¦ý¸g¹L±À½×«á¥u±o¨ì0.7¡A«hªí¥ÜÅv«¤Î°¾¸mȤ£²z·Q¦Ó»Ýn½Õ¾ã¡A¦¹®É®Ú¾Ú®tÈ (1.0 ¡V 0.7 = 0.3)¥Ñ¥k¦Ü¥ª½Õ¾ã¨CÓÅv«Wi¤Î°¾¸mÈbi¡C¦P¼Ëªº¨BÆJ¦A¿é¤J²Ä¤G²Õ¸ê®Æ¶i¦æ½Õ¾ã¡Aª½¨ì©Ò¦³°V½m¸ê®Æ³£°µ¹L¤@½ü¡C
¬°¤F¯à³B²z§ó½ÆÂøªº°ÝÃD¡A¤@Ó¯«¸g¤¸¥iÂX®i¦¨¤@²Õ²³æ (³æ¼h) ¯«¸gºô¸ô¡A¦p¹Ï4©Ò¥Ü¡A·|¦³¿é¤J¼h¡BÁôÂüh (Hidden Layer) ¤Î¿é¥X¼h¡A±À½×©M°V½mªº·§©À»P¤@Ó¯«¸g¤¸¤jP¬Û¦P¡C
¦AªÌ¡A«Ý¸Ñ¨Mªº°ÝÃD¦pªG§ó¥[§xÃø¡A¦p¹Ï5©Ò¥Ü¡A«h¥i¼W¥[¨C¤@¼hªº¯«¸g¤¸©ÎÁôÂüh¼Æ¨Ó¸Ñ¨M¡A¦Ó§Î¦¨½ÆÂø²`¼hªº¯«¸gºô¸ô¡AºÙ¤§Ãþ¯«¸gºô¸ô¼Ò«¬¡A§Y©Ò¿×ªº²`«×¾Ç²ß¡CÃþ¯«¸gºô¸ô¼Ò«¬¡A«Y¥Ñ¿é¤J¼h (Input layer)¡BÁôÂüh (Hidden layer) ¤Î¿é¥X¼h (Output layer) µ¥¤T¼h©Òºc¦¨¡CÃþ¦ü©ó¤H¸£¤º³¡·|¦³¤j¶qºÙ¬°¡u¯«¸g¤¸¡vªº¯«¸g²ÓM©¼¦¹¬Û³s¡C¯«¸g¤¸«ö¼h²Õ´¡A«H¸¹¸g¾ú³v¼h¶Ç°e±q²Ä¤@¼h (§Y¿é¤J) ¦æ¶i¨ì³Ì«á¼h (§Y¿é¥X)¡CÃþ¯«¸gºô¸ô§Q¥Î¹q¸£¼Ò¥é¤HÃþ¯«¸g²ÓM±µ¦¬°T®§ªºµ²ºc©M¥\¯à¹ï°T®§¶i¦æ³B²z¡AµM«á±N¨äµo°eµ¹¨ä¥L¯«¸g¡C¯«¸gºô¸ô¤¤¼Æ¥H¤dpªº³B²z¸`ÂIt³d©w¦ì¹Ï¹³¡B¤å¥»©M»yµµ¥¸ê®Æ¤¤ªº¼Ò¦¡¡C³o¨Ç¼Ò¦¡¨ó§Uºtºâªk½T©w¤U¤@¨B¦p¦ó°µ¡A¨Ã±Ð¾É¦Û¤v¦p¦ó³z¹L¨C¦¸¦¨¥\°õ¦æ¥ô°È°µ±o§ó¦n¡C·í¤@Ó¯«¸g¤¸±q¨ä¥L¯«¸g¤¸¨º¸Ì±µ¦¬ªº¹q®ð«H¸¹¶q¹F¬Y¤@©wÈ¥H¤W¡A´N·|¯«¸g½Ä°Ê(¿³¾Ä)¡F¦b¬Y¤@©wÈ¥H¤U¡A´N¤£·|¿³¾Ä¡C¿³¾Ä°_¨Óªº¯«¸g¤¸¡A·|±N¹q¾¹«H¸¹¶Ç°eµ¹¤U¤@Ó¬Û³sªº¯«¸g¤¸¡C¤U¤@Ó¯«¸g¤¸¦P¼Ë·|¦]¦¹¿³¾Ä©Î¤£¿³¾Ä¡C¤]´N¬O¡A©¼¦¹¬Û³sªº¯«¸g¤¸¡A·|§Î¦¨Áp¦X¶Ç»¼¦æ¬°¡A³z¹L±N³oºØ¬Û³sªºµ²ºc¨Ó¼Æ¾Ç¼Ò«¬¤Æ¡A«K§Î¦¨¤F¡uÃþ¯«¸gºô¸ô¡v¡C¦]¦¹¡A²`«×¾Ç²ß¬O¤@ºØ¥HÃþ¦ü¤HÃþ¯«¸gºô¸ô (²ºÙ¡uÃþ¯«¸gºô¸ô (Neural Network)¡v) ªº¤è¦¡«Øºc¡u¦h¼h¯«¸gºô¸ô¡v¡AÅý¤H¤u´¼¼z¥i¥H¤@¼h¤@¼h³v¼h¾Ç²ß¡A¨Ã¶i¤Æ¨ì¥i¥H¦Û§Ú¾Ç²ßªº¶¥¬q¡A¥B¥un¦³¥R¤Àªº¾Ç²ß¸ê®Æ¿é¤JÃþ¯«¸gºô¸ô¡A¤£»Ýn¤H¬°ªºÀ°§U¥¦´N¯à¦Û¦æ¤ÀªR¸ê®Æ§ä¥X¯S¼xÈ¡C
¥Ñ©ó²`«×¾Ç²ß¬O³z¹L¼Ò¥é¤H¸£ªº¡uÃþ¯«¸gºô¸ô¡v«Øºc¦h¼h¯«¸gºô¸ô¨Ó³v¼h¾Ç²ß¤j¶q¸ê®Æ¡A¨Ã±N¨äÀ³¥Î©ó¸Ñ¨M°ÝÃDªº¤âªk¡C¯«¸gºô¸ô¼Ò«¬¥i¥Î©ó³B²z¦UºØÃþ«¬¤§¸ê®Æ¡C¾Ç²ß¸ê®Æ«h¬O¥Ñ¿é¤J¸ê®Æ¥H¤Î¬Û¹ïÀ³ªº¥¿½T¸Ñµª©Ò²Õ¦¨¡A¦Ó°V½m¤Î¾÷¾¹¾Ç²ß¥i¥Î¥H§P©w«ÝÂǥѼҫ¬¥Î©ó³B²z¿é¤J¸ê®Æ¤§¤@«Y¼Æ¶°¡A§Y¯«¸gºô¸ô¼Ò«¬¤§¯«¸g¤¸¤§¶¡ªºÅv«¡C¥H¼v¹³¿ëÃѬ°¨Ò¡A¯«¸gºô¸ô¼Ò«¬¥i¸g°V½m¥H¿ëÃѸg±µ¦¬¿é¤J¼v¹³¤º¬O§_¦s¦b¬Y¨ÇÃþ«¬¤§ª«¥ó¡A¬°¤FÅýAI¾Ç²ßÃþ¯«¸gºô¸ôªº¼Ò«¬¡Aº¥ý¥²¶·¥ý±N¼v¹³¾Ç²ß¸ê®Æ¤À³Î¦¨¹³¯À¸ê®Æ¡AµM«á±N¦U¹³¯ÀÈ¿é¶i¿é¤J¼h¡C±µ¨ü¤F¸ê®Æªº¿é¤J¼h¡A±N¹³¯Àȼ¤W¡uÅv«¡v«á¡A«K¶Ç°eµ¹«á¤èÁôÂühªº¯«¸g¤¸¡CÁôÂühªº¦UÓ¯«¸g¤¸·|²Ö¥[«e¤@¼h©Ò±µ¦¬¨ìªºÈ¡A¨Ã±N¨äµ²ªG¦A¼¤W¡uÅv«¡v«á¡A¶Ç°eµ¹«á¤èªº¯«¸g¤¸¡C³Ì«á¡A¸g¥Ñ¿é¥X¼hªº¯«¸g¤¸ªº¿é¥X¡A«K¥i±o¨ì¼v¹³¿ëÃѪº¹w´úµ²ªG¡C¬°¤FÅý¿é¥X¼hªºÈ¸ò¦UÓ¿é¤J¸ê®Æ©Ò¹ïÀ³ªº¥¿¸Ñ¸ê®Æ¬Ûµ¥¡A·|¹ï¦UÓ¯«¸g¤¸ªº¿é¤Jpºâ¥X¾A·íªº¡uÅv«¡vÈ¡C³oÓÅv«ªºpºâ¡A¤@¯ë¬O¨Ï¥Î¡u»~®t˶ǻ¼ºtºâªk¡v¡]Error Back Propagation¡^¡A¨Ï¥Î»P¥¿¸Ñ¸ê®Æ¤§¶¡ªº»~®t¡A±q¿é¥X¼h°f±À¦^¥h¡C³z¹L¦U¡uÅv«¡vªº½Õ¾ã¡A¨ÓÁY¤p¿é¥X¼hªºÈ»P¥¿¸Ñ¸ê®ÆªºÈ¤§¶¡ªº»~®t¡A¥H«Ø¥ß¥X§¹¦¨¾Ç²ßªº¼Ò«¬¡C
2006¦~¨¯¹y (Geoffrey Hinton) ¶}µo¥X¦Û°Ê½s½X¾¹ (Autoencoder¡A²ºÙ¦Û½s½X)¡A§Q¥Î¥H¦Û°Ê½s½X¾¹©ÒÀò±oªºÃþ¯«¸gºô¸ôÅv«°Ñ¼Æȶi¦æªì©l¤Æ¡AÀ³¥Î¦b¡u»~®t˶ǻ¼ºtºâªk¡v¥H´£°ª¦h¼hÃþ¯«¸gºô¸ôªº¾Ç²ß·Ç½T«×¡C§Y¦bÃþ¯«¸gºô¸ôªº¿é¤J¼h©M¿é¥X¼h¨Ï¥Î¬Û¦P¸ê®Æ¡A¨Ã±NÁôÂüh³]¸m©ó¤GªÌ¤§¶¡¡AÂǦ¹¥Î¨Ó½Õ¾ãÃþ¯«¸gºô¸ô¤§¶¡ªºÅv«°Ñ¼Æ¡A±q¦Ó¨Ï²`«×¾Ç²ß¦¨¬°¤@ºØ¥un±N¸ê®Æ¿é¤JÃþ¯«¸gºô¸ô«K¯à¦Û¦æ©â¥X¯S¼xªº¤H¤u´¼¼z¡AºÙ¤§¬°¡u¯S¼x¾Ç²ß¡v(feature learning)¡C
»P¾÷¾¹¾Ç²ß¤£¦P¡A²`«×¾Ç²ß¤ÀÃþ¤£®e©ö®Ú¾Ú¾Ç²ß¤èªk©Î¥Ø¼Ð¶i¦æ¤ÀÃþ¡CAsohµ¥¤H±N²`«×¾Ç²ß¤À¬°½T©w©Ê¼Ò«¬ (deterministic model) ©M·§²v¼Ò«¬ (probabilistic model)¡C½T©w©Ê¼Ò«¬¤¤ªº¥Nªí©Êºtºâªk¦Ü¤Ö¥]¬A²`«×¯«¸gºô¸ô (deep neural network)¡B»¼Âk¯«¸gºô¸ô (recurrent neural network¡ARNN)¡B¨÷¿n¯«¸gºô¸ô (convolution neural network¡ACNN)¥H¤Î¦Û°Ê½s½X¾¹ (autoencoder)¡C¦Ü©ó·§²v¼Ò«¬¡A³q±`¨Ï¥Î¨ü¬Áº¸¯÷°Ò¾÷ (restricted boltzmann machine¡ARBM)¡B²`«×¬Áº¸¯÷°Ò¾÷ (Deep Boltzmann Machines¡ADBM) ©M²`«×¸m«Hºô (deep belief network¡ADBN)¡Cªí1¾ã²z¥XAI¾÷¾¹¾Ç²ß©M²`«×¾Ç²ß¤¤±`¨£¤§ºtºâªk¡C