深層å¦ç¿ (æ©æ¢°å¦ç¿ãããã§ãã·ã§ãã«ã·ãªã¼ãº)
- ä½è : 岡谷貴ä¹
- åºç社/ã¡ã¼ã«ã¼: è¬è«ç¤¾
- çºå£²æ¥: 2015/04/08
- ã¡ãã£ã¢: åè¡æ¬ï¼ã½ããã«ãã¼ï¼
- ãã®ååãå«ãããã° (13件) ãè¦ã
ãããªäººã«ãããã
- 人工ç¥è½ãªã¸ãµã³*1ã«ãªããªãããã«
- ç·å½¢ä»£æ°ã¨å¾®åã®åºç¤çãªç´ é¤ã®ãã人
- ãã£ã¼ãã©ã¼ãã³ã°ã®åºç¤ãå¦ã³ãã人
ç« ç«ã¦
- 第1ç« ãã¯ããã«
- 第2ç« ãé ä¼æåãããã¯ã¼ã¯
- 第3ç« ã確ççå¾é éä¸æ³
- 第4ç« ã誤差éä¼ææ³
- 第5ç« ãèªå·±ç¬¦å·åå¨
- 第6ç« ãç³è¾¼ã¿ãã¥ã¼ã©ã«ããã
- 第7ç« ãå帰åãã¥ã¼ã©ã«ããã
- 第8ç« ããã«ããã³ãã·ã³
æ¸è©
1å¹´åã«æ
å ±ç³»ã«å°éãå¤ãããã£ã¼ãã©ã¼ãã³ã°é¢ä¿ã§åãã«èªãã æ¬ãã¡ããã©1å¹´åã¯ççºçã«æµè¡ãå§ããããã ã£ãã
ä»ã¯ããå°ãããããã¨æ¥æ¬èªã®æ¬ãããã£ã¦ãã¦ãããããããªããããã®æ¬ã¯åºç¤ããæ°å¼ã追ããªãããã¡ãã¨å¦ã¹ãæ¬ã ã¨æãã
æ°å¦ã®è¦æãªäººã§ãé°å²æ°ã¯ããããããªæ§æã«ãªã£ã¦ããããææ°ã®ãã£ã¼ãã©ã¼ãã³ã°ã®ç 究ãå§ãããï¼ã¨ãã人ã§ããããããå
¥ãã®ã¯ã¨ã¦ãè¯ãã¨æãã
以ä¸åç« ã®ç°¡åãªèª¬æã
â»ä½æ°ãªãæ¸ãã¦ããã©ãããã£ã¨ç解ããããªãããã ãèªãã°ãããããå³ã¨ããªãã®ã§é©å®ãããã¤ã¤ã
æºåã®1-4ç«
Deep learningã«ããã¦å¤§äºãªåºç¤äºé
ãæ¸ãã¦ããã¦ãããåå¦è
ã¯ãã£ããèªãã¨ããã
ãã£ã£ã£ã£ã£ããã¾ã¨ããã¨ã以ä¸ã®ãã¨ãæ¼ããã¦ããã¦ã»ããã
- Deep learningã¨ã¯ãè¡åè¨ç®+éç·å½¢è¦ç´ (æ´»æ§é¢æ°)ã«ãããæ®éã®æ©æ¢°å¦ç¿ãããæ ¼æ®µã«åé¡ã¨ãã®è½åãä¸ããããããã¤
- å¿ é ã¯ãè¡åè¨ç®ã»åæé¢æ°ã®å¾®å
- è¦ããã¹ãç¨èª*2
- ãã¥ã¼ã©ã«ããã
- æ´»æ§åé¢æ°(ç¹ã«æè¿ã¯ReLU)
- 誤差éä¼æ¬
- å¾é å¹ææ³
- ãã精度ãä¸ããããã®å·¥å¤«
- ããããã
- ããããã¢ã¦ã
- æ£åå
åºæ¬çã«è¡åã¨å¾®åãããã£ã¦ãã°èªãããªã¨ããæããæ®éã®ç·å½¢åé¡å¨ã«æ¯ã¹ã¦è¯ãã¨ããããã®ã¯ ãã¶ãæ´»æ§åé¢æ°ãå¹ãã¦ãããã ãããªãã¨ããå°è±¡ã
Autoencoderã¨ãã次å åæ¸ã®æ¹æ³ã
ãµã¤ãºã®å¤§ããç»åã¨ãã ã¨ãå¦ç¿ã®ããã«æ¬¡å
ãæ¸ãããããªã¨ãããã¨ãããã¯ãã
Autoencoderã¨ããæ¹æ³ã使ãã°ã次å
åæ¸ãè¡ãããã¨ãããã¨ãè¨ã£ã¦ããã
ç°¡åã«è¨ãã¨ãããããã¨
- å ¥åã¨åºåãä¸ç·ã«ãã¦ããã¦ãä¸é層ã®å±¤ã®æ°ãæ¸ããã¦å¦ç¿ããã
- ä¸é層ãæãã¦æ¥ãã°å ¥åãã次å ã®ä½ããã®ã«ãªã£ã¦ããï¼
æ°å¼ã§ã主æååæã¨ä¸ç·ã«ãªãã¨ãããã¨ã示ãã¦ããã®ã¨ããã¥ã¼ã©ã«ãããã®ã»ããèªç±åº¦ãé«ã
æ¹è¯ã§ããã¨ãããã¨ãæ¸ãã¦ãã(denoising autoencoder)
Deep learningã®ç骨é ï¼ç»ååæã®CNNã«ã¤ãã¦æ¸ãã¦ãã6ç«
ç»ååé¡ã³ã³ãã¹ããå¸å·»ãã¦ããCNNã«ã¤ãã¦ã®èª¬æããã®è¾ºãããã£ã¼ãã©ã¼ãã³ã°ã£ã½ããªã£ã¦ãã
ãã£ããããã¨ã以ä¸ã®ãã¨ãåãã£ã¦ããã°ããã
- ç³ã¿è¾¼ã¿:ãã£ã«ã¿ã¼ãç»åã«ããããã¨ã§ãå±æçãªç¹å¾´ãæã
- ãã¼ãªã³ã°ï¼é£æ¥ããç»ç´ ãã¾ã¨ãããã¨ã§ãããã¹ãæ§ãå¾ã
å®ã¯ã¾ã ã¾ã é²åãéãã¦ãã¦ã
ææ°ã®state of the artã¯152層(!)ã¨ãã«ãªã£ã¦ãããResidualNet(åæ)
ã¨è¨ããå®ã¯ãã®ResidualNetãé¢ç½ã工夫ããã¦ããã®ã§ããã®ãã¡ç´¹ä»ãã(ããã)
ãã¨ãCNNã¯æè¿ã§ã¯DQN*3ã®æèã§åºã¦ãããã¨ãå¤ããªã£ã¦ããã å¼·åå¦ç¿ãããããã¶ãæµè¡ããããããã®ãã¡ç´¹ä»ãã(ããã)
æç³»åã«ã ã£ã¦ä½¿ããï¼RNNã«ã¤ãã¦æ¸ãã¦ãã7ç«
æç³»åæ
å ±ãå
¥ãã¦ããã¨åé¡ã次ã®äºæ¸¬ã¨ããã§ãããã¥ã¼ã©ã«ããããæè¿ã®ã°ããã¼ã©IoTã§ã使ããããã
ãã ãã¯ããã¯ããã®ã ããLSTMãCTCã®èª¬æã¯å°ãä¸è¦ªåã«æããããããã¯ã¼ã¯ããããããã®ã§ãããããªãæ°ããããã
ãã£ããããã¨ä¸ã®ãããªæã
- RNN(Recurrent Neural Network)ã¯ãæç³»åæ¹åã«ãã¥ã¼ã©ã«ããããã¤ãªããã¨ã§ãæç³»ååæãå¯è½ã«
- LSTM(Long-Short Term Memory)ã使ããã¨ã§ãã©ãããæ å ±ãè¦ãã¦ããã°ããããå¦ç¿ãã
- CTC(Connectionist Temporal Classification)ã使ããã¨ã§ãé³å£°èªèã«ä½¿ããããã«ãªã*4
çæã¢ãã«ã ã£ã¦è¨è¿°ã§ããï¼RBMã«ã¤ãã¦ã®8ç«
ãã®ç« ã¯æ£ç´é£ããã£ããå¤åããã«ã³ãå ´ã¿ãããªã®ãä»®å®ãã¦ã¨ãã«ã®ã¼ã®è¨ç®ããã¦ããã¿ãããªæãã ã¨ããã*5 çæã¢ãã«ã«ãããã¨ã§ã確çåå¸ã¨ãã¦å¤ãæ±ããã¨ãã§ããã®ãããããã(7ç« ã¾ã§ã®ã¢ãã«ã¯å ¨é¨èå¥ã¢ãã«ï¼ç¢ºççãªèãã¯ãªã)
http://qiita.com/sergeant-wizard/items/0a57485bc90a35efcf26 ãã ãæè¿ã¯GAN(åæ)ã¨ã VAE(åæ)ã¨ãçæã¢ãã«ãæµè¡ã£ã¦ããã®ã§ããããèªãå段éã§ãããèªãã®ã¯ããããã
GAN : Generative Adversarial Netwokrk
VAE : Variational Autoencoder
*1:ä¸ã®ä¸ã®ãã¨ã¯æ·±ãæã«çªã£è¾¼ãã°è§£æ±ºã§ããã¨æã£ã¦ãããä¼ç¤¾ã®ä¸ã®äººã«ããããã¿ã¤ã
*2:ãã¨ã§ã°ã°ããªããæ¸ç±èªããªããã¦ã
*3:DQNãããªããï¼Deep Q Networkã ãï¼
*4:ä¸è¬ã®RNNã¯ããããã¨ããåèªããã®ã¾ã¾ã«ãããããããCTCã使ãã¨ããããâãã¨ãã¦èªèãã¦ããããæéã«å¯¾ãã¦ããã¹ãã«ãªãã®ã§ãé³å£°èªèã«ä½¿ãããããªã
*5:ä¿è¨¼ã¯ã§ããªããçæã¢ãã«ãã¬ãã