å¤åè¿ä¼¼ï¼Variational Approximationï¼ã®åºæ¬ï¼ï¼ï¼
ãã¦ãååã¯å¤åè¿ä¼¼ã®ç®çï¼è¤ééãã¦è§£æ解ãå¾ãããªããããªç¢ºçåå¸ã®è¿ä¼¼ï¼ã¨ãè¿ä¼¼ã®ããã®ææ¨ï¼KL divergenceï¼ã«é¢ãã¦è§£èª¬ãã¾ããã
ä»åã¯ãå¤åè¿ä¼¼ã®ãå ¬å¼ããå°ãã¦ã¿ããã¨æãã¾ããè¿ä¼¼åå¸$q(z)$ã«é¢ãã¦ãå解ã®ä»®å®ããç½®ããã¨ã«ãããæ°å¤æé©åã«ãããå¾é æ³ã®ãããªç¹°ãè¿ãã®æ´æ°æç¶ããå¾ãããè¿ä¼¼åå¸ãæ°å¤çã«è¨ç®ã§ãããã¨ã示ãã¾ãã
Â
[å¿ è¦ãªç¥è]
ä¸è¨ãããã£ã¨ã ã確èªãã¦ããã¨ããã§ãã
- ååã®è¨äºã®å 容
- å¾é æ³
Â
ååã¯ãè¿ä¼¼åå¸ã¨çã®äºå¾åå¸ã®éã®KL divergenceããªãã¹ãå°ãããããã¨ã§ããã¾ãè¿ä¼¼åå¸ãå¾ããã¨ããæ¹éã«ãã¾ããã
\[ KL(q(z)||p(z|x)) = - \int q(z) \ln \frac{p(z|x)}{q(z)} dz \]
ã§ããããä¸çªã®ãã½ãªã®ã§ãããå¤åè¿ä¼¼ã§ã¯ãè¿ä¼¼ãããåå¸ã次ã®ããã«è¤æ°ã®åå¸ã«å解ãããã¨ããä»®å®ãç½®ãã¾ãã
\[ p(z|x) \approx q(z) = \prod_{i} q(z_i) \]
ä»åã¯ç°¡åã®ããã«ãï¼ã¤ã®åå¸ã«å解ããã¨ãããã¨ã«ãã¦ããã¾ãããã*1
\[ q(z) = q(z_1)q(z_2) \]
å¤åè¿ä¼¼ã«ãããKL divergenceã®æå°åæ¹æ³ã¯ãæ°å¤æé©åã«ãããåå¾®åã使ã£ãå¾é æ³ã¨ä¼¼ããããªç¹°ãè¿ãæ´æ°ã®æ¦ç¥ãåãã¾ããã¤ã¾ãã$q(z_1)$ã¨$q(z_2)$ãä¸åº¦ã«ããã£ã¨æ´æ°ããã®ã§ã¯ãªããä¸æ¹ãåºå®ãã¦ä¸æ¹ãæ´æ°ãã¨ããã®ãç¹°ãè¿ãã¾ãã
- $q(z_2)$ãã©ã³ãã ã«åæåããã
- $q(z_2)$ãåºå®ããä¸ã§ãKL divergenceãæå°åãã$q(z_1)$ãæ±ããã
- $q(z_1)$ãåºå®ããä¸ã§ãKL divergenceãæå°åãã$q(z_2)$ãæ±ããã
- 以ä¸ã®2ï¼3ãååãªåæ°ã¾ã§ç¹°ãè¿ãã*2
 ã§ã¯ãå®éã«ä¸è¨ã®ã¹ããã2ã®ããã®æ´æ°å¼ãå°ãã¦ã¿ã¾ããããï¼ã¹ããã3ã¯ã¹ããã2ã¨åãã§ããï¼
ããããã¡ãã£ã¨æ°å¦ã«ãªãã¾ãã$q(z_1)$ãæ´æ°ããããã«ã¯ã$q(z_2)$ãåºå®ããKL divergenceã$q(z_1)$ã®ã¿ã®é¢æ°ã§ããã¨èãã¾ã*3ã å¼ãå¼ã«ä»£å ¥ããKL divergeneã®å¼ãã$z_1$ã«ç¡é¢ä¿ãªé ã$c$ã«ã¾ã¨ãã¦ãã¾ãã¨ã
\[ KL(q(z)||p(z|x)) = - \int \int q(z_1)q(z_2) \ln \frac{p(z_1, z_2|x)}{q(z_1)q(z_2)} dz_1dz_2 \\ = - \int \int q(z_1)q(z_2) \{ \ln p(z_1, z_2|x) - \ln q(z_1) - \ln q(z_2) \} dz_1dz_2 \\ = - \int q(z_1) \bigl\{ \int q(z_2) \ln p(z_1, z_2|x) dz_2 - \int q(z_2) \ln q(z_1)dz_2 \bigr\}dz_1 \\ + \int q(z_1)q(z_2)\ln q(z_2) dz_1dz_2 \\ = - \int q(z_1) \{\int q(z_2) \ln p(z_1, z_2|x) dz_2 - \ln q(z_1) \}dz_1 + c \;\;\;\;(1) \]
ç©åè¨ç®ãå ¥ãåã«ãªã£ã¦ããã®ã§æ³¨æãã¦ãã ãããããã«è¨ç®ãé²ããã¨ããã®å¼ã¯çµå±æ¬¡ã®ãããªéå§å°ç¹ã¨ã¯ã¾ãå¥ã®KL divergenceã«è½ã¡çããã¨ãåããã¾ãã
\[ (1) = - \int q(z_1) \ln \frac{exp \{\int q(z_2) \ln p(z_1, z_2|x) dz_2 \}}{q(z_1)}dz_1 + c \\ =KL(q(z_1) || exp \{ \int q(z_2) \ln p(z_1, z_2|x) dz_2 \} ) +c \]
KL divergenceã¯2ã¤ã®åå¸ãä¸è´ããã¨ããæå°å¤0ãåãã¾ãããããã£ã¦å¼ã®æå°å¤ã¯æ¬¡ã®ããã«ãã¦å¾ããã¨ãã§ãã¾ãã
\[ \ln q(z_1) = \int q(z_2) \ln p(z_1, z_2|x) dz_2 + c \]
ããã§ã®å®æ°é $c$ã¯åå¸ãæ£è¦åï¼ç©åã1ã«ãªãï¼ããããã®å®æ°ã§ããä¸è¨ã®æå¾ å¤è¨ç®ï¼ç©åï¼ã¯ã¡ãã£ã¨é·ã£ããããã®ã§ããã©ã±ãã$\langle \cdot \rangle$ã使ã£ã¦æ¸ãç´ãã®ã便å©ã§ãã
\[ \ln q(z_1) = \langle \ln p(z_1, z_2|x) \rangle_{q(z_2)} + c \]
ä¸å¿ãã®å¼ãå¤åè¿ä¼¼ã®å ¬å¼ã«ãªãã¾ãããèªå以å¤ã®åå¸ã§æå¾ å¤ãåããã¨è¦ãã¦ããã°OKã§ãã
å®éã®å¿ç¨ã§ã¯ãå ·ä½çãªç¢ºçã¢ãã«ãè¨å®ããä¸ã®æå¾ å¤è¨ç®ãé å¼µã£ã¦æè¨ç®ãã¾ããä¸ã¤å¤§äºãªãã¤ã³ãã¯ããã®å¼ãè©ä¾¡ããçµæãç°¡åã«æ£è¦åã§ããï¼ç©åã§ããï¼ç¢ºçåå¸ã«ãªã£ã¦ãããã¨ã§ããããã«é¢ãã¦ã¯å®éãã¢ãã«ãå解ã®ä»®å®ãè²ã å¤ãã¦æè¨ç®ãã¦ã¿ãã¨ãããå°éãªè©¦è¡é¯èª¤ãå¿ è¦ãªå ´åãããã¾ããä¸æ¹ã§ãç·å½¢ã¬ã¦ã¹ã¢ãã«ãæ··åã¢ãã«ããã©ã¡ã¼ã¿ã«å¯¾ããå ±å½¹äºååå¸ã®è¨å®ãªã©ããã¾ãè¨ç®ã§ãããã¨ãè¯ãç¥ããã¦ããã¢ãã«ã®æ§ç¯æ¹æ³ãããã®ã§ããã£ã½ã©ã®ãã¨ããªãéãããããã£ã便å©ãªé¨åãçµã¿åãããã®ãè¯ãã¨æãã¾ãã*4
Â
ãã¦ãã²ã¨ã¾ã確çã¢ãã«ã«åºã¥ãå¤åè¿ä¼¼ã®é©ç¨æ¹æ³ãã¾ã¨ãã¦ããã¾ãããã
- 確çã¢ãã«ãå®ç¾©ãããï¼æ··åã¬ã¦ã¹åå¸ãHMMãªã©ï¼
- äºå¾åå¸ã«é¢ããå解ã®ä»®å®ãããã
- å ¬å¼ã使ã£ã¦æ´æ°å¼ãå°åºããã
- å®è£ ãã¦åããã*5
Â
ä¸å¿å¤åè¿ä¼¼ã使ã£ãè¿ä¼¼ã¢ã«ã´ãªãºã ãåããã®ã«å¿ è¦ãªç¥èã¯ããã ããªã®ã§ããããªãã ãæ½è±¡çã§ããªãã ããã£ããããªããªãã¨ããæããããããããªããã¨æãã¾ãã
次åã¯ããã®æ¹æ³ã使ã£ã¦ç°¡åãªå¤åè¿ä¼¼ã®ä½¿ç¨ä¾ã¨ãç°¡åã§ã¯ãªãï¼ãã©ç¾å®çãªï¼å¤åè¿ä¼¼ã®ä½¿ç¨ä¾ãç´¹ä»ãããã¨æãã¾ãã
[ç¶ãã»é¢é£]
å¤åè¿ä¼¼ï¼Variational Approximationï¼ã®åºæ¬ï¼ï¼ï¼ - ä½ã£ã¦éã¶æ©æ¢°å¦ç¿ã
*1:å®éã®å¿ç¨ã§ã¯ãå ·ä½çãªã¢ãã«ã«åããã¦é©åãªå解ãæãã¤ãå¿ è¦ãããã¾ããä¾ãã°æ¨æºçãªæ··åã¢ãã«ã§ã¯ããã©ã¡ã¼ã¿ã¨é ãå¤æ°ã®ï¼ã¤ã«å解ããã°å¹çã®è¯ãè¿ä¼¼ã¢ã«ã´ãªãºã ãå¾ããã¾ãï¼EMã¢ã«ã´ãªãºã ï¼ããã£ã¨è¤éãªã¢ãã«ã ã¨ãï¼ã¤ä»¥ä¸ã®ç´°ããå解ãå¿ è¦ã«ãªã£ã¦ããå ´åãããã¹ã¦ã®å¤æ°ã¬ãã«ã§ãã©ãã©ã«ããå ´åï¼å¹³åå ´è¿ä¼¼ãmean field approximationï¼ãããã¾ããåºæ¬çã«ã¯è¨ç®ãã§ããæä½éã®å解æ°ã«åãã¦ããã®ãæé©åã®è¦³ç¹ã§ã¯è¯ãã®ã§ãããããå¤ãã®ã¡ã¢ãªã使ç¨ããã¨ããæ¬ ç¹ãããã¾ãã
*2:ç¹°ãè¿ãè¨ç®ãè¡ãåæ°ã¯ã"å¤åä¸é"ã¨å¼ã°ããå®æ°å¤ãæ¯ã¹ããããã¨ã«è©ä¾¡ãã¦æ±ºããã®ãä¸è¬çã§ããå¤åä¸éã¯å¤åè¿ä¼¼ã«ãããç®çé¢æ°ã§ãå ·ä½çãªè¨ç®æ¹æ³ã¯ã¾ãå¥ã®æ©ä¼ã«è§£èª¬ãã¾ãã
*3:ãã®å ´åKL divergenceã¯é¢æ°$q(z_1)$ã®é¢æ°ã«ãªã£ã¦ããã®ã§ãæ±é¢æ°ã¨å¼ã°ãã¾ããæ±é¢æ°ãæé©åããåé¡ãªã®ã§"å¤åæ³"ã¨ããååãã¤ãã¦ãã¾ããä»åã®è¨äºã§ã¯KL divergenceã®ä¸è¬çãªæ§è³ªã使ã£ã¦æ´æ°åãå°ãã¾ãããæ°å¦çãªå¤åã使ã£ã¦åãæ´æ°åãå°ããã¨ãå¯è½ã§ãã
*4:å ¬å¼ã®å³è¾ºãæ£è¦åã§ããããã«ã¢ãã«ãçµãã®ãçæ³ã§ãããã©ããã¦ãæ£è¦åã§ããªããããªè¤éãªã±ã¼ã¹ã«ã¯ãããªãè¿ä¼¼ãèããããããã¾ãããä¾ãã°ãããã«ç´°ããå解ã®ä»®å®ãèãã¦ã¿ãã$q(z)$ã«é©å½ãªãã©ã¡ããªãã¯ãªåå¸ï¼ã¬ã¦ã¹åå¸ãªã©ï¼ãããã¦å¾é æ³ãªã©ã®æ°å¤æé©åæ³ãé©ç¨ãã¦ãã©ã¡ã¼ã¿ãæ±ããããµã³ããªã³ã°ãç¨ãã¦å¿ è¦ãª$q(z)$ã®çµ±è¨éãæ±ãããã©ãã©ã¹è¿ä¼¼ãç¨ããããªã©ã§ãã
*5:ããã§ãå¤åä¸éã®è¨ç®ãå®è£ ãã¦ããã¨ããããã°ã«ä¾¿å©ã§ããã¾ãå¥ã®æ©ä¼ã«ç´¹ä»ãã¾ãã