[è«–æ–‡] Parallel Inference for Latent Dirichlet Allocation on Graphics Processing Units - tsubosakaã®æ—¥è¨˜NIPS 2009ã®online papersãŒã™ã§ã«ãƒ€ã‚¦ãƒ³ãƒãƒ¼ãƒ‰ã§ãるよã†ã«*1ãªã£ã¦ãŸã®ã§ã‚¿ã‚¤ãƒˆãƒ«ã‚’眺ã‚ã¦ã„ãŸã‚‰GPUã§LDAを並列化ã™ã‚‹ã¨ã„ã†è«–æ–‡ãŒã‚ã£ã¦èªã‚“ã ã®ã§å°‘ã—紹介ã—ã¦ã¿ã‚‹ã€‚ ã¾ãšã€å½¼ã‚‰ã®è«–æ–‡[1]ã§ã¯LDAã®æŽ¨è«–アルゴリズムã®ã†ã¡ã€Collapsed Gibbs sampling(CGS)ã¨Collapsed Variational Bayes(CVB)ã®2ã¤ã«é–¢ã—ã¦ä¸¦åˆ—化を試ã¿ã¦ã„ã‚‹ãŒCollapsed Gibbs samplingã®æ–¹ã«é–¢ã—ã¦ã®ã¿ç´¹ä»‹ã™ã‚‹ã€‚ã¾ãŸã€å½¼ã‚‰ã®è«–æ–‡ã§ã¯GPGPUã®çµ±åˆé–‹ç™ºç’°å¢ƒã¨ã—ã¦CUDAを用ã„ã¦ã„る。 LDAã«ã¤ã„㦠LDAã¯è«–æ–‡[2]ã§æ案ã•ã‚ŒãŸã€æ–‡ç« 生æˆãƒ¢ãƒ‡ãƒ«ã§ãƒˆãƒ”ック分æžãªã©ã«åºƒã用ã„られã¦ã„る。 モデルã®æŽ¨è«–アルゴリズムã¨ã—ã¦ã¯å¤‰åˆ†ãƒ™ã‚¤ã‚º[1]ã€ã‚®ãƒ–スサンプリング[4]ã€EPã€collapsed 変分ベイズ[5]ãªã©ãŒçŸ¥ã‚‰ã‚Œã¦ã„る。 ã“
![[è«–æ–‡] Parallel Inference for Latent Dirichlet Allocation on Graphics Processing Units - tsubosakaã®æ—¥è¨˜](https://cdn-ak-scissors.b.st-hatena.com/image/square/a5f4c2888f93e54cefd6facbcae7e335bcda3a58/height=288;version=1;width=512/https%3A%2F%2Fcdn.image.st-hatena.com%2Fimage%2Fscale%2F49f7342dee132dfc4fd17d92c2ccabf202e67ce7%2Fbackend%3Dimagemagick%3Bheight%3D1200%3Bversion%3D1%3Bwidth%3D1200%2Fhttps%253A%252F%252Fcdn-ak.f.st-hatena.com%252Fimages%252Ffotolife%252Ft%252Ftsubosaka%252F20091129%252F20091129221225.jpg)
{{#tags}}- {{label}}
{{/tags}}