Preferred Networksï¼PFNï¼ãéçºãããã£ã¼ãã©ã¼ãã³ã°ï¼æ·±å±¤å¦ç¿ï¼ãã¬ã¼ã ã¯ã¼ã¯ãChainerããéããªããåæ£å¦çã«å¯¾å¿ãããã¨ã§å¤§å¹ ã«é«éåãããPFNã®è¥¿å·å¾¹ç¤¾é·ã2017å¹´1æ26æ¥ï¼ç±³å½æéï¼ã«ç±³ãµã³ãã©ã³ã·ã¹ã³ã§éå¬ããããDeep Learning Summit 2017ãã§æããã«ããï¼åç1ï¼ã ç¾å¨ã®Chainerã¯ããã«ããã¼ãï¼è¤æ°ãµã¼ãã¼ï¼ã§ã®å¦ç¿ã«å¯¾å¿ã«ãã¦ãããã1å°ã®ãã¼ãä¸ã§è¤æ°ã®GPUï¼Graphics Processing Unitï¼ã使ç¨ããå ´åããã©ã®GPUã§ä½ã®å¦çãå®è¡ããã®ããããã°ã©ãã¼ãè¨è¿°ããå¿ è¦ããã£ããPFNã¯ãã«ããã¼ãã§ã®å¦ç¿ã«å¯¾å¿ããåæ£ãã¼ã¸ã§ã³ã®Chainerãéçºä¸ã§ã社å ã§ã¯ãã¹ããå§ãã¦ããããã®åæ£ãã¼ã¸ã§ã³ã®Chainerã§ã¯è¨å®ãå¤æ´ããã ãã§ããã«ããã¼ãç°å¢ããã«ãGPUç°å¢ã«å¯¾å¿ãã
2. â¾èªâ¼°å·±ç´¹ä» æµ·é  è£ä¹ lï¬â¯ ï¼æ ªï¼ããªãã¡ã¼ãã¤ã³ãã©ã¹ãã©ã¯ãã£ã¼ lï¬â¯ â¾èªç¶â¾è¨èªå¦ç理ãæ å ±æ¤ç´¢ï¥ªãæ©æ¢°å¦ç¿ãããã¹ãã㤠ãã³ã°ãªã©ã®ç 究éçº lï¬â¯ ç»å解æã¨ããããã¾ã lï¬â¯ Jubatusã®éçºãªã© NLPè¥ï¥´â¼¿æã®ä¼å ±åå§å¡â»â¾§é·ï¼2014-ï¼ 2 3. 宣ä¼ï¼NLPè¥ï¥´â¼¿æã®ä¼ï¼YANSï¼ lï¬â¯ YANSã·ã³ãã¸ã¦ã ï¼9â½æï¼ lï¬â¯ è¥ï¥´â¼¿æç 究è ï¼40æ³æªæºãããï¼ã®è¥ï¥´â¼¿æç 究è ããâ¾èª â¾èº«ã®ç 究ãé²ããããã®éã¾ã lï¬â¯ å¦â½£çãã¨ã³ã¸ãã¢ã®åå ãæè¿ lï¬â¯ ä»å¹´ï¦ãå宿ãããäºå®ã§ã lï¬â¯ ã¹ãã³ãµã¼ãåéããäºå®ã§ã lï¬â¯ YANSæï¼3â½æï¼ lï¬â¯ â¾è¨èªå¦ç理å¦ä¼å ¨å½â¼¤å¤§ä¼æéä¸ã«æ親ä¼ããã¾ã lï¬â¯ åãªã飲ã¿ä¼ã§ã  Jï 3
Preferred Networks ã¨DeNAã人工ç¥è½æè¡ãç¨ãã ä¼æ¥åãã½ãªã¥ã¼ã·ã§ã³ãæä¾ããåå¼ä¼ç¤¾PFDeNAãè¨ç« ï½ãã£ã¼ãã©ã¼ãã³ã°ãä¸å¿ã¨ããé«åº¦ãªæ©æ¢°å¦ç¿æè¡ã§æ§ã ãªãã¼ã¿ã®ä¾¡å¤ãæ大åï½ æ ªå¼ä¼ç¤¾Preferred Networksï¼ããªãã¡ã¼ãã»ãããã¯ã¼ã¯ã¹ãæ¬ç¤¾ï¼æ±äº¬é½å代ç°åºãåµæ¥è 代表åç· å½¹ç¤¾é·ï¼è¥¿å· å¾¹ã以ä¸PFNï¼ã¨æ ªå¼ä¼ç¤¾ãã£ã¼ã»ã¨ãã»ã¨ã¼ï¼æ¬ç¤¾ï¼æ±äº¬é½æ¸è°·åºã代表åç· å½¹ç¤¾é·å ¼CEOï¼å®å® åã以ä¸DeNAï¼ã¯ããã£ã¼ãã©ã¼ãã³ã°ï¼æ·±å±¤å¦ç¿ï¼â»ï¼ãé«åº¦ãªæ©æ¢°å¦ç¿â»2ãªã©æå 端ã®äººå·¥ç¥è½â»3ï¼AIï¼æè¡ãæ´»ç¨ããä¼æ¥åãã½ãªã¥ã¼ã·ã§ã³ãæ¶è²»è åãååã»ãµã¼ãã¹çã®æä¾ãç®çã¨ãã¦ãæ¬æ¥ãåå¼ä¼ç¤¾ãè¨ç«ãã¾ããã æ°ä¼ç¤¾ã§ããæ ªå¼ä¼ç¤¾PFDeNAï¼ãã¼ã»ã¨ãã»ãã£ã¼ã»ã¨ãã»ã¨ã¼ã以ä¸PFDeNAï¼ã§ã¯ãDeNAãå¤å½©ãªã¤ã³ã¿ã¼ããããµã¼ãã¹ã®éå¶ãé
Preferred Networksã¨ã深層å¦ç¿æè¡ãæ´»ç¨ãã¦ã å°éã®è¡æ¶²ãã14種ã®ããã®æç¡ãå¤å®ã§ãã é«ç²¾åº¦ãªã·ã¹ãã ã®éçºãç®æãã å ±åç 究ãéå§ãã¾ããã PFDeNAã¯ãæ ªå¼ä¼ç¤¾ãã£ã¼ã»ã¨ãã»ã¨ã¼ï¼DeNAï¼ã¨æ ªå¼ä¼ç¤¾Preferred Networksï¼PFNï¼ã®åå¼ä¼æ¥ã§ããDeNAãå¤å½©ãªã¤ã³ã¿ã¼ããããµã¼ãã¹ã®éå¶ãéãã¦èç©ãã¦ããæ§ã ãªãã¼ã¿ãè¤æ°äºæ¥é åã§ã®çµé¨ã¨ãPFNã®æ©æ¢°å¦ç¿ã»æ·±å±¤å¦ç¿æè¡ã«é¢ããåºç¯å²ãªç¥è¦ãæå 端ã®æè¡ãçµã¿åããããã¨ã«ãããããããäºæ¥é åã§ä¼æ¥åãã½ãªã¥ã¼ã·ã§ã³ãæ¶è²»è åããµã¼ãã¹ãæä¾ãã¦ããã¾ãã DeNADeNA ã¯ãåµæ¥ä»¥æ¥ã¤ã³ã¿ã¼ãããé åã§æ§ã ãªãã¸ãã¹ã«ææ¦ãç¶ããããã«2004å¹´ããããã¡æ©ãã¢ãã¤ã«ã«éä¸ããäºæ¥ãå±éãã¦ãããã¨ã§ãã¢ãã¤ã«ã²ã¼ã ãã¯ããã¨ããã¤ã³ã¿ã¼ããããµã¼ãã¹ã®åéã«ããã¦ãä¸çããã
DeNAã¨Preferred Networksã人工ç¥è½æè¡ãç¨ãã ä¼æ¥åãã½ãªã¥ã¼ã·ã§ã³ãæä¾ããåå¼ä¼ç¤¾PFDeNAãè¨ç« æ ªå¼ä¼ç¤¾ãã£ã¼ã»ã¨ãã»ã¨ã¼ï¼æ¬ç¤¾ï¼æ±äº¬é½æ¸è°·åºã代表åç· å½¹ç¤¾é·å ¼CEOï¼å®å® åã以ä¸DeNAï¼ã¨æ ªå¼ä¼ç¤¾Preferred Networksï¼ããªãã¡ã¼ãã»ãããã¯ã¼ã¯ã¹ãæ¬ç¤¾ï¼æ±äº¬é½å代ç°åºãåµæ¥è 代表åç· å½¹ç¤¾é· CEOï¼è¥¿å· å¾¹ã以ä¸PFNï¼ã¯ããã£ã¼ãã©ã¼ãã³ã°ï¼æ·±å±¤å¦ç¿ï¼â»1ãé«åº¦ãªæ©æ¢°å¦ç¿â»2ãªã©æå 端ã®äººå·¥ç¥è½â»3ï¼AIï¼æè¡ãæ´»ç¨ããä¼æ¥åãã½ãªã¥ã¼ã·ã§ã³ãæ¶è²»è åãååã»ãµã¼ãã¹çã®æä¾ãç®çã¨ãã¦ãæ¬æ¥ãåå¼ä¼ç¤¾ãè¨ç«ãã¾ããã æ°ä¼ç¤¾ã§ããæ ªå¼ä¼ç¤¾PFDeNAï¼ãã¼ã»ã¨ãã»ãã£ã¼ã»ã¨ãã»ã¨ã¼ã以ä¸PFDeNAï¼ã§ã¯ãDeNAãå¤å½©ãªã¤ã³ã¿ã¼ããããµã¼ãã¹ã®éå¶ãéãã¦èç©ãã¦ããæ§ã ãªãã¼ã¿ãè¤æ°äºæ¥é åã§ã®çµé¨ã¨ãPFNã®AIæè¡ãç¹ã«
Chainer â A flexible framework of neural networks¶ Chainer is a powerful, flexible and intuitive deep learning framework. Chainer supports CUDA computation. It only requires a few lines of code to leverage a GPU. It also runs on multiple GPUs with little effort. Chainer supports various network architectures including feed-forward nets, convnets, recurrent nets and recursive nets. It also supports
ããã«ã¡ã¯ãå¾å± ã§ããæè¿ã¯æ¯æ©ã¤ã«ã«ãªã£ã¦æ¦å ´ãé§ãã¾ãã£ã¦ãã¾ãã æ¬æ¥ãDeep Learning ã®æ°ãããã¬ã¼ã ã¯ã¼ã¯ã§ãã Chainer ãå ¬éãã¾ããã Chainer å ¬å¼ãµã¤ã GitHub â pfnet/chainer Chainer Documentation Chainer ã¯ããã¥ã¼ã©ã«ãããã誤差éä¼ææ³ã§å¦ç¿ããããã®ãã¬ã¼ã ã¯ã¼ã¯ã§ãã以ä¸ã®ãããªç¹å¾´ãæã£ã¦ãã¾ãã Python ã®ã©ã¤ãã©ãªã¨ãã¦æä¾ï¼è¦ Python 2.7+ï¼ ãããããã¥ã¼ã©ã«ãããã®æ§é ã«æè»ã«å¯¾å¿ åçãªè¨ç®ã°ã©ãæ§ç¯ã«ããç´æçãªã³ã¼ã GPU ããµãã¼ãããè¤æ° GPU ãã¤ãã£ãå¦ç¿ãç´æçã«è¨è¿°å¯è½ ãã¥ã¼ã©ã«ããããã©ã®ããã«æ¸ããã 次ã®ã³ã¼ãã¯å¤å±¤ãã¼ã»ãããã³ã®å¾é ãè¨ç®ããä¾ã§ãã from chainer import FunctionSet, Vari
æ¯æ¥æãã§ãããæ¯æ¸ã§ãã ã¡ããã©ä»é±ã·ã«ã´ã§éããã¦ããSIGKDD2013ã§Best research paperã«é¸ã°ããEdo Libertyæ° (Yahoo! Haifa Labs)ã®âSimple and Deterministic Matrix Sketchingâã®ã¢ã«ã´ãªãºã ãå®è£ ãã¦å ¬éãã¦ã¿ã¾ããã å è«æPDFã¯èè ãµã¤ããããç§ãæ¸ããPythonã³ã¼ãã¯Githubããããããå ¥æã§ãã¾ãã SIGKDD (ACM SIGKDD Conference on Knowledge Discovery and Data Mining)ã¯ACM主å¬ã§è¡ããããç¥èçºè¦ï¼ãã¼ã¿ãã¤ãã³ã°ã«ããããããä¼è°ã§ããæè¿ã¯æ©æ¢°å¦ç¿ã¨ã®å¢ç®ãææ§ã«ãªã£ã¦ãã¾ããããæ»èªæã«ã¯çè«çãªæ°ããã ãã§ãªããå®ãã¼ã¿ï¼ç¹ã«å¤§è¦æ¨¡ãã¼ã¿ï¼ã使ã£ãå®é¨ã§ã®è©ä¾¡ãå¿ è¦ã¨ãããã®ãç¹å¾´ã§ãã
ãä¹ ãã¶ãã§ãã徳永ã§ãã è¬è«ç¤¾ãããªã³ã©ã¤ã³æ©æ¢°å¦ç¿ã¨ããã¿ã¤ãã«ã®æ¬ãåºçããã¾ããèè ã¯Preferred Infrastructure/Preferred Networksã®æµ·é, 岡éå, å¾å± , 徳永ã®4人ã§ãã æ©æ¢°å¦ç¿ã®ä¸ã§ããªã³ã©ã¤ã³æ©æ¢°å¦ç¿ã«ç¹åããæ¬ã§ãåç´ãã¼ã»ãããã³ããå§ã¾ããPassive Aggressive, Confidence Weighted, AROW, Soft Confidence Weightedãªã©ï¼Passive Aggressive, Confidence Weighted, AROWã¯åæ£ãªã³ã©ã¤ã³æ©æ¢°å¦ç¿ãã¬ã¼ã ã¯ã¼ã¯Jubatusã§ãå®è£ ããã¦ãã¾ãï¼ã«ã¤ãã¦ã¢ã«ã´ãªãºã ã®æ¦è¦ã説æãããããªã°ã¬ãã解æã«ããæ§è½è§£æã«ã¤ãã¦èª¬æãã¦ãã¾ããã¾ããåæ£ç°å¢ã§ã®ãªã³ã©ã¤ã³æ©æ¢°å¦ç¿ãã深層å¦ç¿ã§ã®å¿ç¨ãå¹ççãªå®è£ æ¹æ³ãªã©ãå¿ç¨çãª
2. èªå·±ç´¹ä» æµ·é è£ä¹ (@unnonouno) Preferred Infrastructure èªç¶è¨èªå¦çãæ©æ¢°å¦ç¿ãªã©ã®ç 究éçº ç»åå¦çãã¡ãã£ã¨ããã¾ã NLPè¥æã®ä¼å ±åå§å¡é·ï¼2014~ï¼
7. :::::::: ââââââââââââââââ â :::::::: ï½ OOPSLAãããããããã ãªâ¦ â ::::: ââââââââââââââââï½ââââ¬â ::::: ï½ãããâ¦å¥´ã¯SIGPLANå天çã®ä¸ã§ãæå¼± ⦠â ââââââââââââï½ââ¬âââââââââ ï½ ãã£ã24åã§ãªãã¥ã¼ã¢ã«ã¨ã¯â ï½ OOPã®é¢æ±ãã⦠â âââââï½âââââââââ |ï¾, ï¼ ï½ã½ /! ,.ââ、 |彡/äºOãã|ã /ä¸ä¸ä¸!, |! `,â \、、_,|/-ャ ã `=j r=ï¾ /ã !彡 T çª| / /ï¿£|/´__,ャ |`ä¸ä¸â/ |`=、|,=â| /人 ã½ ã=âï¼|`:::::::/イ__ ãï½ã¼ ã__,-, 、 _!_ / / `ã¼âââ |_,.ã¤ï½¤ | |/、 Y /| | | j / ï¾`â´â彡\ POPL PLDI ICFP 8.
åç2âãã¬ã¸ã£ã¼ãã¼ã¿ã®ã¡ã³ãã¼ãCEOã¯ç±³ã¬ããããããªã©ã§å¤åããè³å·è£èª æ°ãCTOã®å¤ªç°ä¸æ¨¹æ°ã¯ããªãã¡ã¼ãã¤ã³ãã©ã¹ãã©ã¯ãã£ã¼ã®åCTOã楽天ã§åæ£ãã¼ã»ããªã¥ã¼ã»ã¹ãã¢ã®ROMAãéçºãã西澤ç¡ææ°ãOSSã®ãã°åéãã¼ã«Fluentdãã¡ãã»ã¼ã¸ã³ã°ããã«ã¦ã¨ã¢MessagePackãéçºããå¤æ©è²ä¹æ°ãMongoDB JPã®è¨ç«ã¡ã³ãã¼äºä¸æ¬æµ©æ°ãªã©ãåå ãã¦ãã ç±³ãã¬ã¸ã£ã¼ãã¼ã¿ã¯2012å¹´9æ27æ¥ï¼ç±³å½æéï¼ããã¼ã¿ã¦ã¨ã¢ãã¦ã¹ï¼DWHï¼ã®ã¯ã©ã¦ããµã¼ãã¹ã§ãããTreasure Data Cloud Data Warehouseï¼DWHï¼ Serviceããéå§ããã¨çºè¡¨ããããHadoopãããã¼ã¹ã«ããDWHã ããHadoopç¬èªã®ãMapReduceãã§ã¯ãªããSQLã®ã¯ã¨ãªã¼ããJDBCããODBCããªã©ã使ã£ã¦èç©ãããã¼ã¿ãæ´»ç¨ã§ãããã¨ãç¹å¾´ãç±³å½
ã天æãã¨å¼ã°ããããã°ã©ããã¡ãèªãä¼ç¤¾ãç«ã¡ä¸ããã¨ã©ããªããããã®çããç¥ããããã°ããããã³ãã£ã¼ä¼æ¥ã«æ³¨ç®ããã°ãããITæ¥çã§æ³¨ç®ãéãããã¯ããã¸ã¼ãã³ãã£ã¼ãããªãã¡ã¼ãã¤ã³ãã©ã¹ãã©ã¯ãã£ã¼ãã®ç¤¾é·ã«è©±ãèããã ç¹éãå¦çèµ·æ¥å®¶ãã¡ã®èåããæçµåã¯ç¾å¨ãæã注ç®ãéãã¦ãããã¯ããã¸ã¼ãã³ãã£ã¼ã®1ã¤ãããªãã¡ã¼ãã¤ã³ãã©ã¹ãã©ã¯ãã£ã¼ï¼PFIï¼ã®ä»£è¡¨åç· å½¹ç¤¾é· è¥¿å·å¾¹æ°ã«ç»å ´ãã¦ããã£ããæ±äº¬å¤§å¦å¤§å¦é¢ã«å¨å¦ä¸ãACMå½é大å¦å¯¾æããã°ã©ãã³ã°ã³ã³ãã¹ãï¼ACM/ICPCï¼ã®ä¸ç大ä¼ã«åºå ´ããã¡ã³ãã¼ã¨è¨ç«ããä¼ç¤¾ã¯ãèªç¶è¨èªå¦çã®åéã§æ¥å¢ãã«ãã®åå¨æã大ãããã¦ããã ä»åã¯ç¹å¥ã«ãé£è¼ã天æããã°ã©ãã«èã10ã®è³ªåãã§ããªãã¿ãLispããã«ã¼ã®ç«¹å ééæ°ã«ã¤ã³ã¿ãã¥ã¢ã¼ããé¡ãããã2人ã®æ¿å¯ãªå¯¾è«ã®æ§åãä½ãã¨ãããªããå±ãããï¼ä»¥ä¸ãæ¬ç§°ç¥ï¼ã åºä¼
ä»åã®ã²ã¹ãã¯ãï¼â æ ªï¼ããªãã¡ã¼ãã¤ã³ãã©ã¹ãã©ã¯ãã£ã¼ï¼PFIï¼ã®æé«æè¡è²¬ä»»è 太ç°ä¸æ¨¹ããããã§ãã¼ã»ç¹å¥ç ç©¶å¡ å²¡éå大è¼ãããã¨ã³ã¸ã㢠ç°ä¸è±è¡ãããæ¬èªã§ãVol.49 ç¹å¥ä¼ç»ãâ ï¼»â éç¿ï¼½ã¬ã³ã¡ã³ãã¨ã³ã¸ã³ããã¯ããã¨ããè¨äºãå·çããã ãã¦ãã¾ããããã£ã¼ããªè©±é¡ã§å¯¾è«ãçãä¸ããã¾ããã ï¼å·¦ããï¼å¤ªç°ä¸æ¨¹æ°ãç°ä¸è±è¡æ°ã岡éå大è¼æ°ãå°é£¼å¼¾æ°ï¼æ®å½±ï¼æ¦ç°åº·å®ï¼ å°å¦çã®ã¨ããããã¼ã¿å§ç¸® å¼¾ï¼ããããèªå·±ç´¹ä»ããé¡ããã¾ãã 岡éåï¼ä»¥ä¸ã岡â ï¼â ï¼ï¼PFIã«ã¯ï¼ä¼ç¤¾è¨ç«ã®ã¨ããããã¾ãã¦ãä»ã¯å¤§å¦ã®å士課ç¨3å¹´[1]ã«ãå¨ç±ãã¦ãç 究ã¨ãç 究ãããã«ãã¦ä¸ã®ä¸ã®ãã¨ã«å¿ç¨ã§ããå¯è½æ§ãããããå®éã«ãããã試ãã¨ãããã¨ããã¦ãã¾ããèªåã®é¢å¿ã¨ãã¦ã¯ãæãããã¼ã¿å§ç¸®ã大好ãã§ãå°ã»ä¸å¦æ ¡ã®ã¨ãã«ãNIFTY SERVE[2]ã®éä¿¡ãé ãã®ããã£ããã§lhaã¨
ãªãªã¼ã¹ãé害æ å ±ãªã©ã®ãµã¼ãã¹ã®ãç¥ãã
ææ°ã®äººæ°ã¨ã³ããªã¼ã®é ä¿¡
å¦çãå®è¡ä¸ã§ã
j次ã®ããã¯ãã¼ã¯
kåã®ããã¯ãã¼ã¯
lãã¨ã§èªã
eã³ã¡ã³ãä¸è¦§ãéã
oãã¼ã¸ãéã
{{#tags}}- {{label}}
{{/tags}}