ãŸããç¬èªã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ã©ã€ãã©ãªãããããšãæ確ã«ããå¿ èŠããããŸãã ãã¡ãããããã»ã©åºå€§ã§éå¿çãªãã®ã§ã¯ãããŸããããã¯ãŒããã®ç¹å®ã®ã¿ã¹ã¯ã解決ããããã«èšèšãããŠããŸãã ãã€ãŠãæ¢è£œã®ãœãªã¥ãŒã·ã§ã³ã䜿çšãã代ããã«ãç¬èªã®ãœãªã¥ãŒã·ã§ã³ãäœæããããšã«ããŸããã ã¯ãã¹ãã©ãããã©ãŒã ãã³ã³ãã¯ãããæ®ãã®ã³ãŒããžã®çµ±åã®å®¹æããããŸããŸãªãµãŒãããŒãã£ã®äœè ããã®äžè¬çã«äžèŠãªããŒã«ãžã®å€æ°ã®äŸåé¢ä¿ãæã€çŸçäžæ¬æãªã©ãããŸããŸãªçç±ã§ã ãã®çµæã䟿å©ãªFïŒã§èšè¿°ãããããŒã«ãäœæãããçŽ2MBã®ã¹ããŒã¹ãåããå¿ èŠãªåŠçãå®è¡ããŸããã äžæ¹ã§ãé床ã®é¢ã§ã¯ããªãé ããGPUã³ã³ãã¥ãŒãã£ã³ã°ããµããŒããããæ©èœãããªãã³ã°ãããŠãããçŸä»£ã®çŸå®ãžã®æºæ ã«çåãæ®ããŸãã
åé¡ã¯ããã€ãŒã«ãåçºæããããããã«äœ¿çšã§ããããã«ããããäžè¬çã«ã¯æ°žé ã§ãã ãããŠãçãã®ã¯ãŒã ã¯ãèªåèªèº«ã®æ段ãç¶æããããšã¯äžåœãªãã®ã§ããããªãœãŒã¹ãæ¶è²»ããå¯èœæ§ãå¶éããŠããããšãåžžã«ããããŸããã å®æçãªæªåã«ããããã¹ãŠã®æ®éã®äººãããã§ããããã«ãããšãã°ãTheanoãTorchã䜿çšããããã«ã¢ã€ãã¢ãå€ãããŸããã ããããæã¯å±ããŸããã§ããã ãããŠãTensorFlowã®ãªãªãŒã¹ã«ããããã1ã€ã®è¿œå ã®åæ©ãçããŸããã
ãã®æèã§ãç§ã¯ãã®ã·ã¹ãã ã«å¯ŸåŠãå§ããŸããã
ã€ã³ã¹ããŒã«ããã»ã¹ã«é¢ããçãã¡ã¢
TensorFlowã¯ãããããä»ã®å€ãã®ææ°ã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ã©ã€ãã©ãªãããç°¡åã«ã€ã³ã¹ããŒã«ã§ããŸãã éããããã°ãåé¡ã¯Linuxã¿ãŒããã«ã©ã€ã³ã«å ¥åããã1è¡ã«å¶éãããå ŽåããããŸãã ã¯ããWindowsã¯äŒçµ±çã«ãµããŒããããŠããŸãããããã®ãããªäºçŽ°ãªããšã¯ç¢ºãã«æ¬åœã®éçºè ãæ¢ããŸããã
Ubuntu 11ã§ã¯ãTensorFlowã¯ã€ã³ã¹ããŒã«ãæåŠããŸããã ãããã04/14ãžã®ã¢ããã°ã¬ãŒããšã¿ã³ããªã³ãšã®ãã³ã¹ã®åŸãäœãããŸã æ©èœããŠããŸããã å°ãªããšããå ¥éã»ã¯ã·ã§ã³ããã³ãŒããã©ã°ã¡ã³ããå®è¡ããããšã¯å¯èœã§ããã ãããã£ãŠãç¹ã«æ°ãããã£ã¹ããªãã¥ãŒã·ã§ã³ããããPython 2ããŒãžã§ã³ã2.7.9以äžã§ããå ŽåãTensorFlowã®ã€ã³ã¹ããŒã«ã¯ç°¡åã§é£ãããªãããšãå®å šã«èšè¿°ã§ããŸãã Linuxã®å Žåãè€éãªãœãããŠã§ã¢ããã±ãŒãžãããã«ã€ã³ã¹ããŒã«ãããªãå Žåãããã¯æ£åžžã§ãïŒãã¶ããæ£åžžã§ã¯ãªããããããŸããããããã¯æ£åžžãªç¶æ³ã§ãããããèµ·ãããŸãïŒã
äœæ¥ã®æ€èšŒã
ããã§ç§ã¯æ¬¡ã®ããšãèšããªããã°ãªããŸããã 次ã®ãã¹ãŠã¯ãããªãçãé åã§å人çãªå¥œå¥å¿ãæºããããã«äœæããããå人çãªçµéšããã®å人çãªäŸãšèŠãªãããå¿ èŠããããŸãã èè ã¯ã調æ»çµæãäžççã«éèŠã§ãããšããäžè¬çã«ã¯äŸ¡å€ããããšããµããããŸããã çµæã®æ¬ ç¹ã«ã€ããŠã®è°è«ã¯ãTensorFlowã·ã¹ãã èªäœïŒåºæ¬çã«ã¯å®å šã«æ©èœããæãéèŠãªã®ã¯è¿ éã«æ©èœããïŒã«èµ·å ãããã®ã§ã¯ãªããç¹å®ã®ã¢ãã«ãšãã¬ãŒãã³ã°äŸã«èµ·å ãããã®ã§ãã
ãã£ããããã
å°æ¥ã®ããã«èšå®ãããMNISTããæ°åãèªèããããšã«ã€ããŠã®å ¥éã¬ãã¹ã³ãçµäºããããã«ã·ãŒã±ã³ã¹ããã·ãŒã±ã³ã¹ãžã®åŠç¿ã«é¢ããã»ã¯ã·ã§ã³ãéããŸããã å±éºã«ãããããŠãããã®ãæ確ã«ããããã«ãåé¡ã®æ¬è³ªãããå°ã詳ãã説æããŸãã
ã·ãŒã±ã³ã¹ããã·ãŒã±ã³ã¹ãžã®åŠç¿ã®äžè¬çãªã¿ã¹ã¯ã¯ãæåã®å ¥åã·ãŒã±ã³ã¹ã«åºã¥ããŠæ°ããæåã·ãŒã±ã³ã¹ãçæããããšã§ãã ç¹å®ã®ã±ãŒã¹ã§ã¯ãã·ã³ãã«ã¯åèªã§ãã ãã®ã¿ã¹ã¯ã®æãæåãªã¢ããªã±ãŒã·ã§ã³ã¯ãããããæ©æ¢°ç¿»èš³ã§ããããèšèªã®æãã¢ãã«ã®å ¥åã«éä¿¡ãããåºåã§å¥ã®èšèªãžã®ç¿»èš³ãååŸãããå Žåã§ãã å ¥éã¬ãã¹ã³ã®å³ã«ç€ºãããã«ããã®ãããªã¢ãã«ã«ã¯2ã€ã®äž»èŠãªã¯ã©ã¹ããããŸãã 1ã€ç®ã¯ãšã³ã³ãŒããŒãã³ãŒããŒãªãã·ã§ã³ã䜿çšããŸãã å ã®ã·ãŒã±ã³ã¹ã¯ãã¥ãŒã©ã«ãããã¯ãŒã¯ã䜿çšããŠåºå®é·è¡šçŸã«ãšã³ã³ãŒãããã2çªç®ã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ã¯ãã®è¡šçŸãæ°ããã·ãŒã±ã³ã¹ã«ãã³ãŒãããŸãïŒå³1ïŒ[1]ã 人ãããã¹ããäžããããèšæ¶ããããã«æ±ãããã次ã«éžæããŠå¥ã®èšèªãžã®ç¿»èš³ãæžãããã«æ±ãããããããªãã®ã§ãã 2çªç®ã®ã¯ã©ã¹ã¯ããã³ãŒããŒãæäœäžã«å ã®ã·ãŒã±ã³ã¹ã«ãã¹ãã€ãã§ããå Žåãéžæç泚æã®ã¢ãã«ã䜿çšããŸãïŒå³2ïŒ[2]ã

å³1.ãã¥ãŒã©ã«ã·ãŒã±ã³ã¹ã€ã³ã¢ãã«ããã¬ãŒãã³ã°ããããã®åºæ¬ã¢ãã«ã å·ŠåŽïŒæåã®åïŒã«åæã·ãŒã±ã³ã¹ãå³åŽã«-çæãããã·ãŒã±ã³ã¹ã tensorflow.orgããã®ç»å

å³2.ãã¥ãŒã©ã«ã·ãŒã±ã³ã¹ã€ã³ã¢ãã«ããã¬ãŒãã³ã°ããããã®åºæ¬ã¢ãã«ã tensorflow.orgããã®ç»å
ãã£ãããããã¯ãã·ãŒã±ã³ã¹ã€ã³ã·ãŒã±ã³ã¹åŠç¿ã¿ã¹ã¯ã®ç¹å¥ãªã±ãŒã¹ã§ãïŒå ¥åã§ã®è³ªåãåºåã§ã®åçïŒã åã«ã ãã®ã¡ãœããã䜿çšãããã£ãããããã®å®è£ ã«ã€ããŠæžããŸããã å°ãåŸã«ãGoogle瀟å¡ã«ããèšäº[3]ãåããããã¯ã«ç»å ŽããŸããããå³ã«ç€ºããã®ãšåæ§ã®ããé«åºŠãªïŒèãããšããã®ïŒã¢ãã«ã䜿çšããŸããã 1ã
Googleã®ã¢ãã«ã¯LSTMã»ã«ã䜿çšããŠããŸãããç§ã¯éåžžã®ãªã«ã¬ã³ããããã¯ãŒã¯ã䜿çšããŠãã£ããããããæ§ç¯ããæé©ãªäœæ¥ã®ããã«1ã€ã®å€æŽã®ã¿ããã蟌ã¿ãŸããã èšäº[3]ã«ç€ºãããŠãããã€ã¢ãã°ã¯ãç§ãæã«å ¥ããããšãã§ãããã®ãããå°è±¡çã§èå³æ·±ãããã«èŠããŸãïŒããã«ããŠãŒã¶ãŒãµããŒããµãŒãã¹ã®ãã€ã¢ãã°ã§ãã¬ãŒãã³ã°ããããã£ããããããææ矩ãªãã«ããæäŸã§ãããšããäºå®ã«ã€ããŠè©±ããŸãïŒã ããããGoogleã®ãã£ãããããã¯ã以åã®ãµã³ãã«ãããã¯ããã«å€ãã®ããŒã¿ã®ã³ã¬ã¯ã·ã§ã³ã§ãã¬ãŒãã³ã°ãããŠããŸãã
æ©æ¢°ç¿»èš³ãç®çãšããTensorFlowã¹ã€ãŒãã®æšæºäŸãä¿®æ£ããã®ã§ããã£ãããããã®ãã¬ãŒãã³ã°ã«äœ¿çšãããã€ã¢ãã°ã®ã³ã¬ã¯ã·ã§ã³ããããŒã¿ãããŠã³ããŒãããŸããïŒGoogleãã£ãããããããã¬ãŒãã³ã°ãããæ°åäžã®å Žåãšã¯å¯Ÿç §çã«ã3000ã®äŸãããããŸããïŒã Googleã®ã¬ãã¹ã³ã§ã¯ã[2]ããã®ã¢ãã«ãšéžæçãœããããã¯ã¹[4]ãäŸã§å®è£ ãããŠãããã€ãŸãããã®åéã®ææ°ã®çµæããã¹ãŠå®éã«é©çšããããšè¿°ã¹ãããŠããŸãã
ç§ã®ãã£ãããããã¯ãåè¿°ã®ãšãããç³ã¿èŸŒã¿ãããã¯ãŒã¯ããšã³ã³ãŒããŒãšããŠäœ¿çšãïŒãµã€ãºã2ã¯ãŒããš3ã¯ãŒãã®16åã®ãã£ã«ã¿ãŒãåãã1ã€ã®ã¬ã€ã€ãŒ+æ倧ãŠããªã³ã¬ã€ã€ãŒïŒããã³ãŒããŒã¯åçŽãªElmanãªã«ã¬ã³ããããã¯ãŒã¯ã䜿çšããŸãã ä»ã®ç 究è ã«ãããšãã©ã¡ããã©ã¡ããã·ãŒã±ã³ã¹ããã·ãŒã±ã³ã¹ãžã®åŠç¿ã¿ã¹ã¯ã§ããèªäœãããŸã瀺ããŠããŸããã ãã®ãããã·ã¹ãã ã«1ã€ã®å€æŽãé©çšããã1幎以äžåã«å¥ã®ã¿ã¹ã¯ïŒã¬ãã¥ãŒã®çæïŒã®ããã«çºæããŸããã 1ã€ã®ç³ã¿èŸŒã¿ãããã¯ãŒã¯ïŒãšã³ã³ãŒããŒïŒã®ä»£ããã«ã2ã€ã®ç³ã¿èŸŒã¿ãããã¯ãŒã¯ã䜿çšãããŸãã1ã€ã¯ãœãŒã¹ããã¹ãã®ãšã³ã³ãŒãçšã§ããã1ã€ã¯çæãããã°ããã®æ°ãããã®ã®ãšã³ã³ãŒãçšã§ãã ãŠããªã³ã®æåŸã®ã¬ã€ã€ãŒã®åºåã¯ã次ã®ã¬ã€ã€ãŒã§ãã¢ã§æ¥ç¶ãããŸãïŒã€ãŸãã1ã€ã®ãã¥ãŒãã³ãæåã®ãããã¯ãŒã¯ã®ãã¥ãŒãã³ãã1ã€ã®å ¥åãåãåãã2çªç®ã®ãã¥ãŒãã³ãã1ã€ã®å ¥åãåãåããŸãïŒã å ¥åä¿¡å·ã«å¯Ÿå¿ããäœããçæããããšããã®ä¿¡å·ã¯2çªç®ã®ãããã¯ãŒã¯ã«ãã£ãŠæå¶ãããã·ã¹ãã ã¯æ®ãã®ããã¹ããçæãç¶ãããšããèãæ¹ã§ããã ç§ã®ç¥ãéãããã®ãããªè§£æ±ºçã¯ã©ãã«ãèšèŒãããŠããããå®è£ æã«ã¯æããã«èšèŒãããŠããŸããã§ããã ïŒãã£ãããããã¯1ã€ã®ãšã³ã³ãŒããŒãããåªããŠãããšã¯ããïŒæªãèŠããããã«æ©èœããéžæçã¢ãã³ã·ã§ã³ã¡ã«ããºã ã«é¢ããèšäºãèŠããšãã«ãããæŸæ£ããŸããã
çµæã¯æ¬¡ã®ãšããã§ãã

å³3. TensorFlowã¹ã€ãŒãã®seq2seq.embedding_attention_seq2seqã¢ãã«ã䜿çšãããã£ããããããšã®ãã€ã¢ãã°
ãã£ããããããšã®ãã€ã¢ãã°ãããããã®ç¿»èš³ïŒ
TensorFlow Kitã®ã¢ãã« | ç§ã®ã¢ãã« |
HïŒããã«ã¡ã¯ïŒããã«ã¡ã¯ïŒ
å®å ïŒããã«ã¡ã¯ïŒããã«ã¡ã¯ïŒ HïŒå æ°ã§ããïŒ ïŒå æ°ã§ããïŒ KïŒå æ°ã§ãïŒçŽ æŽãããïŒ HïŒã©ãã«ããã®ïŒ ïŒã©ãã«ããŸããïŒ KïŒå æ°ã§ãïŒå æ°ã§ãïŒ DïŒäœãæããŠïŒäœãèšã£ãŠïŒ CïŒã¯ãïŒã¯ãïŒ HïŒããã£ããããã«äŒããŸãããïŒããã«äŒããŸãããïŒ KïŒãããã§ããã HïŒãã€ïŒãã€ïŒ KïŒãã€ïŒãã€ïŒ | HïŒããã«ã¡ã¯ïŒããã«ã¡ã¯ïŒ
KïŒããã«ã¡ã¯ïŒããã«ã¡ã¯ïŒ HïŒå æ°ã§ããïŒ ïŒå æ°ã§ããïŒ KïŒéåæãæããïŒéåæãæããïŒ HïŒã©ãã«ããã®ïŒ ïŒã©ãã«ããŸããïŒ KïŒç§ã¯ä»å®¶ã«ããŸãïŒç§ã¯å®¶ã«ããŸãïŒ DïŒäœãæããŠïŒäœãèšã£ãŠïŒ å®å ïŒäžçïŒäžçïŒ HïŒããã£ããããã«äŒããŸãããïŒããã«äŒããŸãããïŒ KïŒåãããªã...ïŒåãããªãïŒ HïŒãã€ïŒãã€ïŒ KïŒããã£ãïŒããã£ãïŒ |
äž¡æ¹ã®ã¢ãã«ã¯æ©èœããŸãããæ®å¿µãªããã倧ããªæ¹åã¯åŸãããŸããã§ããã ããã«ãç§ã®äž»èŠ³çãªæèŠã§ã¯ãå³åŽã®å¯Ÿè©±ã¯ããèå³æ·±ããããæ£ç¢ºã«èŠããŸãã ã»ãšãã©ã®å Žåãåé¡ã¯ããŒã¿ã®éã«ãããŸãã å³ããã®ã¢ãã« 1.ããªãã®ä»£è¡šåãæã£ãŠããŸãããæå³ã®ããçµæãçæãå§ããã«ã¯å€ãã®ããŒã¿ãå¿ èŠã§ãïŒåã®èšäºã§èª¬æããç¶æ³ïŒã ç§ã®ãã£ãããããã¢ãã«ã¯ããã»ã©è¯ããªããããããŸããããããŒã¿äžè¶³ã«çŽé¢ããŠãæå³ã®ããçµæãçã¿åºãããšãã§ããŸãã ããã«ãããããšãã°ãéããããã€ã¢ãã°ã§ããŸããŸãªäººã ã®éã®ã³ãã¥ãã±ãŒã·ã§ã³ã®ã¢ãã«ãäœæã§ããŸãã ããšãã°ã3000çµã®ãã¬ãŒãºã®ã»ããããã察話è ã®1人ã ãã®ãã¬ãŒãºã®åçãååŸãããšã次ã®ããã«ãªããŸãã

å³4
ç§ã¯ïŒåã³äž»èŠ³çã«ïŒããç©æ¥µçã§å奜çãªã³ãã¥ãã±ãŒã·ã§ã³ãæããŠããŸãã ããªãã¯ã©ãã§ããïŒ TensorFlowãããã®ã¢ãã«ããã¯ãè¡šã«ç€ºãããŠããæé«ã®ãã€ã¢ãã°ãéæã§ããŸããã§ãããã5ã€ã®ç°ãªãæ§æã®ã¿ããã§ãã¯ããŸããããããã䜿çšããçµéšã®ãã人ãããè¯ãçµæãåŸãããšãã§ããŸãã
äžé£ã®åèªããã®ãã¬ãŒãºã®åæ§ç¯
åèªã®ã»ããããã®ãã¬ãŒãºã®åæ§ç¯ã¯åæã¿ã¹ã¯ã§ããã顧客ã®1人ãèšå®ããå®éã®ã¿ã¹ã¯ã®ä»£ããã«ããã§äœ¿çšããŸãã 顧客ã¯ãã¿ã¹ã¯èªäœãšç¬èªã®ããŒã¿ã䜿çšããäŸã®äž¡æ¹ã®å ¬éã«å察ããŠããŸããããã®ããããã®èšäºã§ã¯ããã©ãŒã ã䌌ãŠããå¥ã®ã¿ã¹ã¯ãæãã€ããŸããã
ãã¬ãŒãºãšããŠãç§ã¯ãŠãŒã¶ãŒã¯ãšãªãæ€çŽ¢ãšã³ãžã³ã«äœ¿çšããŸããããªããªãããããã¯çãæå³ã®ãããã¬ãŒãºã®è¯ããœãŒã¹ã§ãããããã«ãååãªéãæå ã«ãã£ãããã§ãã ã¿ã¹ã¯ã®æ¬è³ªã¯æ¬¡ã®ãšããã§ãã ã泚æã®ããã«åæ¥èšŒæžãäœæããããšãããªã¯ãšã¹ããããããã®å°ç¡ãããŒãžã§ã³ããåæ¥èšŒæžãäœæããããšãããªã¯ãšã¹ãããããšããŸãã å ã®æå³ãä¿æããªãããç ŽæããããŒãžã§ã³ããæå³ã®ããèŠæ±ãå床è¡ãå¿ èŠããããŸãã ã€ãŸããããŠã£ã³ããŠãé çªã«äžŠã¹ããã¯èª€ã£ãçµæãšèŠãªãããŸãã çãããããŠè²ã£ãããŒãžã§ã³ã¯ãåèªã®åé 眮ãæ§å¥ãçªå·ã倧æåå°æåã®å€æŽãããã³4æåããçããã¹ãŠã®åèªã®åé€ã«ãã£ãŠèªåçã«çæãããŸããã åèšã§120,000ã®ãã¬ãŒãã³ã°ãµã³ãã«ããã®æ¹æ³ã§äœæããããã®ãã¡1,000ããã¹ãçšã«ç¢ºä¿ãããŸããã ã¿ã¹ã¯ã¯æ©æ¢°ç¿»èš³ã®åé¡ãããç°¡åã«æããŸãããåæã«ãããšå ±éç¹ããããŸãã
å ã®åé¡ã解決ããã«ã¯ãäžèšã®ãã£ãããããã®ã»ã¯ã·ã§ã³ã§èª¬æããã¢ã€ãã¢ã«åºã¥ããŠç¹å¥ãªã¢ãã«ãéçºããå¿ èŠããããŸããã 顧客ã®ããŒãºã«å¯Ÿããå質ãäžååã ã£ããããã»ã¬ã¯ãã£ããœããããã¯ã¹ã®ã¢ã€ãã¢ãæãèµ·ãããããéåžžã«å€§ããªèŸæžãæäœããããŒã«ãè¿œå ããŸããã ãšããã§ãTensorFlowãã¥ãŒããªã¢ã«ããã4æ¥åã«åããŠã»ã¬ã¯ãã£ããœããããã¯ã¹ã«ã€ããŠåŠã³ãŸããã ãã®åã«ã幞éã«ãçµæã«ã€ããŠã¯å€æããŠããŸããããçµæã«ã€ããŠã¯å°ãåŸã«ãªã£ãŠãç§ã¯åœŒã«ã€ããŠã®èšäºãèŠãŸããã ãã®ã¢ãã«ã«ã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®ããã¡ã³ã¿ãžãŒãã®åºŠåããå¶åŸ¡ããæ段ãšã圢æ ãèæ ®ããã ãã§ãªãã圢æ ãèæ ®ããåé¡ãåé¿ããæ段ãè£ åãããŠããŸããããã«é¢é£ããåé¡ããããŸãã ãã ãããã®ãœãªã¥ãŒã·ã§ã³ã«ã¯ãéžæçãªæ³šæã®éåžžã®ã¡ã«ããºã ããªããå ã®ã·ãŒã±ã³ã¹ãè¡šãããã®åºæ¬çãªã¡ã«ããºã ããããLSTMãŸãã¯GRUã¢ãžã¥ãŒã«ã¯äœ¿çšããŸããã ãã®ãããç§ã®å®è£ ã«ããã圌ã®äœæ¥ã®é床ã¯éåžžã«é©åã§ãã
çµæã¯æ¬¡ã®ãšããã§ãã

å³5
TensorFlowãããã®ã¢ãã«ãè¡ã£ãããšã¯æ¬¡ã®ãšããã§ãã çæãããæ€çŽ¢ã¯ãšãªã¯ãå ã®ããŒã¯ãŒããšé¢ä¿ããããäžè¬ã«ããã·ã¢èªã®èŠåã«åŸã£ãŠäœæãããŸãã ãããŠããã©ãã«ã®æå³ãæã¡ãŸãã ãå±æ ¹ãåãããã®ä¹³æ¯ãã¯ããã ãã®äŸ¡å€ããããŸãã äžæ¹ããããã®æ§æã§ç§ãåã°ããã®ã¯ãèšèªã®ååã®ã¢ãã«ç解ãšåµé çãªã¢ãããŒãã§ãã ããšãã°ãããã¿ã€ãããéé±ç³é±ç©Fe2O3ãã§ãããããã¿ã€ãã¢ãã¿çšã®è£ 眮ããªãããšãç¥ããªãå Žåããããã¿ã€ãã¢ãã¿è£ 眮ãã¯ãã£ãšãããããå»åŠçã«ã¯ã©ãããèãããŸãã ããããå®éçãªç®çã§ã¯ãããã¯é©åã§ã¯ãããŸããã ãã¹ãããã100ã®ãã¹ãã±ãŒã¹ã®ãã¡ã1ã€ãæ£ãããããŸããã§ããã
ç§ã®ã¢ãã«ã¯æ¬¡ã®ãªãã·ã§ã³ãäœæããŸããïŒ

å³6
ããã§ããªãªãžãã«ã¯ç ŽæããããŒãžã§ã³ãçæããããã®ã¯çæããããã®ã人éã¯åææ€çŽ¢ã¯ãšãªã§ãã 100ã®æ€èšŒæžã¿ã®äŸã®ãã¡ã72ïŒ ãæ£ããã§ãã TensorFlowãããã®ã¢ãã«ã®æ¹åã¯ããŸããããŸããã§ããã
æ©èœã®ã¹ããŒããšå¹
ãããã®ãã©ã¡ãŒã¿ãŒã§ã¯ãGoogleããã®ããã±ãŒãžã¯ç¢ºãã«ç§ã®ã©ã€ãã©ãªãäžæ¡äžåã£ãŠããŸãã ãŸããææ°ã®ãã¹ãŠã®ã¡ãœãããLSTMããã³GRUãèªå埮åã䜿çšããç»å解æã®ããã®ç³ã¿èŸŒã¿ãããã¯ãŒã¯ãåããŠãããäžè¬ã«ããã¥ãŒã©ã«ãããã¯ãŒã¯ã ãã§ãªããããããçš®é¡ã®ã¢ãã«ãç°¡åã«äœæã§ããŸãã ãã¹ãŠãéåžžã«çŽæçã«è¡ãããååã«ææžåãããŠããŸãã ãã¡ãããLinuxãŸãã¯MacOãããå Žåã¯ãç¹ã«åå¿è åãã«æ©æ¢°åŠç¿ãè¡ãããšããå§ãããŸãïŒãããããcygwinãmingwããŸãã¯ãã®ä»ã®æ¹æ³ã䜿çšããŠãœãŒã¹ã³ãŒããWindowsã§ã³ã³ãã€ã«ã§ããŸãããããã¯å ¬åŒã«ã¯ãµããŒããããŠããŸããïŒã
é床ã«é¢ããŠã¯ããŸã æ£ç¢ºãªæž¬å®ãè¡ã£ãŠããŸããããCPUã®TensorFlowã¢ãã«ã§ã¯ãå®è£ ã®2åãã3åã®é床ã§ãã»ãŒåãæ°ã®ãã©ã¡ãŒã¿ãŒãå®è¡ãããŠããããã«æããŸãïŒããã©ãŒãã³ã¹ã®å€§ããªéããäºæ³ãããŸãïŒã ãŸããå€§å¹ ã«å°ãªãã¡ã¢ãªãæ¶è²»ããŸãã ãŸããGPUããŒãžã§ã³ã¯CPUå®è£ ããã10åé«éã§ãïŒãããããããŸã§ã®ãšãããæ£ç¢ºãªæž¬å®ãªãã®äžè¬çãªå°è±¡ã§ãïŒã Googleã«ã¯ã³ãŒãæé©åã«æºããå€ãã®ãªãœãŒã¹ãšããã°ã©ããŒãããŸãïŒ40åã®éçºè ã®ãªã¹ãã®TensorFlowã«é¢ããããŒãžã§ïŒããç§ã«ã¯ãã®ãããªæ©äŒã¯ãããŸãã-ããã¯ããŸããããŸãã
äžæ¹ãç§ã®ã©ã€ãã©ãªãŒã¯ã¹ããŒã¹ãã»ãšãã©å æãããã¢ãã䜿çšããŠããå Žåã¯Windowsããã³Linuxã§å®è¡ãããŸãã ç¹å®ã®ç¶æ³ã§ã¯ãããã¯ãã©ã¹ã«ãªãå ŽåããããŸãã
ãã¡ããçµæã«ã€ããŠã¯ã现å¿ã®æ³šæãæã£ãŠè§£éããå¿ èŠããããŸãã ãããã¯ç¹å®ã®ç¹æ®ãªã±ãŒã¹ã«é¢é£ããŠãããããã«ããããã¯ã©ã€ãã©ãªå šäœã®ç¹å®ã®ã¢ãã«ã®çµæã§ããããã®æ©èœã¯ã¯ããã«åºããªã£ãŠããŸãã ãããã£ãŠãã¢ãã«ãTensorFlowã«è»¢éãããšãçµæã¯åãã«ãªãããã¹ãŠãã¯ããã«é«éã«ãªããŸãã ãã®æå³ã§ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®æ£ããã¢ãŒããã¯ãã£ã®ç¥èã¯ãç¹å®ã®æè¡ã¹ã¿ãã¯ã®ç¥èãããéèŠã§ãã
çå®ã¯ãå²åŠçãªçåã®1ã€ã§ãã æåã«TensorFlowã«ã¢ã¯ã»ã¹ããå ŽåããŸãã¯åæ§ã®æ¢è£œã®ããŒã«ã䜿çšããå Žåãåãã¢ãã«ãäœæããŠåãçµæãåŸãããšãã§ããŸããïŒ ã·ã¹ãã ããã°ã©ãã³ã°ããŒãããè¡ãããšã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®åºæ¬ãããæ·±ãç解ããã®ã«åœ¹ç«ã¡ãŸããããããšãäœåãªæéã®ç¡é§ã§ããïŒ ããã©ãŒãã³ã¹ã®å¶éã¯ãæ°ããã¢ãã«ãéçºããããã®ã€ã³ã»ã³ãã£ãã§ããããããšãè¿·æãªé害ã§ããïŒ
ãããã«
誰ããèªåã§çµè«ãåºãããšããå§ãããŸãã ãããã®å®éšããç§ã®è³ªåã«å¯Ÿããæ確ãªåçããŸã åãåã£ãŠããŸãããèªè»¢è»ãçºæãã/ããªããšããåé¡ã¯æ®ã£ãŠããŸãããæ å ±ã¯éåžžã«æçãªããã§ãã
åç §è³æ
1. SutskeverãIlyaãOriol Vinyalsãããã³Quoc VV Leã ããã¥ãŒã©ã«ãããã¯ãŒã¯ã䜿çšããã·ãŒã±ã³ã¹åŠç¿ã®ã·ãŒã±ã³ã¹ãã ãã¥ãŒã©ã«æ å ±åŠçã·ã¹ãã ã®é²æ© ã 2014幎ã
2. BahdanauãDzmitryãChounghyun Choãããã³Yoshua Bengioã ã調æŽãšç¿»èš³ã®å ±ååŠç¿ã«ãããã¥ãŒã©ã«æ©æ¢°ç¿»èš³ã ArXiv preprint arXivïŒ1409.0473 ïŒ2014ïŒã
3. VinyalsãOriolãããã³Quoc Leã ãç¥çµäŒè©±ã¢ãã«ãã ArXiv preprint arXivïŒ1506.05869 ïŒ2015ïŒã
4.ãžã£ã³ãã»ãã¹ãã£ã³ãä» ããã¥ãŒã©ã«æ©æ¢°ç¿»èš³ã«éåžžã«å€§ããªã¿ãŒã²ããèªåœã䜿çšããããšã«ã€ããŠã ArXiv preprint arXivïŒ1412.2007 ïŒ2014ïŒã
PS
ãã®èšäºã«èšèŒãããŠãããã¹ãŠã®åæšã¯ãããããã®ææè ã®è²¡ç£ã§ãã