çé æºä¹ (Tomoyuki Mano) <tomoyukimano@gmail.com> version 1.0, 2020-06-19
çé æºä¹ (Tomoyuki Mano) <tomoyukimano@gmail.com> version 1.0, 2020-06-19
æ¥ç«è£½ä½æ ç 究éçºã°ã«ã¼ã èªåé転ç 究é¨ã¯ãGPUã«ããCNNï¼Convolutional Neural Networkï¼ç³ã¿è¾¼ã¿ãã¥ã¼ã©ã«ãããã¯ã¼ã¯ï¼å¦çã«ããã¦æå¾ ãããããªæ§è½ãåºãªãçç±ããã®å¯¾å¦æ¹æ³ã«ã¤ãã¦ããDAã·ã³ãã¸ã¦ã 2022 âã·ã¹ãã ã¨LSIã®è¨è¨æè¡âãï¼æ å ±å¦çå¦ä¼ ã·ã¹ãã ã¨LSIã®è¨è¨æè¡ç 究ä¼(SLDM)ã2022å¹´8æ31æ¥ï½9æ2æ¥ã«éå¬ï¼ã§çºè¡¨ãããCNNå¦çã¯ãADASï¼å é²é転æ¯æ´ã·ã¹ãã ï¼ãèªåé転ã«ãããç»åèªèã§ãã使ãããæ¼ç®ã§ããã çºè¡¨ããæ¥ç«ã®å³¶æå 太éæ°ã«ããã°ãCNNå¦çãGPUã§å®è¡ããã¨ããã®æ¼ç®æ§è½ãGPUã¡ã¼ã«ã¼çºè¡¨ã®ãã¼ã¯æ§è½ããæ¡éãã«ä½ãå ´åãããã¨ããï¼å³1ï¼ãä¾ãã°ããã¼ã¿ã»ã³ã¿ã¼ã§ã®æ¨è«å¦çã§ãã使ããã¦ããç±³NVIDIAï¼ã¨ãããã£ã¢ï¼ã®GPUã«ã¼ããTesla T4ãã®16ãããæµ®åå°æ°ç¹æ¼ç®ï¼FP1
Generate unique, expressive AI-generated faces in real time
æ°ããªæ°å¦ã®å®çã®çºè¦ããæªè¨¼æã®äºæ³ã®è§£æ±ºã«AIãå½¹ç«ã¤ââãããªç 究çµæããå²ç¢AIãAlphaGoããªã©ã§ç¥ãããè±DeepMindãçºè¡¨ãããé åã«é¢ããæ°ããå®çãçºè¦ããä»ãã²ãã®çµã³ç®ãæ°å¦çã«ç 究ãããçµã³ç®çè«ãã«ã¤ãã¦ããç°ãªãæ°å¦ã®åéãã¤ãªããäºæ³ãã¦ããªãã£ãé¢ä¿æ§ãè¦ã¤ããã¨ããã DeepMindã¯ã豪ã·ããã¼å¤§å¦ã¨è±ãªãã¯ã¹ãã©ã¼ã大å¦ã®æ°å¦è ã¨ã¨ãã«æ°å¦ç 究ãæ¯æ´ããããã®æ©æ¢°å¦ç¿ãã¬ã¼ã ã¯ã¼ã¯ãæ§ç¯ãããã¾ã§ãæ°å¦è ã¯ãç 究対象ã調ã¹ãããã«ã³ã³ãã¥ã¼ã¿ã使ãããã¾ãã¾ãªãã¿ã¼ã³ãçæãããã¨ã§çºè¦ã«å½¹ç«ã¦ã¦ãããããã®ãã¿ã¼ã³ã®æ義ã¯æ°å¦è èªèº«ãèå¯ãã¦ãããããããç 究対象ã«ãã£ã¦ã¯ä½åãã®æ¬¡å ããããã¨ããã人éã«ããèå¯ãéçããã£ãã ä»åéçºããã¢ã«ã´ãªãºã ã¯ããããããã¿ã¼ã³ãæ¤ç´¢ããä»ãæ師ããå¦ç¿ãåºã«ãã®æå³ãç解ãããã¨è©¦ã¿ãã¨ãã
ã¯ããã« å»å¹´ããéåã³ã³ãã¥ã¼ã¿ã®åå¼·ä¼ã«åå ããããã«ãªã£ãã®ã§ãããããã§ã¨ãããããã¦ããéåGANã®ã話ãé£ãããªãããã¨ã¦ãé¢ç½ãã£ãã®ã§ãè«æãèªãã§ã¾ã¨ãããã¨ã«ãã¾ããã ã¾ããéåGANå®è£ ã®æ¹æ³ãæ¸ãã¦ãã£ãã®ã§ãMDR社ã®blueqatã¨pytorchã§ã³ã¼ããæ¸ãã¦ã¿ããã¨ã«ãã¾ããã è«æã®ãªã³ã¯ââ https://arxiv.org/pdf/1807.01235.pdf è«æã®è¦æ¨ GANã¯Discriminatorã¨Generatorã交äºã«ç«¶åãããªããå¦ç¿ãé²ãããã¨ã§ãæ¬ç©ã«è¿ããã¼ã¿ã®çæãå¯è½ã«ãããããã¯ã¼ã¯ã§ããæ¬è«æã§ã¯Discriminatorã¯ãã®ã¾ã¾ã«ãã¦ãGeneratorãéåã²ã¼ãã§ããããã¦ãã¾ããããã«ããæ¢åã®GANããéåæ©æ¢°å¦ç¿ã¨æ¯ã¹ã¦ããã¤ãã®ç¹ã§æç¨ã ããã§ãã å¤å ¸çãªGANã§ã¯å¾é æ¶å¤±åé¡ã«ãããé¢æ£çãªå¤ãç
KARTEã«ãããããã³ãã¨ã³ãã¨é£æºãããKubeflow Pipelineã®æ´»ç¨æ¹æ³ã«ã¤ãã¦è§£èª¬ãã¾ã
以åã社å ã®åå¼·ä¼ç¨ã®è³æã¨ãã¦ä½æããã確ççµ±è¨-æ©æ¢°å¦ç¿ãã®åã«ãã¨ããè³æãä»åå¤§å¹ ã«æ¹å®ãã¦å ¬éãã¾ããã æ¹å®ã®ä¸çªã®ãã¤ã³ãã¯ãçµ±è¨åæã®ç·è¦§çãªã¬ã¤ãã¨ãã¦ä½¿ããããã«ãããã¨ã§ããæ£ç´ãªè©±ãå 容çã«ç§èªèº«ãååã«ç解åºæ¥ã¦ããªãé¨åãå¤ãã®ã§ããããããã調ã¹ãçµæããªãã¨ãªãçµ±è¨åæã®ææ³ãããçãªãã®ãé ã®ä¸ã«åºæ¥ä¸ãã£ã¦ããã®ã§ããããã¢ã¦ããããã¨ãã¦ã¾ã¨ãã¦ã¿ã¾ããã 確ççµ±è¨ï¼æ©æ¢°å¦ç¿ãã®åã« v2.0 from Hidekatsu Izuno 確ççµ±è¨ã«éã£ã話ã§ã¯ããã¾ããããæ°ããåéãå¦ã¶åå¦è ã«ã¨ã£ã¦ãå ¨ä½åãããããããªãããæ··ä¹±ãããã¨ãå°ãªãããã¾ããããã®è³æãèªããã¨ã§ãªãã¨ãªãã§ãå ¨ä½ãææ¡ã§ãã¦ããã°ãä»ã®æç®ãèªãéã«ãç解ã容æã«ãªãã®ã§ã¯ã¨æã£ã¦ãã¾ãã ä¾ã«ãã£ã¦ã確ççµ±è¨ã«ã¤ãã¦ã¯å®å ¨ã«ç´ 人ãªã®ã§ãééããããããããã¾ãããã
ã³ã³ãã¥ã¼ã¿ã¼ã«ããæ©æ¢°å¦ç¿ã®æè¡ãç¨ãããã¨ã§ãå¾æ¥ã¨ã¯å ¨ãç°ãªãé³ä½ããå¯è½ã«ããã·ã³ã»ãµã¤ã¶ã¼ãNSynth Superãã®éçºãGoogleãé²ãã¦ãã¾ãããã®ã·ã³ã»ãµã¤ã¶ã¼ã¯ããã¾ãã¾ãªé³è²ãæã¤åºæã®ç¹å¾´ãã³ã³ãã¥ã¼ã¿ã¼ãæ©æ¢°å¦ç¿ã§ç解ãããã¨ãã°ããã«ã¼ããã¨ãã¹ãã¢ãã©ã ãã®ç¹å¾´ãæãåãããå ¨ãæ°ããé³ãçã¿åºããã¨ãå¯è½ã«ãã¦ãã¾ãã NSynth Super https://nsynthsuper.withgoogle.com/ Making music using new sounds generated with machine learning https://www.blog.google/topics/machine-learning/making-music-using-new-sounds-generated-machine-learning/ NSyn
ã¯ããã« å ¨çµå層 å ¨çµå層ã®æ°å¼ éå ¨çµå層 ç³ã¿è¾¼ã¿å±¤ å ¥åæåã®indexã«çç®ããéå ¨çµå層 éå ¨çµåã»éã¿å ±æ層 1Dç³ã¿è¾¼ã¿å±¤ 2Dã®ç³ã¿è¾¼ã¿å±¤ æå¾ã« ã¯ããã« ä»åã¯ãã¥ã¼ã©ã«ãããã¯ã¼ã¯ã®åºæ¬çãªæ§æè¦ç´ ã§ããç·å½¢ï¼å ¨çµåï¼å±¤ã¨ç³ã¿è¾¼ã¿å±¤ã«ã¤ãã¦ç°¡åã«èª¬æãã¾ãã ããã説æããã¢ããã¼ã·ã§ã³ã¯ããç³ã¿è¾¼ã¿å±¤ã¯ç·å½¢å±¤ãããåªããçºå±çææ³ãè¤éãªææ³ã§ããã¨ããåéãã åå¦è ã®ä¸ã§ããã®ã§ã¯ãªããã¨æããããã®è§£æ¶ãããããã§ãã 以éãç³ã¿è¾¼ã¿å±¤ã¨ç·å½¢å±¤ã®æããéãã主張ããããã«ãç·å½¢å±¤ã®ãã¨ãå ¨çµå層ã¨å¼ã¶ãã¨ã«ãã¾ãã ãã®å¼ã³åã¯ãTensorFlowãKerasãªã©ã®ãã¬ã¼ã ã¯ã¼ã¯ã§æ¡ç¨ããã¦ãã¾ãï¼layers.denseï¼ã å ¨çµå層 å ¨çµå層ã®æ°å¼ ã¾ãå ¨çµå層ã«ã¤ãã¦æ°å¼ã以ä¸ã«ç¤ºãã¾ãã å ¥åãã¯ãã«$x \in \mathbb R^{D}$ã«å¯¾ã
#æ¦è¦ æ©æ¢°å¦ç¿ããã¸ã§ã¯ãã«é¢ãã¦ã®èªåç¨ã®åå¿é²ã§ããã¾ã ãããã¸ã§ã¯ãèªä½ã«ã¢ãµã¤ã³ãããã°ãããªã®ã§ãã¨ãããã調ã¹ç©ã®éä¸çµéçãªãªã¹ãã§ãã ããã§ãçããã®å¦ç¿ãããã¸ã§ã¯ãã®åèã«ãªãã°å¹¸ãã§ãããã®ãã¡ããããæç®ã®ã¾ã¨ãè¨äºãä½æãããã¨èãã¦ãã¾ãã KDNuggetsï¼è±æï¼ æ©æ¢°å¦ç¿ããã£ã¼ãã©ã¼ãã³ã°ãªã©ãã¼ã¿ãµã¤ã¨ã³ã¹ã«é¢ããè¨äºãã¾ã¨ãããã¦ãããµã¤ããææ°ã®ãã£ã¼ãã©ã¼ãã³ã°ã®ãã¬ã¼ã ã¯ã¼ã¯ã®ååã«ã¤ãã¦ã®è¨äºã ã£ãããIoTã¤ãã³ãæ å ±ãè¼ã£ã¦ããã§æ å ±è±å¯ãæ¯æ¥ãã§ãã¯ãã¦ãã¾ãã Advice For New and Junior Data Scientists ï¼è±æï¼ TwitterããAirbnbã¸ã¨ãã£ãªã¢ãã§ã³ã¸ãããRobert Changæ°ã®è¨äºãæ©æ¢°å¦ç¿ããã¸ã§ã¯ãã®é£ããç¹ãè¨èªã®å¦ç¿ã«ã¤ãã¦ãç¾å ´ãµã¤ãããã¢ãã´ã¡ã¤ã¹ãã¦ãã¾ãã
Deleted articles cannot be recovered. Draft of this article would be also deleted. Are you sure you want to delete this article? ç¾å¨AIã«å¯¾ãã¦å¤ãã®äººããã¤ã¤ã¡ã¼ã¸ã¨ãã¦ãOver Estimate(éå°ãªæå¾ ) ã¨Under Estimate(éå°ãªæå¾ )ã¨ããåé¡ãããã¾ããOver Estimate(éå°ãªæå¾ )ã¯AIã使ãã¨å ¨ã¦ã®åé¡ãèªåçã«è§£æ±ºãããã¨ãããã®ã§ãUnder Estimate(éå°ãªæå¾ )ã¯AIã使ãã°å®ã¯ç°¡åã«è§£æ±ºãããåé¡ãªã®ã«ãã®ãã¨ã«æ°ã¥ãã¦ãããªãã¨ãããã¨ã§ãã ä»æ¥ã§ã¯ãæ¯è¼çç°¡åã«AIã使ããã¨ã®ã§ãããã¼ã«ãããããããã®ã§ãããããæ©ä¼ãå©ç¨ãã¦ãèªåãã¡ã®ãã¼ã¿ããã¨ã«ã©ãã©ã使ã£ã¦ã¿ã¦ãAIã使ãã¨å®éã«ä½ã
(編注ï¼2020/10/01ã2016/07/29ãããã ãããã£ã¼ãããã¯ããã¨ã«è¨äºãä¿®æ£ãããã¾ããã) ç®æ¬¡ï¼ ãã¾ãã¾ãªå¾é éä¸æ³ ãããå¾é éä¸æ³ 確ççå¾é éä¸æ³ ãããããå¾é éä¸æ³ èª²é¡ å¾é éä¸æ³ãæé©åããã¢ã«ã´ãªãºã Momentum(æ £æ§) Nesterovã®å éå¾é éä¸æ³ Adagrad Adadelta RMSprop Adam ã¢ã«ã´ãªãºã ã®å¯è¦å ã©ã®ãªããã£ãã¤ã¶ãé¸ã¶ã¹ãï¼ SGDã®ä¸¦ååã¨åæ£å Hogwild! Downpour SGD SGDã®ããã®é 延èæ§ã¢ã«ã´ãªãºã TensorFlow Elastic Averaging SGD æé©åãããSGDã«å¯¾ããæ´ãªãæ¦ç¥ ã·ã£ããã«å¦ç¿ã¨ã«ãªãã¥ã©ã å¦ç¿ ãããæ£è¦å æ©æçµäº å¾é ãã¤ãº çµè« åèæç® å¾é éä¸æ³ã¯ãæé©åã®ããã®æãç¥ãããã¢ã«ã´ãªãºã ã®1ã¤ã§ããããã¾ã§ã¯ãã¥ã¼ã©ã«ããã
åå¼·ãã¦ããã¨ãããªã®ã§ãééã£ã¦ããç®æãããããããã¾ãã. æ©æ¢°å¦ç¿ã¯ãé«åº¦ãªæ°å¦ç¥èã使ã£ã¦ãå¤æ°ã®ãã¼ã¿ãããæç¨ãªç¹å¾´ãç®åºããããã¤ãã®ã¢ã«ã´ãªãºã ã使ã£ãææ³ã®ãã¨ããã. ä¾ãã°ããã¥ã¼ã©ã«ãããã¯ã¼ã¯ã使ã£ã深層å¦ç¿ã¯ãæ©æ¢°å¦ç¿ã®ä¸åéã§ãã. éåç¥ããã°ã©ãã³ã°ã¯ããããªæ©æ¢°å¦ç¿ã®ãµã³ãã«ãããããããæ¸ç±ã ã£ããã©ããããããã«å¤ãå 容ãã. anacondaanacondaã¯ãPythonã®ä¸»è¦ãªã©ã¤ãã©ãªãããã±ã¼ã¸ãããªã¼ãã³ãã¼ã¿ãµã¤ã¨ã³ã¹ã®ãã©ãããã©ã¼ã . ã ããããããã¤ã³ã¹ãã¼ã«ãã¦ããã°ãç°å¢ãæ´ã. Anaconda jupyterjupyterã¯ãã¤ã³ã¿ã©ã¯ãã£ããªããã°ã©ãã³ã°è¨èªã®å®è¡ç°å¢. PythonãRãªã©ã®ã¤ã³ã¿ã©ã¯ãã£ãå®è¡ç°å¢Jupyter Notebookãã¤ã³ã¹ãã¼ã«ãã
2. éå»ã®çºè¡¨ 2014å¹´11æ29æ¥ TokyoWebMining #40 2 å°éã¨ä¸åé¢ä¿ãã· 2chããã¹ããã¤ãã³ã°ã¨ã¾ã¨ããµã¤ãã®èªåçæ ã»ã¯ã·ã¼å¥³åªã§å¦ã¶ç»ååé¡å ¥é 3. æå± èªå·±ç´¹ä» 2014å¹´11æ29æ¥ TokyoWebMining #40 3 Twitter ID ï½ï½ï½ï¼ï¼ï¼ï¼ å°é çµå¶å·¥å¦/æé©å æãã¼ã¿åæä¼ç¤¾ æ¥å åæä½ã§ãå±ãã æ©æ¢°å¦ç¿ã¨ã®åºä¼ã å½æã®ç 究ãå®ç¨æ§ çç¡ ç²¾ç¥ã®éãéã¨ã㦠æ©æ¢°å¦ç¿ ãéå§ ç 究ã è©°ãã§ã 趣å³ãæ¬è·ã« è¨èªãç»åã¨å¹ åºã éãã§ã¾ã
9. èæ¯ ï¼ï¼ï¼å¤§è¦æ¨¡ãå°è¦æ¨¡ãªã²ã¼ã éçºã«ããããããã¾ãåæ¥ãã¤ã³ ãã£ã¼ãºãåãããã²ã¼ã ã¨ã³ã¸ã³ã®ä¸ã§éçºãããããã¨ãé常åã ã¦ããï¼ï¼Ex.ï¼Unity3D, Unreal Engine 4 ï¼ï¼ï¼ã²ã¼ã ä¼ç¤¾ããªãªã¸ãã«ã®ã²ã¼ã ã¨ã³ã¸ã³ãä½ãå ´åãããã°ã å¸è²©ã®ã²ã¼ã ã¨ã³ã¸ã³ãå§åçãªã·ã§ã¢ã伸ã°ãã¦ããã ä»çµ¦é» é(2016) ã²ã¼ã ã¨ã³ã¸ã³ã®æ´å²æ¦è¦ æ¥æ¬ãã¸ã¿ã«ã²ã¼ã å¦ä¼å¤§ä¼äºç¨¿é pp. 87-90. ï¼ï¼ï¼ã²ã¼ã ã¨ã³ã¸ã³ã®æ´å²ãè¨è¨ã«ã¤ãã¦ãç 究ãé²ãããã¦ãã ã¸ã§ã¤ã½ã³ã»ã°ã¬ã´ãªã¼(2015) ã²ã¼ã ã¨ã³ã¸ã³ã»ã¢ã¼ããã¯ã㣠SBã¯ãªã¨ã¤ãã£ã ä¸å® é½ä¸é(2016) ã¿ã¹ã¯ã·ã¹ãã ã®èµ·æºã«ã¤ã㦠æ¥æ¬ãã¸ã¿ã«ã²ã¼ã å¦ä¼å¤§ä¼äºç¨¿é pp. 83-86. 10. èæ¯ ï¼ï¼ï¼éçºã¯ã¼ã¯ããã¼ã«ããã¦äººå·¥ç¥è½ãå°å ¥ããä¾ããã¨ãã°å質 管çã«äººå·¥ç¥è½ãå°å ¥ãã
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}}