涓昏浜猴細瀹嬪＋鍚夋暀鎺?br />銆€銆€鏃墮棿錛?017騫?0鏈?8鏃?3錛?0
銆€銆€鍦扮偣錛氭満鐢?shù)瀛﹂櫌湓掍华鍗楁?07-2
銆€銆€嬈㈣繋騫垮ぇ甯堢敓鍙傚姞錛?/p>
銆€銆€銆€銆€銆€銆€銆€Topic: Stochastic Depth and Densely Connected Convolutional Networks
銆€銆€銆€銆€銆€銆€銆€銆€銆€銆€銆€銆€銆€銆€銆€銆€闅忔満娣卞害鍜岀瀵嗚仈緇撶殑鍗風(fēng)Н緗戠粶
銆€銆€
銆€銆€Abstract: Deep learning methods have dramatically improved the state-of-the-art in speech recognition, visual object recognition, object detection and many other domains. In particular, Convolutional Neural Networks (CNNs) were popularized within the vision community in 2009 through AlexNet and its celebrated victory at the ImageNet competition. In this talk, we will introduce two effective convolutional neural networks: Deep Networks with Stochastic Depth and Densely Connected Convolutional Networks (DenseNet). Stochastic Depth enables the seemingly contradictory setup to train short networks and use deep networks at test time. With stochastic depth we can increase the depth of residual networks even beyond 1200 layers and still yield meaningful improvements in test error (4.91% on CIFAR-10). DenseNet connects each layer to every other layer in a feed-forward fashion, which could alleviate the vanishing-gradient problem, strengthen feature propagation, encourage feature reuse, and substantially reduce the number of parameters. DenseNet obtains significant improvements over the state-of-the-art on most of object recognition benchmark tasks, whilst requiring less memory and computation to achieve high performance.
銆€銆€璁插駭浜虹畝浠嬶細
銆€銆€瀹嬪＋鍚? 鐢鳳紝1965騫?鏈堢敓錛屾竻鍗庡ぇ瀛﹁嚜鍔ㄥ寲緋繪暀鎺堛€佸崥澹敓瀵煎笀銆?996騫磋幏寰楀搱灝?dāng)婊ㄥ伐涓氬ぇ瀛︽暟瀛﹀崥澹浣嶅Q?996鑷?000騫達紝鍦ㄤ腑鍥芥搗媧嬪ぇ瀛︾墿鐞嗘搗媧嬩笓涓氥€佷笢鍗楀ぇ瀛︽帶鍒剁悊璁轟笌搴旂敤涓撲笟2嬈″畬鎴愬崥澹悗鐮旂┒銆?br />銆€銆€闀挎湡浠庝簨澶嶆潅鍒墮€犵郴緇熷緩妯′紭鍖栦笌鎺у埗鎶€鏈€佹満鍣ㄥ涔?fàn)涓庢晠闅滆瘖鏂€佹按涓嬫満鍣ㄤ漢鏅鴻兘鎺у埗絳夋柟鍚戠爺絀跺伐浣溿€傚湪鍥藉唴澶栭噸瑕佸鏈湡鍒婁細璁彂琛ㄨ鏂?00浣欑瘒錛屽叾涓璉EEE Transactions 緋誨垪鏈熷垔闀挎枃銆佸浗闄呰憲鍚嶄笓涓氭湡鍒奡CI媯€绱㈣鏂囪繎100綃囷紝鎷呬換銆奍EEE T SMC:Systems銆嬶紝銆婁腑鍥界瀛?淇℃伅縐戝銆嬪強銆婅嚜鍔ㄥ寲瀛︽姤銆嬬瓑鏈熷垔緙栧錛屻€婁漢宸ユ櫤鑳戒笌鏈哄櫒浜虹爺絀躲€嬪壇涓葷紪銆傝幏寰楁暀鑲查儴鑷劧縐戝浜岀瓑濂栧姳2欏癸紝榛戦緳姹熺敓鑷劧縐戝浜岀瓑濂?欏癸紝姹熻嫃鐪佽嚜鐒剁瀛︿竴絳夊1欏廣€?br />銆€銆€涓繪寔瀹屾垚浜嗗浗瀹惰嚜鐒剁瀛﹀熀閲戦噸澶х瀛︿華鍣ㄧ爺鍒墮」鐩€侀噸鐐歸」鐩€侀潰涓婇」鐩€佺鎶€閮ㄩ噸鐐逛笓欏廣€佷腑鍥藉ぇ媧嬪崗浼氫笓欏硅棰樼瓑30浣欓」銆?/p>
