电子科技大学研究生《机器学习》精品课程 第14讲深度CNN 14 Deep CNN 郝家胜(Jiasheng Hao) Ph.D.,Associate Professor Email:hao@uestc.edu.cn School of Automation Engineering University of Electronic Science and Technology of China,Chengdu 611731 Aug.2015第一稿;Apr2021第四稿
电子科技大学研究生《机器学习》精品课程 Email: hao@uestc.edu.cn School of Automation Engineering University of Electronic Science and Technology of China, Chengdu 611731 郝家胜 (Jiasheng Hao) Ph.D., Associate Professor Aug. 2015 第一稿;Apr. 2021 第四稿 第14讲 深度CNN 14 Deep CNN
LeNet:( CNN C3:f.maps 16@10x10 INPUT C1:feature maps S4:f.maps 16@5x5 6@28x28 32x32 S2:f.maps 6@14x14 C5:layer F6:layer 120 OUTPUT 84 10 Full connection Gaussian connections Convolutions Subsampling Convolutions Subsampling Full connection CNNs are basically layers of convolutions followed by su bsampling and fully connected layers. 口三大关键思想 ■ 局部感受野(卷积操作) ■ 权值共享 ■ 池化 2
LeNet:CNN 2 CNNs are basically layers of convolutions followed by su bsampling and fully connected layers. o 三大关键思想 n 局部感受野(卷积操作) n 权值共享 n 池化
CNNs in ilsvro Participation and Performance 0.28 0.66 5 0.03 35 0.23 29 2010 2011 2012 2013 2014 2015 2016 Number of Classification Average Precision Entries Errors(top-5) For Object Detection 3
CNNs in ILSVRC 3
What can CNN do perso car person dog norse 0.5 02 城市道 0.3 9, 050 nx k representation of Comvolutional layer with 0515225335445 sentence with static and multiple finer widths and non-stalic channels eature map到 x10
What can CNN do
What can CNN do
5 What can CNN do
Outline l:What does CNN Learn? ll:How does CNN go deep? lll:Variants of CNN
Outline III: Variants of CNN II: How does CNN go deep? I: What does CNN Learn?
The whole cat dog … Convolution Max Pooling Can repeat Fully Connected many times Feedforward network Convolution Max Pooling Flatten
The whole CNN Fully Connected Feedforward network cat dog …… Convolution Max Pooling Convolution Max Pooling Flatten Can repeat many times
The whole Property 1 >Some patterns are much Convolution smaller than the whole image Property 2 The same patterns appear in Max Pooling Can repeat different regions. many times Property 3 Convolution > Subsampling the pixels will not change the object Max Pooling Flatten
The whole CNN Convolution Max Pooling Convolution Max Pooling Flatten Can repeat many times Ø Some patterns are much smaller than the whole image ØThe same patterns appear in different regions. ØSubsampling the pixels will not change the object Property 1 Property 2 Property 3
Three Steps for Deep Learning Steo 1 Step 3. define a Step 2: pick the Convolutional goodness best Neural Network of function function Deep Learning is so simple ..... 1 2 3 CDC.TENCENT.COM
Step 1: define a set of function Step 2: goodness of function Step 3: pick the best function Three Steps for Deep Learning Deep Learning is so simple …… Convolutional Neural Network
Why CNN for ImageB..ECCv2014 。00。。0 000.00 :: : Represented as pixels The most basic Use 1st layer as module Use 2nd layer as classifiers to build classifiers module… Can the network be simplified by considering the properties of images?
Why CNN for Image? Can the network be simplified by considering the properties of images? 1 x 2 x … … Nx … … … … … … …… …… …… The most basic classifiers Use 1st layer as module to build classifiers Use 2nd layer as module …… [Zeiler, M. D., ECCV 2014] Represented as pixels