Convolutional Networks - Case Studies

Convolutional Networks - Case Studies

1. Classic Networks

   • LeNet 5
     • Aim to n_h and n_w goes down, and number of filters increase
     • Conv_pool -> conv_pool -> FC -> FC -> output
     • LeCun et al: Gradient-based machine learning applied to document recognition (1998)
   • AlexNet
     • Named after Alex Krizhevsky
   • VGG-16

2. Residual Networks

   • Network of Residual Blocks; made by adding input variable into next layer of neural network...
   • Facilitate the possibiliy of adding a lot of levels into neural network. Skip Connection (input being being given to second layer) helps two added layers provide (at least) identity function of input, so that it does not loose on precision.

3. 1x1 Convolution

   • Used to shrink amount of filters, so 1x1x32 filter would create a same size matrix with n_c[2] amound of filters

4. Inception Network (GoogleNet)

   • Using multiple kinds of convolutions/pool at the same time in one single step
   • Used to lower the computational cost, but shrinking amount of filter using 1x1 convolution
   • Inception Module
     • Used to combine two technics mentioned bellow to quicker develop a new layer of network with lo computation cost

Images and files in the map