CNN
1. Overall architecture
It is my first time to learn the CNN, which is hot field in machine learning. I will try to speak detail of CNN.
> Q1: what is the channel?
The channel, which is the number of the input for special layer, could be image or feature map. For example(1), The number of channel is three in the first layer of CNN if the network is used to deal with RGB picture. (2) In the hidden layers, if the number of output of K layer is 32, then the number of channel of K+1 layer is previous output of K layer, that is 32.
> How the feature maps in k layer are performed by convolution layer?
When performing the input features, the convolutional network deal with the patches from the same area of all previous layers.
> what is the convolution kernel?
Generally, the convolution kernel is be selected from the point that it is similar to the mechanism of brain which recognize the object feature. Based on the analysis, the Sobel or Laplace operator are used as convolution kernel.
>what is the parameter sharing
Units organized into the same “feature map” share parameters
In experiment, the distribute of parameter follows the Guussian or Xarvier distribution.
> what is the feature map?
For CNN, one convolution kernel means one feature map that is convolution kernel perform input data or feature maps to get certain feature map.
where every element of the set (64 64 256 256) is the number of feature maps in currency layer.
> what is the local connectivity
Each hidden layer is only connected with the subregion (patch) of the input image. Besides, it is connected to all channels.
>what is the pooling
The pooling is performed on the non-overlapping neighborhoods. However, there will has overlapping neighborhoods.
> what is the active function?