CNN Notes [001]

CNN convolutional Neural network key points:

{1} is a non-fully connected network (sparse network), compared to the BP neural network (fully connected network), of course, CNN can also have a portion of the layer is the full connection layer.

{2} weight sharing (the same weight factor), which is the same as the convolution kernel (location independent), 1 convolution cores. 1 convolution cores correspond to a feature map, and a feature extraction is performed to obtain
A feature extraction diagram, of course, a layer can be placed on multiple convolution cores.

The {3} convolution process is the weight and process, and the weight matrix is the convolution kernel.

{4} CNN is a supervised learning algorithm.

The {5} CNN layer can be divided into feature extraction layer (convolutional layer), Feature computing layer (sampling layer), input layer, full connection layer, and output layer.

{6} CNN has position, scale, deformation invariance.

The key to the {7} CNN algorithm is the training process, which is the learning process.

{8} parameters to be trained,
Convolution layer: (1) convolution kernel weight coefficient, (2) convolution core offset,
Sampling layer: (1) sample weight factor, (2) sample offset,

The {9} CNN algorithm was originally designed for two-dimensional handwritten images.

{Ten} The Parallel learning strategy of CNN training is mainly embodied in the independent parallel between different convolution cores in the same layer. Or a different strategy?

CNN Notes [001]