The previous paragraph, convolution can extract features, but for the real world among the massive gallery, we do not know which local features effectively, we still hope to train the neural network, auto-learning out how to do it? BP had to use the algorithm learned earlier, but now the problem is the convolution of the neural network and what does it matter? Look at the following two charts show that, in fact, the convolution operation is the summation after multiplying, and neural network effect is the same. And convolution kernel convolution result correspond to the parameters and the results of the hidden layer in the neural network. This return to BP algorithm you learned earlier, the approach is the same, the first initialization parameters, and through the training so that the error getting smaller and smaller.
Reprinted from the original article: https://blog.csdn.net/qq_43650923/article/details/100630276