首先将前面的图中卷积核和特征图调换一下位置, 将特征图矩阵的元素替换为

First, change the position of the convolution kernel and the feature map in the previous figure, replace the elements of the feature map matrix with δ, and replace the elements of the convolution kernel matrix with the partial derivative of E to w. According to the partial derivative formula obtained above, It can be found that the convolution of the feature map to the input matrix can obtain the yellow partial derivative matrix on the right. That is

First, replace the convolution core and feature diagram in the previous diagram, replace the elements of the feature map matrix with , and replace the elements of the convolution core matrix with the bias of E to w. According to the partialization formula obtained earlier, it can be found that the pattern-to- input matrix convolution can obtain the yellow bias matrix on the right. That is,

Firstly, the convolution kernel and the eigengraph in the previous graph are replaced by a position, the elements of the eigengraph matrix are replaced by δ, and the elements of the convolution kernel matrix are replaced by the partial derivative of e to W. according to the partial derivative formula obtained previously, it can be found that the partial derivative matrix of the right yellow color can be obtained by the convolution of the eigengraph to the input matrix<br>