Jack Dermody


Convolutional Neural Networks

Max pooling is like a convolutional layer, but rather than sliding a filter around and multiplying the input images by the filter weights, max pooling layers take the highest pixel value in each filter and passes it to the next layer. The expected data table format is a 3D tensor column with the image data (a 3D tensor is a list of matrices, one for each input image channel (red, green blue) or in this case since MNIST is black and white, a list with a single matrix for the black and white pixel values) followed by the output vector column. The other pixels will simply be dropped from consideration. In this case the box of four pixels will measure every pixel once in the image without overlap and only one in four of the input pixels will be preserved in the output image.