Misha Orel
Sep 8, 2018 · 1 min read

Each one has its own purpose. Pooling, be it average or max, is designed to erase any information about positions of the features within the kernel. On the other hand, convolution pooling net will learn this information. The result is like this: for a classifier, the net with convolutions would require more training data to generalize over positions within the kernel; for an autoencoder, like in your case, where position does matter anyway, convolution would probably be a strictly better approach.

    Misha Orel

    Written by