Deep convolutional networks usually require a pooling operation so that the spatial size of each successive layer decreases. The strategy behind probabilistic max pooling is to constrain the detector units so at most one may be active at a time.