Channel-Wise Activation Map Pruning using MaxPool for Reducing Memory Accesses

While neural network pruning can reduce the amount of data transfers, pruning techniques such as fine-grained pruning cannot be efficiently implemented as indexing overhead is high. In order to design a hardware friendly pruning technique, structured pruning, which removes groups of data together to minimize the indexing overhead, is highly required. In this paper, to reduce the overall number of main memory accesses when implementing convolutional neural network (CNN) accelerators, we propose a hardware friendly L2-norm based structured channel-wise activation pruning using max-pooling. Max-pooling is typically deployed in CNN to decrease the height and width of CNN activation maps. The simulation results of the proposed technique shows that the activation maps of ResNet20 and ResNet56 can be reduced to 48% and 51%, respectively, with less than 1% accuracy degradation on CIFAR-10.