Customization of structured neural network weight pruning method

Abstract

Weight pruning is a widely used model optimization technique for reducing the number of weights in neural networks. Weights represent the magnitudes of connections between neurons, and their values determine how much they contribute to the output of a neural network layer. This thesis provides an overview of weight pruning and introduces a custom weight pruning method, which uses a restricted number of patterns for sparsity. The proposed structured pruning enables model acceleration on a specific digital signal processor.

This work briefly evaluates the custom pruning method using a feed-forward model trained on the MNIST handwritten digits dataset. The custom pruning method effectively reduces half of the weights from each fully-connected layer by using specific patterns. The accuracy of the test model, which uses the custom pruning method, does not differ significantly from that of the model with the baseline structured pruning method and the original fully-connected test model.