Análise de técnicas de compressão em redes neurais profundas por poda em dataset de imagens

Abstract

Deep neural network techniques have been successfully used for image classification using convolutional neural networks. However, deep learning algorithms perform a lot of mathematical operations. These operations can be a bottleneck in the process of large amounts of images. In low-cost microcontrollers, these operations can result in a significant increase in energy consumption, showing the need to apply compression techniques for these networks. Currently, most of the deep networks used or image classification are not optimized. The purpose of this work is to optimize a convolutional neural network using the technique of data compression by pruning. During training, the technique is to remove the weights at each batch, instead of removing weights only in the first batch of each epoch. This strategy was applied to classify 10,000 images from 10 different classes. It was possible to remove approximately 82% of the parameters from the deep neural network while maintaining high accuracy. These results shows that the batch weight removal technique proved to be effective for this application.