Compression of Deep Neural Networks with Structured Sparse Ternary Coding

Abstract

Deep neural networks (DNNs) contain large number of weights, and usually require many off-chip memory accesses for inference. Weight size compression is a major requirement for on-chip memory based implementation of DNNs, which not only increases inference speed but also reduces power consumption. We propose a weight compression method for deep neural networks by combining pruning and quantization. The proposed method allows weights to have values of + 1 or − 1 only at predetermined positions. Then, a look-up table stores all possible combinations of sub-vectors of weight matrices. Encoding and decoding structured sparse weights can be conducted easily with the table. This method not only allows multiplication-free DNN implementations but also compresses the weight storage by as much as x32 times more than that in floating-point networks and with only a tiny performance loss. Weight distribution normalization and gradual pruning techniques are applied to lower performance degradation. Experiments are conducted with fully connected DNNs and convolutional neural networks.