Learning Sparse Low-Precision Neural Networks With Learnable Regularization

TL;DR

This paper introduces a learnable regularization approach to train low-precision neural networks, improving accuracy and compression ratios by aligning high-precision weights with their quantized counterparts.

Contribution

It proposes a novel MSQE regularizer with a learnable coefficient and integrates weight pruning, quantization, and entropy coding for effective low-precision DNN compression.

Findings

Achieved state-of-the-art compression ratios of 7.13 and 6.79 on ImageNet with MobileNet and ShuffleNet.

Produced 8-bit low-precision models for super-resolution with negligible performance loss.

Enhanced training convergence and accuracy of low-precision neural networks.

Abstract

We consider learning deep neural networks (DNNs) that consist of low-precision weights and activations for efficient inference of fixed-point operations. In training low-precision networks, gradient descent in the backward pass is performed with high-precision weights while quantized low-precision weights and activations are used in the forward pass to calculate the loss function for training. Thus, the gradient descent becomes suboptimal, and accuracy loss follows. In order to reduce the mismatch in the forward and backward passes, we utilize mean squared quantization error (MSQE) regularization. In particular, we propose using a learnable regularization coefficient with the MSQE regularizer to reinforce the convergence of high-precision weights to their quantized values. We also investigate how partial L2 regularization can be employed for weight pruning in a similar manner. Finally, combining weight pruning, quantization, and entropy coding, we establish a low-precision DNN compression pipeline. In our experiments, the proposed method yields low-precision MobileNet and ShuffleNet models on ImageNet classification with the state-of-the-art compression ratios of 7.13 and 6.79, respectively. Moreover, we examine our method for image super resolution networks to produce 8-bit low-precision models at negligible performance loss.