Training and approximation methods for realizing computationally-cheaper binarized neural networks

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Binarized Convolutional Neural Network (BCNN) is a promising technology to realize fast and energy-efficient inference in various applications such as robots and virtual reality (VR) in edge AI devices. A low-bit quantized CNN approximation method using random forest (RF) models has been proposed, which can significantly reduce the number of required multiply-accumulate (MAC) operations and the area of a BCNN accelerator. However, the accuracy reduction is a problem. In this paper, we propose two methods to realize computationally-cheaper BCNN models while maintaining sufficient accuracy with random forest approximation (RFA). The first method is approximate-aware training for BCNN models. The second method is a channel-wise RFA method based on the imbalance rate of the output feature map’s values. We demonstrate that for binarized VGG16, the proposed approximate-aware training enables the approximation of intermediate layers that are computationally heavy and achieves 13.9% reduction in the computational cost of feature extraction by RFA while maintaining the original inference accuracy. Moreover, we demonstrate that channel-wise RFA approximation can select channels that are more suitable for RFA approximation compared with channel selection based on the output prediction accuracy of random forest models.

This publication has 10 references indexed in Scilit: