Stochastic Precision Ensemble: Self-Knowledge Distillation for Quantized Deep Neural Networks

TL;DR

This paper introduces SPEQ, a novel stochastic ensemble knowledge distillation method for quantized deep neural networks that improves performance by using a dynamically changing teacher formed from the student model itself.

Contribution

SPEQ employs stochastic bit-precision variation to create a self-knowledge distillation scheme, eliminating the need for separate teacher networks and enhancing quantized model accuracy.

Findings

Outperforms existing quantization methods in image classification.

Effective in question-answering and transfer learning tasks.

Reduces activation quantization noise through stochastic soft labels.

Abstract

The quantization of deep neural networks (QDNNs) has been actively studied for deployment in edge devices. Recent studies employ the knowledge distillation (KD) method to improve the performance of quantized networks. In this study, we propose stochastic precision ensemble training for QDNNs (SPEQ). SPEQ is a knowledge distillation training scheme; however, the teacher is formed by sharing the model parameters of the student network. We obtain the soft labels of the teacher by changing the bit precision of the activation stochastically at each layer of the forward-pass computation. The student model is trained with these soft labels to reduce the activation quantization noise. The cosine similarity loss is employed, instead of the KL-divergence, for KD training. As the teacher model changes continuously by random bit-precision assignment, it exploits the effect of stochastic ensemble KD. SPEQ outperforms the existing quantization training methods in various tasks, such as image classification, question-answering, and transfer learning without the need for cumbersome teacher networks.