A simple approach for quantizing neural networks

TL;DR

This paper introduces a simple, deterministic method for quantizing neural network weights that preserves performance without hyper-parameter tuning, supported by theoretical guarantees and applicable to deep networks.

Contribution

A new straightforward quantization approach that simplifies analysis and guarantees performance preservation, unlike previous methods requiring hyper-parameter tuning.

Findings

The method preserves network performance after quantization.

Theoretical guarantees show error decay with more parameters.

Applicable to deep networks through layer-wise quantization.

Abstract

In this short note, we propose a new method for quantizing the weights of a fully trained neural network. A simple deterministic pre-processing step allows us to quantize network layers via memoryless scalar quantization while preserving the network performance on given training data. On one hand, the computational complexity of this pre-processing slightly exceeds that of state-of-the-art algorithms in the literature. On the other hand, our approach does not require any hyper-parameter tuning and, in contrast to previous methods, allows a plain analysis. We provide rigorous theoretical guarantees in the case of quantizing single network layers and show that the relative error decays with the number of parameters in the network if the training data behaves well, e.g., if it is sampled from suitable random distributions. The developed method also readily allows the quantization of deep networks by consecutive application to single layers.