StatQAT: Statistical Quantizer Optimization for Deep Networks

TL;DR

This paper introduces a statistical error analysis framework and novel quantizers for deep neural network quantization, improving accuracy and stability across various data distributions and formats.

Contribution

It presents a new theoretical analysis and iterative quantizers tailored for arbitrary and Gaussian-like data distributions, enhancing low-precision neural network training.

Findings

Improved accuracy in quantized neural networks.

Enhanced stability during training with new quantizers.

Effective quantization across multiple data formats.

Abstract

Quantization is essential for reducing the computational cost and memory usage of deep neural networks, enabling efficient inference on low-precision hardware. Despite the growing adoption of uniform and floating-point quantization schemes, selecting optimal quantization parameters remains a key challenge, particularly for diverse data distributions encountered during training and inference. This work presents a novel statistical error analysis framework for uniform and floating-point quantization, providing theoretical insight into error behavior across quantization configurations. Building on this analysis, we propose iterative quantizers designed for arbitrary data distributions and analytic quantizers tailored for Gaussian-like weight distributions. These methods enable efficient, low-error quantization suitable for both activations and weights. We incorporate our quantizers into quantization-aware training and evaluate them across integer and floating-point formats. Experiments demonstrate improved accuracy and stability, highlighting the effectiveness of our approach for training low-precision neural networks.