[Paper Review] Post-training 4-bit quantization of convolution networks for rapid-deployment

Ron Banner, Yury Nahshan+2 more|arXiv (Cornell University)|Oct 2, 2018

Advanced Neural Network Applications16 references125 citations

TL;DR

Presents a practical 4-bit post-training quantization framework for CNNs without retraining, using Analytical Clipping for Integer Quantization (ACIQ), per-channel bit allocation, and bias-correction to preserve accuracy.

ABSTRACT

Convolutional neural networks require significant memory bandwidth and storage for intermediate computations, apart from substantial computing resources. Neural network quantization has significant benefits in reducing the amount of intermediate results, but it often requires the full datasets and time-consuming fine tuning to recover the accuracy lost after quantization. This paper introduces the first practical 4-bit post training quantization approach: it does not involve training the quantized model (fine-tuning), nor it requires the availability of the full dataset. We target the quantization of both activations and weights and suggest three complementary methods for minimizing quantization error at the tensor level, two of whom obtain a closed-form analytical solution. Combining these methods, our approach achieves accuracy that is just a few percents less the state-of-the-art baseline across a wide range of convolutional models. The source code to replicate all experiments is available on GitHub: \url{https://github.com/submission2019/cnn-quantization}.

Motivation & Objective

• Motivate rapid deployment of low-bit quantized CNNs without access to full training data.
• Minimize quantization error at the tensor level through analytical clipping and channel-aware bit-width selection.
• Enable accurate 4-bit quantization of both activations and weights without re-training.
• Provide bias correction to mitigate quantization-induced bias in weights.

Proposed method

• Introduce ACIQ: analytically determine clipping threshold to minimize mean-squared error for activations.
• Propose per-channel bit allocation to assign optimal bit-width per channel under a fixed average bit-budget.
• Apply bias-correction to compensate for quantization bias in weights.
• Quantize both weights and activations using the proposed methods in a joint deployment pipeline.
• Use a per-channel quantization scheme with fused ReLU to reduce noise.
• Demonstrate that combining these methods restores most degradation without fine-tuning.

Experimental results

Research questions

• RQ1 Can 4-bit post-training quantization approach floating-point accuracy closely for CNNs without using the full training dataset?
• RQ2 How much do analytical clipping, per-channel bit allocation, and bias-correction individually and jointly affect accuracy in 4-bit quantization?
• RQ3 Is it feasible to quantize both weights and activations to 4 bits with acceptable loss across common CNN architectures?
• RQ4 What are the practical gains in deployment speed and memory when applying these post-training techniques?

Key findings

• ACIQ and weight bias-correction improve 4-bit baselines on average by about 3.2% and 6.0%, respectively.
• Per-channel bit allocation improves activation quantization by about 2.85% (and weight quantization by about 6.3%).
• Combining all three methods for both weights and activations recovers most of the degraded accuracy without retraining.
• Across six ImageNet models, 4-bit post-training quantization achieves accuracy close to a state-of-the-art baseline with minimal retraining—enabling rapid deployment.