Task-Guided Quantization Strategies

Abstract

Conventional image sensors employ uniform quantization, a process agnostic to the downstream task, often discarding task-critical information. We propose a framework for end-to-end learning of a non-uniform quantization strategy, co-designed with a neural network for visual recognition. Our method replaces the standard analog-to-digital converter with a differentiable, learned module that optimizes the allocation of discrete levels for a specific machine vision task. Extensive experiments on ImageNet, CIFAR-100, and SID demonstrate that our approach outperforms uniform quantization and other baselines at ultra-low bit-depths (4-8 bits), achieving superior accuracy and enhanced robustness to noise while significantly reducing data bandwidth.