WrapNet: Neural Net Inference with Ultra-Low-Resolution Arithmetic

TL;DR

WrapNet introduces a method for neural network inference that employs ultra-low-resolution (8-bit) additions in accumulators, maintaining accuracy while significantly reducing computational complexity, suitable for efficient hardware implementation.

Contribution

The paper proposes a novel approach to neural network inference that replaces high-resolution accumulation with low-resolution (8-bit) addition, using cyclic activation and regularization to preserve accuracy.

Findings

Achieves comparable accuracy to 32-bit models with 8-bit accumulation.

Demonstrates effectiveness on both software and hardware platforms.

Reduces inference complexity significantly through ultra-low-resolution arithmetic.

Abstract

Low-resolution neural networks represent both weights and activations with few bits, drastically reducing the multiplication complexity. Nonetheless, these products are accumulated using high-resolution (typically 32-bit) additions, an operation that dominates the arithmetic complexity of inference when using extreme quantization (e.g., binary weights). To further optimize inference, we propose a method that adapts neural networks to use low-resolution (8-bit) additions in the accumulators, achieving classification accuracy comparable to their 32-bit counterparts. We achieve resilience to low-resolution accumulation by inserting a cyclic activation layer, as well as an overflow penalty regularizer. We demonstrate the efficacy of our approach on both software and hardware platforms.