AuGQ: Augmented quantization granularity to overcome accuracy degradation for sub-byte quantized deep neural networks

초록

Deployment of neural networks on IoT devices unleashes the potential for various innovative applications, but the sheer size and computation of many deep learning (DL) networks prevented its widespread. Quantization mitigates this issue by reducing model precision, enabling deployment on resource-constrained edge devices. However, at extremely low bit-widths, such as 2-bit and 4-bit, the aggressive compression leads to significant accuracy degradation due to the reduced representational capacity of the neural network. A critical aspect of effective quantization is identifying the range of real values (FP32) that impact model accuracy. To address accuracy loss at sub-byte levels, we introduce Augmented Quantization (AuGQ), a novel granularity technique tailored for low bit-width quantization. AuGQ segments the range of real-valued (FP32) weight and activation distributions into small uniform intervals, applying affine quantization in each interval to enhance accuracy. We evaluated AuGQ using both post-training quantization (PTQ) and quantization-aware training (QAT) methods, achieving accuracy levels comparable to full precision (32-bit) DL networks. Our findings demonstrate that AuGQ is agnostic to the training pipeline and batch normalization folding, distinguishing it from conventional quantization techniques. Furthermore, when integrated into state-of-the-art PTQ algorithms, AuGQ necessitates only 64 training samples for fine-tuning which is 16× fewer than traditional methods. This reduction facilitates the application of high-accuracy quantization at sub-byte bit-widths, making it suitable for practical IoT deployments and enhancing computational efficiency on edge devices. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2025.

키워드