Sep 2026· 2026 9th International Conference on Computer Network, Electronic and Automation (ICCNEA)· pp. 1-5· 0 citations· 16 references
TL;DR
This paper proposes Quantization Stability Score (QSS), a mixed-precision PTQ method that adaptively assigns bit-widths based on per-layer quantization stability, and demonstrates that QSS achieves superior accuracy-compression trade-offs.
Abstract
Deploying deep neural networks (DNNs) on resource-constrained embedded AI platforms requires effective model compression techniques. Post-training quantization (PTQ) reduces model size by converting 32-bit floating-point parameters to low-precision integers without retraining, but conventional methods apply uniform bit-widths across all layers, ignoring the inherent variation in quantization sensitivity among different network layers. This paper proposes Quantization Stability Score (QSS), a mixed-precision PTQ method that adaptively assigns bit-widths based on per-layer quantization stability. QSS jointly evaluates the Kullback-Leibler (KL) divergence and cosine similarity between full-precision and quantized layer activations to construct a comprehensive stability score. A threshold-based iterative search strategy with a rollback mechanism then allocates lower bit-widths to stable layers while preserving higher precision for sensitive ones. Extensive experiments on ResNet-18, ResNet-50, MobileNet-V2, and YOLOv8n demonstrate that QSS achieves superior accuracy-compression trade-offs. On ResNet-50 at 4-bit quantization, QSS achieves 74.02% top-1 accuracy with 8.31× compression, outperforming BRECQ (73.96%). On MobileNet-V2, QSS reaches 70.02% top-1 accuracy, surpassing BRECQ by 0.70%. Ablation studies validate the complementary roles of KL divergence and cosine similarity in capturing distinct aspects of quantization-induced perturbation.
A new approach, called Q-PACE, consisting of a second-order sensitivity model that predicts the loss increase as a sum of quantization noise MSE weighted by per-layer curvature coefficients, and its stability during training allows for infrequent, cheap recalibration.
Alexandra Volkova, Matin Ansaripour, Erik Schultheis et al.· 0 citations
A new LLM compression method that co-optimizes pruning and quantization in a unified framework that performs favorably against two-stage baselines, even when subjected to extreme quantization settings designed for ultra-efficiency.
A novel quantization and accelerator design harnessing adaptive-precision (HAP), leverages a per-group dynamic zero-point to generalize APQ and achieves superior accuracy compared with existing PTQ solutions at the level of 4 or 5 bits for a variety of DNN families.
The results show that compression and execution layout must be co-designed to convert model-size reduction into practical parallel inference acceleration, and that compression and execution layout must be co-designed to convert model-size reduction into practical parallel inference acceleration.
Qian-Sheng Song, Guo-Lin Tang· International Conference on...· 0 citations
Mixed-precision post-training quantization is a network compression method that assigns bits layer by layer, under a global memory budget using a small calibration set. The main difficulties are to overcome the combinatorial nature of the allocation problem and to manage the sensitivity to small, potentially corrupted...
Samy Houache, Y. Traonmilin, Jean-François Aujol· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.