RAP: A Reliability-aware Pruning Framework for Deep Neural Networks

Setareh Ahsaei

doi:10.1016/j.sysarc.2026.103829

RAP: A Reliability-aware Pruning Framework for Deep Neural Networks

2026-05-02

0

PRE

AI

DOI：10.1016/j.sysarc.2026.103829

Original

Original request for help

Save

Abstract

En 中文

Deep neural networks (DNNs) are widely used for machine intelligence and automation, but their large number of parameters necessitates significant memory and computation resources. Pruning is an effective technique for reducing model size and enabling deployment on resource-constrained embedded platforms. However, existing pruning techniques have overlooked the possible reliability challenges originated from soft errors during pruning DNNs, making serious challenges towards their application in safety-critical embedded systems. This paper presents a reliability-aware pruning framework, called RAP, which incorporates the reliability of DNNs throughout the entire pruning process. RAP introduces a novel pruning criterion that evaluates the impact of soft errors on DNN reliability in addition to classification accuracy. Moreover, a reliability-aware post-pruning fine-tuning is proposed to fine-tune the remaining weights of the model while preserving their resilience against soft errors. The proposed framework is evaluated on VGG11, ResNet50, ViT-Base, and DeiT-Base models trained on the CIFAR-100 dataset. Experimental results demonstrate that RAP consistently enhances the reliability of deep neural networks across a wide range of pruning ratios (PRs) and Bit Error Rates (BERs), while maintaining competitive accuracy. These findings highlight the effectiveness of incorporating reliability awareness into the pruning and fine-tuning process, enabling more robust DNNs under fault-prone conditions.

Keywords:

Reliability-aware pruning

Deep neural networks

Soft errors

Model compression

Fault resilience

RAP: A Reliability-aware Pruning Framework for Deep Neural Networks

Abstract

Journal

Organization

Cited Papers