Skip to content
Open access

Revisiting Structural Pruning via Layer-Wise Feature Map Similarity for Edge AI: A Self-Adaptive Methodology to Minimize Search Overhead

2026 · IEEE Access · Vol 14, pp. 136523-136543 · 0 citations · 54 references

Abstract

Efficient deployment of Deep Neural Networks (DNNs) on edge devices requires effective model compression with low optimization cost. However, existing approaches often force users into a costly trial-and-error process to find a suitable balance between compression and accuracy. This paper presents Self-Adaptive Structural Pruning (SASP), a self-adaptive compression methodology instantiated in this work through structured pruning. Rather than asking the user to predefine a target pruning ratio, SASP automatically searches for a highly compressed feasible model that satisfies a user-defined accuracy constraint. The proposed methodology starts from an aggressively compressed model and progressively relaxes compression where needed, using layer-wise feature map similarity, computed between corresponding layers of the original and pruned networks, to guide layer-wise adaptation. In this way, it treats compression as a search-efficiency problem, reducing the need for exhaustive search and long retraining cycles. As a result, SASP can automatically determine non-uniform layer-wise compression levels while keeping the overall optimization overhead low. We evaluate SASP on several DNN models and compare it against recent one-shot (OS) and iterative pruning methods, as well as approaches provided by industrial tools (AIMET tool by Qualcomm). The results show that, under the same accuracy constraint, SASP achieves highly competitive compression levels, with up to 92.79% parameter reduction and 83% FLOPs reduction, while providing up to a $55.66\times $ reduction in compression-search overhead compared to state-of-the-art iterative methods. Furthermore, SASP achieves higher compression with a search overhead reduction of up to $35.12\times $ compared to AIMET’s iterative gradual pruning, demonstrating its robustness for diverse real-world deployment scenarios. These results show that SASP offers a practical and scalable solution for automated model compression for Edge AI, while using structured pruning as its realization in this work.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.