A Generic Layer Pruning Method for Signal Modulation Recognition Deep Learning Models
Yao Lu, Yutao Zhu, Yuqi Li, Dongwei Xu, Yun Lin, Qi Xuan, Xiaoniu Yang
TL;DR
The paper tackles the high computational demand of deep AMR models by proposing a generic layer pruning framework that partitions networks into blocks of semantically similar layers using representational similarity measures, selects important layers per block with a training-free SynFlow proxy, and reassembles a compact, fine-tuned model. The approach, grounded in Fisher's optimal segmentation and CS-based similarity, yields efficient pruning across five AMR benchmarks, beating state-of-the-art channel-pruning and several layer-pruning baselines while preserving accuracy. The methodology supports edge deployment by achieving substantial reductions in FLOPs and parameters with minimal performance loss, and includes extensive ablations confirming robustness to the number of blocks and choice of similarity metric. Overall, this work provides a general, scalable pipeline for layer-wise model slimming in DL-based signal processing tasks.
Abstract
With the successful application of deep learning in communications systems, deep neural networks are becoming the preferred method for signal classification. Although these models yield impressive results, they often come with high computational complexity and large model sizes, which hinders their practical deployment in communication systems. To address this challenge, we propose a novel layer pruning method. Specifically, we decompose the model into several consecutive blocks, each containing consecutive layers with similar semantics. Then, we identify layers that need to be preserved within each block based on their contribution. Finally, we reassemble the pruned blocks and fine-tune the compact model. Extensive experiments on five datasets demonstrate the efficiency and effectiveness of our method over a variety of state-of-the-art baselines, including layer pruning and channel pruning methods.