End-to-End Differentiable Channel Fusion for Structured CNN Pruning

Channel pruning is a widely adopted strategy for reducing the computational cost of convolutional neural networks (CNNs), improving computational efficiency under limited resource budgets. Existing methods often rely on hand-crafted criteria and non-differentiable clustering, hindering joint endto-end optimization of channel grouping and pruning under high compression ratios. In this work, we introduce a fully differentiable channel fusion framework for structured CNN pruning. Interchannel dependencies across layers are captured through an inter-layer feature matrix and optimized via a Gumbel-Softmax-based differentiable clustering module. This unified end-to-end design enables adaptive fusion of redundant channels according to learned clustering assignments, preserving informative features while substantially reducing computational complexity. Comprehensive experiments on representative CNN architectures, including ResNet, VGG and MobileNet, show that the proposed method can reduce floating-point operations by up to 54% while maintaining Top-1 accuracy loss within 1.4%. Additional runtime evaluations on an edge platform further confirm that the theoretical computational reductions translate into measurable inference improvements. These results demonstrate that the proposed approach provides an effective and general framework for structured CNN pruning, achieving a favorable balance between model compactness and predictive performance.