[Paper Review] CHIP: CHannel Independence-based Pruning for Compact Neural Networks

Yang Sui, Miao Yin+4 more|arXiv (Cornell University)|Oct 26, 2021

Advanced Neural Network ApplicationsComputer Science62 references25 citations

TL;DR

CHIP introduces channel independence as an inter-channel metric to prune filters, achieving high accuracy with substantial reductions in parameters and FLOPs on CIFAR-10 and ImageNet benchmarks.

ABSTRACT

Filter pruning has been widely used for neural network compression because of its enabled practical acceleration. To date, most of the existing filter pruning works explore the importance of filters via using intra-channel information. In this paper, starting from an inter-channel perspective, we propose to perform efficient filter pruning using Channel Independence, a metric that measures the correlations among different feature maps. The less independent feature map is interpreted as containing less useful information$/$knowledge, and hence its corresponding filter can be pruned without affecting model capacity. We systematically investigate the quantification metric, measuring scheme and sensitiveness$/$reliability of channel independence in the context of filter pruning. Our evaluation results for different models on various datasets show the superior performance of our approach. Notably, on CIFAR-10 dataset our solution can bring $0.90\%$ and $0.94\%$ accuracy increase over baseline ResNet-56 and ResNet-110 models, respectively, and meanwhile the model size and FLOPs are reduced by $42.8\%$ and $47.4\%$ (for ResNet-56) and $48.3\%$ and $52.1\%$ (for ResNet-110), respectively. On ImageNet dataset, our approach can achieve $40.8\%$ and $44.8\%$ storage and computation reductions, respectively, with $0.15\%$ accuracy increase over the baseline ResNet-50 model. The code is available at https://github.com/Eclipsess/CHIP_NeurIPS2021.

Motivation & Objective

• Motivate filter pruning from an inter-channel perspective to capture cross-channel redundancy.
• Propose Channel Independence (CI) as a metric to quantify filter importance.
• Develop a low-cost, robust scheme to compute CI and select filters for pruning.
• Demonstrate that CI-based pruning preserves accuracy while reducing model size and FLOPs across datasets.

Proposed method

• Define Channel Independence (CI) as the change in nuclear norm when removing a feature map, i.e., CI(A_i^l) = ||A^l||_* - ||M_i^l ∘ A^l||_* (Eq. 3).
• Compute CI for each feature map, average across input samples, and prune filters with the smallest CI values.
• Approximate the CI for pruning multiple filters by summing individual CI values (Eq. 4).
• Use a one-shot pruning procedure (no heavy retraining for mask learning) with subsequent fine-tuning of pruned networks.
• Algorithm 1 (CHIP) outlines per-layer steps: compute, average, sort CI, prune, and fine-tune.

Experimental results

Research questions

• RQ1 Can cross-channel feature information (inter-channel perspective) yield better filter importance than intra-channel methods?
• RQ2 Is channel independence a reliable and robust proxy for filter importance across data distributions and input batches?
• RQ3 How to efficiently compute CI for multiple-filter pruning without combinatorial explosion?
• RQ4 Does CHIP achieve substantial compression with minimal accuracy loss on standard benchmarks?
• RQ5 Is further data-driven refinement of pruning masks beneficial beyond the proposed one-shot CI-based scheme?

Key findings

• On CIFAR-10, CHIP yields up to 0.90%–0.94% accuracy gains with substantial reductions in model size and FLOPs across ResNet-56, ResNet-110, and VGG-16.
• On ImageNet, CHIP achieves 40.8% storage and 44.8% FLOPs reductions with a 0.15% accuracy increase for ResNet-50.
• Compared to state-of-the-art pruning methods, CHIP provides competitive or superior accuracy at similar or better compression ratios (e.g., CIFAR-10 results and ImageNet comparison).
• The CI-based, one-shot pruning scheme remains robust across batches and input distributions, with minimal need for additional mask learning during fine-tuning.