Data-Efficient Structured Pruning via Submodular Optimization

Structured pruning is an effective approach for compressing large pre-trained neural networks without significantly affecting their performance, which involves removing redundant regular regions of weights. However, current structured pruning methods are highly empirical in nature, do not provide any theoretical guarantees, and often require fine-tuning, which makes them inapplicable in the limited-data regime. We propose a principled data-efficient structured pruning method based on submodular optimization. In particular, for a given layer, we select neurons/channels to prune and corresponding new weights for the next layer, that minimize the change in the next layer's input induced by pruning. We show that this selection problem is a weakly submodular maximization problem, thus it can be provably approximated using an efficient greedy algorithm. Our method is one of the few in the literature that uses only a limited-number of training data and no labels. Our experimental results demonstrate that our method outperforms popular baseline methods in various one-shot pruning settings.