MLPruner: pruning convolutional neural networks with automatic mask learning

In recent years, filter pruning has been recognized as an indispensable technique for mitigating the significant computational complexity and parameter burden associated with deep convolutional neural networks (CNNs). To date, existing methods are based on heuristically designed pruning metrics or implementing weight regulations to penalize filter parameters during the training process. Nevertheless, human-crafted pruning criteria tend not to identify the most critical filters, and the introduction of weight constraints can inadvertently interfere with weight training. To rectify these obstacles, this article introduces a novel mask learning method for autonomous filter pruning, negating requirements for weight penalties. Specifically, we attribute a learnable mask to each filter. During forward propagation, the mask is transformed to a binary value of 1 or 0, serving as indicators for the necessity of corresponding filter pruning. In contrast, throughout backward propagation, we use straight-through estimator (STE) to estimate the gradient of masks, accommodating the non-differentiable characteristic of the rounding function. We verify that these learned masks aptly reflect the significance of corresponding filters. Concurrently, throughout the mask learning process, the training of neural network parameters remains uninfluenced, therefore protecting the normal training process of weights. The efficacy of our proposed filter pruning method based on mask learning, termed MLPruner, is substantiated through its application to prevalent CNNs across numerous representative benchmarks.