关键词: 二值神经网络, 剪枝, 可训练门限, 演化

Abstract: BNNs( Binarized Neural Networks) are popular due to their extremely low memory requirements. While BNNs can be further compressed through pruning techniques, existing BNN pruning methods suffer from low pruning ratios, significant accuracy degradation, and reliance depending on fine-tuning after training. To overcome these limitations, a filter-level BNN pruning method is proposed based on evolution from ternary to binary, named ETB ( Evolution from Terry to Binary). ETB is learning-based, and by introducing trainable quantization thresholds into the quantization function of BNNs, it makes the weights and activation values gradually evolve from ternary to binary or zero, aiming to enable the network to automatically identify unimportant structures during training. And a pruning ratio adjustment algorithm is also designed to regulate the pruning rate of the network. After training, all zero filters and corresponding output channels can be directly pruned to obtain a simplified BNN without fine-tuning. To demonstrate the feasibility of the proposed method and the potential for improving BNN inference efficiency without sacrificing accuracy, experiments are conducted on CIFAR-10. ETB is pruned the VGG-Small model by 46. 3% , compressing the model size to 0. 34 MB, with an accuracy of 89. 97% . The ResNet-18 model is also pruned by 30. 01% , compressing the model size to 1. 33 MB, with an accuracy of 90. 79% . Compared with some existing BNN pruning methods in terms of accuracy and parameter quantity, ETB has certain advantages.

Key words: binarized neural network, pruning, trainable threshold, evolution