IR2Net: information restriction and information recovery for accurate binary neural networks.

Weight and activation binarization can efficiently compress deep neural networks and accelerate model inference, but they cause severe accuracy degradation. Existing optimization methods for binary neural networks (BNNs) focus on fitting full-precision networks to reduce quantization errors and suffer from a tradeoff between accuracy and efficiency. In contrast, considering information loss and the mismatch between model capacity and input information quantity caused by network binarization, we propose Information Restriction and Information Recovery Network (IR 2 Net) to stimulate the potential of BNNs and achieve improved network accuracy by restricting the input information and recovering feature information. The proposed approach includes (1) information restriction, which evaluates the feature information extracted from the input by a BNN, discards some of the information it cannot focus on, and reduces the amount of the input information to match the model capacity; and (2) information recovery: due to the information loss incurred during forward propagation, the extracted feature information of the network is not sufficient for supporting accurate classification. Shallow feature maps with richer information are selected, and these feature maps are fused with the final feature maps to recover the extracted feature information and further enhance the model capacity to match the amount of input information. In addition, the computational cost is reduced by streamlining the information recovery method to strike a better tradeoff between accuracy and efficiency. Experimental results demonstrate that our approach still achieves comparable accuracy even with a ∼ 10x floating-point operations (FLOPs) reduction for ResNet-18. The models and code are available at https://github.com/pingxue-hfut/IR2Net. [ABSTRACT FROM AUTHOR]