Structured Binary Neural Networks for Image Recognition

In this paper, we propose to train binarized convolutional neural networks (CNNs) that are of significant importance for deploying deep learning mobile devices with limited power capacity and computing resources. Previous works on quantizing CNNs often seek approximate the floating-point information weights and/or activations using a set discrete values. Such methods, termed value approximation here, typically built same network architecture full-precision counterpart. Instead, take new “structured approximation” view quantization — it is possible valuable exploit flexible transformation when low-bit networks, which can achieve even better performance than original in some cases. particular, “group decomposition” strategy, GroupNet, divides into desired groups. Interestingly, our GroupNet each group be effectively reconstructed by aggregating homogeneous binary branches. We also learn effective connections among groups improve representation capability. To model capacity, dynamically execute sparse branches conditioned input features while preserving computational cost. More importantly, proposed shows strong flexibility few vision tasks. For instance, extend accurate semantic segmentation embedding rich context structure. The object detection. Experiments image classification, segmentation, detection tasks demonstrate superior methods over various quantized literature. Moreover, speedup runtime memory cost evaluation comparing related strategies analyzed GPU platforms, serves as benchmark further research.