Binarizing by Classification: Is Soft Function Really Necessary?

Binary neural networks leverage $\mathrm{Sign}$ function to binarize weights and activations, which require gradient estimators overcome its non-differentiability will inevitably bring errors during backpropagation. Although many hand-designed soft functions have been proposed as better approximate gradients, their mechanism is not clear there are still huge performance gaps between binary models full-precision counterparts. To address these issues reduce error, we propose tackle network binarization a classification problem use multi-layer perceptron (MLP) the classifier in forward pass estimator backward pass. Benefiting from MLP's theoretical capability fit any continuous function, it can be adaptively learned backpropagate gradients without prior knowledge of functions. From this perspective, further empirically justify that even simple linear outperform previous complex Extensive experiments demonstrate method yields surprising both image human pose estimation tasks. Specifically, achieve $65.7\%$ top-1 accuracy ResNet-34 on ImageNet dataset, with an absolute improvement $2.6\%$. Moreover, take lightweighting approach for well-designed SBPN BHRNet. When evaluating challenging Microsoft COCO keypoint enables mAP up $60.6$ first time. Experiments conducted real platforms BNN achieves balance computational complexity, especially when resources extremely low.