ResBinNet: Residual Binary Neural Network

Recent efforts on training light-weight binary neural networks offer promising execution/memory efficiency. This paper introduces ResBinNet, which is a composition of two interlinked methodologies aiming to address the slow convergence speed and limited accuracy of binary convolutional neural networks. The first method, called residual binarization, learns a multi-level binary representation for the features within a certain neural network layer. The second method, called temperature adjustment, gradually binarizes the weights of a particular layer. The two methods jointly learn a set of soft-binarized parameters that improve the convergence rate and accuracy of binary neural networks. We corroborate the applicability and scalability of ResBinNet by implementing a prototype hardware accelerator. The accelerator is reconfigurable in terms of the numerical precision of the binarized features, offering a trade-off between runtime and inference accuracy.