We introduce a novel scheme to train binary convolutional neural networks (CNNs) -- CNNs with weights and activations constrained to {-1,+1} at run-time. It has been known that using binary weights and activations drastically reduce memory size and accesses, and can replace arithmetic operations with more efficient bitwise operations, leading to much faster test-time inference and lower power consumption.

The paper describes a new method for training binary convolutional neural networks. The key insight here is to approximate the real valued weights via a linear combination of M binary basis weights. The coefficients for reconstructing the real weights can be found using least squares in the forward pass, and then pulled outside the convolution to allow for fast binary convolution at test time. A similar approach is taken for the activations, but in this case the weights and shifts are trained as normal during backpropagation. The result is a network that requires M more binary convolutions than a straightforward binary neural network, but it is expected that these will be significantly more hardware friendly.

We introduce a novel scheme to train binary convolutional neural networks (CNNs) -- CNNs with weights and activations constrained to \{-1, 1\} at run-time. It has been known that using binary weights and activations drastically reduce memory size and accesses, and can replace arithmetic operations with more efficient bitwise operations, leading to much faster test-time inference and lower power consumption. In this paper, we address this issue with two major innovations: (1) approximating full-precision weights with the linear combination of multiple binary weight bases; (2) employing multiple binary activations to alleviate information loss. The implementation of the resulting binary CNN, denoted as ABC-Net, is shown to achieve much closer performance to its full-precision counterpart, and even reach the comparable prediction accuracy on ImageNet and forest trail datasets, given adequate binary weight bases and activations. Papers published at the Neural Information Processing Systems Conference.