https://hgpu.org/?p=15446
BinaryNet: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1