binary neural networks (BNNs) represent original full-precision weights and
activations into 1-bit with sign function. Since the gradient of the
conventional sign function is almost zero everywhere which cannot be used for
back-propagation, several attempts have been proposed to allevi