Just saw this. Im not too sure if anything would change? As long as the neural net is still end to end differentiable ie. uses a continuous activation function It would be fine. Im not sure what effect binary weights might have – i’m guessing less complexity but also maybe promotion of sparsity… but again, dunno.