Deep Differentiable Logic Gate Networks

The problem in training networks of discrete components like logic gates, is that they are nondifferentiable and therefore, conventionally, cannot be optimized via standard methods such asgradient descent [5].