Neurosymbolic logic gate networks
Published:
Differentiable logic gate networks learn by relaxing each neuron into a soft two-input Boolean gate, which later hardens into a real circuit. I run logic synthesis on the hardened circuit between gradient steps, so the model keeps re-compressing what it has learned.
Where gradient descent stalls, as it does on parity, an exact GF(2) solver takes over and recovers the function from about n+1 examples. Alternating the two beats either one alone. The finished model is a small circuit that decompiles to a rule set with provable output guarantees.
Stack: PyTorch, Python. In progress.