Fully Trainable Deep Differentiable Logic Gate Networks and Lookup Table Networks

The paper introduces a method for jointly optimizing connectivity and computation in deep differentiable logic gate networks (LGNs) and lookup table networks (LUTNs). Each gate or LUT input pin learns a probability distribution over candidate connections, while gate types or LUT entries are optimized in parallel. The authors report 98.92% accuracy on MNIST with two layers of 8,000 gates and 98.45% with one layer, claiming nearly 50 times fewer gates than fixed-connection LGNs. For LUTNs, two layers of 2,000 six-input LUTs reach 98.88% accuracy.