Input-Anchored Logic Gate Networks Achieve 100+ Layer Depth

A team of researchers led by Taegun An at Korea University has identified why existing Logic Gate Networks (LGNs) fail to benefit from increased depth — and proposed a fix. In their paper submitted to arXiv, they show that deep LGNs suffer from two distinct problems: optimization collapse in relaxed versions, and a topology-induced limitation that persists even with careful initialization. Their solution, Input-Anchored Logic Gate Networks (IALGNs), modifies the network so that each gate combines a hidden feature with a direct connection to the original input. This creates a "computational spine" that preserves information flow while conditioning every layer on raw data.

The results are striking: IALGNs achieve consistent accuracy gains as depth increases beyond 100 layers on standard vision benchmarks (MNIST, CIFAR-10, CIFAR-100), while previous LGN topologies saturate or degrade. The team proves that a depth-D path can depend on up to D+1 input bits, establishing a strict hierarchical advantage. A Random-k anchor relaxation further improves selection without breaking the spine. Layer-wise probes confirm that input anchoring produces progressively more informative representations and longer effective computation paths. The work suggests that scalable depth in Boolean networks requires both stable training and an information-access pattern that supports iterative refinement — a principle that could influence future hardware-aware AI designs.