Input-Anchored Logic Gate Networks Achieve Scalable Depth in AI Models.

Traditional Logic Gate Networks (LGNs), which perform computations through Boolean operations, have struggled to benefit reliably from increased depth. Researchers have pinpointed two primary reasons for this limitation: an optimization collapse in deep relaxed LGNs and a topology-induced constraint that persists even with stabilized training methods. This indicates that merely making a network trainable is insufficient; deeper layers must also receive information that enables meaningful computation. To address these challenges, the paper introduces Input-Anchored Logic Gate Networks (IALGNs). This novel topology ensures that each gate combines an evolving hidden feature with a direct input anchor, thereby maintaining a computational spine and conditioning every layer on the original input. This design allows for progressively more informative representations and longer computational paths. Experiments across MNIST, CIFAR-10, and CIFAR-100 datasets demonstrate that IALGNs achieve consistent improvements in depth-accuracy, even beyond 100 layers, a significant advancement over other LGN topologies that typically saturate or degrade. This research highlights that scalable depth in LGNs requires both stable optimization and an information-access pattern that supports input-conditioned refinement.