Operand Selective Logic Gate Network

Soo Yong Kim

Keywords: Operand

Abstract: We propose Operand-Selective Logic Gate Networks (OSLGN), a symbolic neural architecture that builds differentiable logic circuits via operand and operator selection. Each logic unit dynamically selects two operands from the input and applies one of sixteen predefined binary logic operators, thereby forming a symbolic computation structure that remains trainable through gradient descent. Our operator selection builds upon prior work on differentiable logic gates, while our introduction of operand selection constitutes a novel modular extension. To encourage locally coherent logic formation, we initialize operand selectors with a proximity-based prior inspired by small-world network topology. Specifically, each operand selector is biased toward selecting neighboring input features, allowing the network to efficiently compose local structures and gradually learn long-range dependencies. Experiments on MNIST demonstrate that this initialization improves generalization and stabilizes gradient flow, and we further show that despite modest classification performance, the trained network can be fully converted into compact symbolic logic expressions.

Submission Number: 43