Robust Bidirectional Associative Memory via Regularization Inspired by the Subspace Rotation Algorithm

ICLR 2026 Conference Submission17777 Authors

Keywords: robust neural network, subspace rotation algorithm, bidirectional associative memory

Abstract: Bidirectional Associative Memory (BAM) trained by Bidirectional Backpropagation (B-BP) suffer from poor robustness and sensitivity to noise and adversarial attacks. To address it, we propose a novel gradient-free training algorithm, the Bidirectional Subspace Rotation Algorithm (B-SRA), designed to improve the robustness and convergence behavior of BAM. Through comprehensive experiments, two key principles, orthogonal weight matrices (OWM) and gradient-pattern alignment (GPA), are identified as central to enhancing the robustness of BAM. Motivated by these insights, new regularization strategies are introduced into B-BP, yielding models with significantly improved resistance to corruption and adversarial perturbations. We conduct an ablation study across different training strategies to determine which approach achieves a more robust BAM. Additionally, we evaluate the robustness of BAM under various attack scenarios and across increasing memory capacities, including the association of 50, 100, and 200 pattern pairs. Among all strategies, the SAME configuration—which combines OWM and GPA—achieves the highest resilience. Our findings suggest that B-SRA and carefully designed regularization strategies lead to more reliable associative memories and open new directions for building resilient neural architectures.

Primary Area: generative models

Submission Number: 17777