B-SRA: A Gradient-Free Training Algorithm to Enhance BAM Robustness

A recent paper published on arXiv (ID: 2511.11902) presents the Bidirectional Subspace Rotation Algorithm (B-SRA), a novel gradient-free approach aimed at enhancing the resilience of Bidirectional Associative Memory (BAM) models. Traditional BAMs, which utilize Bidirectional Backpropagation (B-BP), often exhibit vulnerability to noise and adversarial threats. To mitigate these issues, the authors introduce B-SRA, showcasing marked improvements in both robustness and convergence rates. Their research highlights two fundamental concepts—orthogonal weight matrices (OWM) and gradient-pattern alignment (GPA)—as crucial for strengthening BAM's defenses. They also propose new regularization techniques for B-BP, leading to models that show significantly better resistance to adversarial attacks and corruption. An ablation study is included to assess various training strategies for optimal robustness. The paper is classified as a replace-cross type and can be accessed via the provided source URL.