CRHNs cut reconstruction errors 10x under adversarial attacks

Associative memory models are critical for pattern retrieval in AI, but they degrade under adversarial perturbations and severe input corruptions. Existing approaches like Modern Hopfield Networks (MHNs) and Predictive Coding Networks (PCNs) struggle with storage capacity, computational efficiency, and robustness. To address this, researchers Ci Lin, Tet Yeap, and Iluju Kiringa introduce Convolutional Restricted Hopfield Networks (CRHNs). CRHNs combine convolutional feature extraction with attractor-based memory retrieval in a structured latent space, using subspace representations and fixed-point dynamics. They employ a novel gradient-free Subspace Rotation Algorithm (SRA) for training, avoiding backpropagation bottlenecks.

Extensive experiments on the Self-Taught Learning (STL) dataset show CRHNs consistently achieve significantly lower reconstruction error compared to MHNs and PCNs across a wide range of adversarial attacks and input degradations. In many cases, CRHNs reduce reconstruction error by an order of magnitude and maintain stable retrieval performance as perturbation strength increases. Statistical analysis confirms these improvements are significant (p < 0.01). These results highlight the effectiveness of attractor-based memory mechanisms and suggest CRHNs provide a promising framework for building robust and scalable associative memory systems, especially for applications like image recognition and autonomous systems where input noise is common.