Robust Auto-associative Memory via Convolutional Restricted Hopfield Networks

The paper introduces Convolutional Restricted Hopfield Networks (CRHNs), which enhance associative memory performance by integrating convolutional feature extraction with attractor-based memory retrieval. Utilizing a gradient-free Subspace Rotation Algorithm (SRA), CRHNs demonstrate significantly lower reconstruction errors compared to Modern Hopfield Networks (MHNs) and Predictive Coding Networks (PCNs) under adversarial conditions, achieving improvements by an order of magnitude in many cases. This model's robustness and memory capacity make it a valuable framework for practitioners aiming to develop scalable and resilient associative memory systems in AI applications.