Evaluating the Efficacy of Self-Replay for Pattern Recovery in Pruned Modern Hopfield Networks

Modern Hopfield Networks (MHNs) are powerful associative memory systems, but like all neural networks they are vulnerable to structural damage through weight pruning and functional interference through catastrophic forgetting. While replay mechanisms have been studied as a way to prevent forgetting during sequential learning, much less is known about whether replay can repair a network after damage has already occurred. In this paper, we investigate whether self-replay — replaying stored embeddings back through a damaged network — can restore retrieval performance in a hybrid CNN-MHN architecture trained on MNIST. We test two forms of damage: random weight pruning across ten severity levels, and functional interference through a Split-MNIST continual learning protocol. Our results show that self-replay effectively restores retrieval accuracy for structural damage up to around 50–60% pruning severity, beyond which the attractor structure is too degraded to recover. Under functional interference, self-replay recovers old memories without hurting newly learned ones, suggesting the network has enough capacity to hold both tasks when replay is used. We also find that mild pruning followed by replay can deepen the energy landscape compared to the intact baseline, offering an interesting connection to the Lottery Ticket Hypothesis.