The School of Computer Science is pleased to present…
Associative Memory Networks for Representation Learning: From Robustness to Emergent Structure Discovery
MSc Thesis Defense by: Saleh Sargolzaei
Date: Wednesday, June 25, 2025
Time: 10:00 AM
Location: Essex Hall, Room 122
This thesis develops a unified framework for representation learning through associative memory networks, demonstrating how Hopfield networks address fundamental challenges in unsupervised learning through interconnected approaches. We first establish that Modern Hopfield Networks (MHN) enhance model robustness to corrupted data by learning energy-based prototypes that serve as robust attractors for clean representation recovery. Building on this foundation, we show how Hopfield network structure can be systematically exploited for latent discovery by connecting spurious memories to meaningful data directions through spiked covariance models, enabling hierarchical extraction of interpretable prototypes. These complementary capabilities, robust prototype learning and structure discovery, naturally lead to DANCE (Dynamic Associative Network for Cluster Emergence), which unifies both approaches through competing Hopfield networks. DANCE exhibits emergent two-phase dynamics: exploration that identifies latent signals, followed by refinement that improves prototypes through energy-based learning. This transition occurs naturally via adaptive repulsion, creating a self-organizing system for adaptive clustering. The progression from robustness to structure discovery to adaptive clustering demonstrates how associative memory networks provide a principled foundation for representation learning, offering interpretable and adaptive solutions that unify key challenges in unsupervised learning.
Internal Reader: Dr. Dan Wu
External Reader: Dr. Ahmed Hamdi Sakr
Advisor: Dr. Luis Rueda
Chair: Dr. Muhammad Asaduzzaman