Title: Memory as Computation: Associative Memory as a Foundation for Modern AI

Date: Monday, July 6, 2026

Time: 1:00 PM to 3:00 PM Eastern Time (US)

Location: CODA 114 (first floor conference room; just walk in, no special access needed)

Remote attendance: Zoom

Benjamin Hoover (https://bhoov.com)

ML Ph.D. Student

School of Computational Science and Engineering

College of Computing

Georgia Institute of Technology

Committee:

• Dr. Duen Horng (Polo) Chau - Advisor, Georgia Tech, Computational Science & Engineering
• Dr. Judy Hoffman - UC Irvine, Department of Computer Science
• Dr. Zsolt Kira - Georgia Tech, School of Interactive Computing
• Dr. Kartik Goyal - Georgia Tech, School of Interactive Computing
• Dr. Dmitry Krotov - Dynamical Mind, Co-Founder & CEO

Abstract:

Modern AI has achieved extraordinary capabilities, but it has done so through increasingly expensive and opaque systems. At the same time, leading AI architectures strongly resemble older theories of physical memory systems — most famously, the Hopfield Network — that compute by iteratively correcting errors. This thesis shows that using these energy-based associative memories (AMs) as a design language for the next generation of AI architectures points toward a paradigm where interpretable, efficient, and robust AI emerges from physical computation built around memory, energy, and dynamics.

(1) Modernizing Associative Memory.

Classical AMs are too rigid for modern workloads. We propose the Energy Transformer (ET), a novel AM that recurrently minimizes a Lyapunov energy resembling a transformer. ET matches vanilla transformer performance while using >10x fewer parameters that are directly interpretable. Energy GPT (NRGPT) scales ET to language modeling, competitively performing against GPT baselines on key benchmarks. Memory in Plain Sight connects AMs to diffusion modeling, where denoising generation is synonymous with memory retrieval in the absence of Lyapunov stability.

(2) Generalizing Associative Memory.

Modern AMs are built using Dense Associative Memories (DenseAMs) that improve Hopfield Networks, but have limited generalizability and memory capacity tightly coupled to their size. We introduce the Distributed representation for DenseAM (DrDAM) to decouple memory capacity from parameter count, distributing memories across all synaptic connections. In Log-Sum-ReLU DenseAM (LSRDAM), we introduce optimal density estimation kernels to elicit unprecedented numbers of emergent memories, boosting creativity while maintaining exponential memory capacity.

(3) Unifying Associative Memory.

We need a systematic design language to scale DenseAMs. We propose the Hierarchical Associative Memory User eXperience (HAMUX), the first work to distill the primitives of AMs into a library of composable neuron and synaptic-layer energies. By expressing AMs as sums of local energy components, HAMUX provides a framework for building hierarchical AMs atop flexible neural network operations.

This dissertation brings the rigor of physics to AI architectures, grounding the resurgence of energy-based models in a theory of physical computation. Our research has been featured in Nature Reviews and Quanta Magazine (ET and Memory in Plain Sight), and recognized at flagship AI conferences by a NeurIPS Spotlight (LSR-DenseAM, top 3%), two full tutorials at ICML and AAAI, and four dedicated workshops across NeurIPS, ICLR, and ICCV.