andrew smith

An energy landscape. A state settles into a low-energy configuration, then re-forms as the landscape shifts.

ECE PhD Student · UC Santa Cruz

I am a second year ECE PhD student at UCSC advised by Jason Eshraghian. Previously, I worked on local learning rules at the Allen Institute, and industrial edge computing gateways at Microsoft. I did my undergrad at the University of Virginia in Computer Science. I am now working on local learning for Energy-Based Models and Complementary Learning Systems.

Research

Energy-Based Models and Complementary Learning Systems

A lot of my research is about unlocking better forms of locally learned, parsimonious Energy-Based Models (EBMs) — and figuring out how these methods can scale, not on their own, but in conjunction with feature encoders trained with backprop and other state-of-the-art methods.

The research I care about focuses on dynamics, which I think are key to a lot of the hard problems in interacting with the world. I believe EBMs offer a unique advantage here, especially when forming a Complementary Learning Systems (CLS) architecture that pairs fast adaptation with slow consolidation. I believe that unblocking CLS will offer methods where we can then scale compute against arbitrary data.

I am especially interested in the intersection of these three areas:

Machine learning

Local, biologically plausible training procedures for energy-based and recurrent models.

Neuroscience

Learning rules as a lens on cortical credit assignment and memory consolidation.

Hardware

Temporally and spatially local algorithms that cut the dominant on-chip cost: communication.

Working from the inside out

Current foundation models are largely compression engines for the corpus, which we can then task-tune for various domains. Task-tuning a compression engine is a lot different than understanding the fundamental mechanisms of intelligence.

I think about intelligence as a set of nested systems — an inner adaptive loop wrapped in memory, salience, world-modeling, and abstraction. Foundation models sit outside that stack, learning from the outputs it emits; my interest is in the mechanisms inside the onion. I believe there should be a stronger bias to incrementally work from the inside of the onion, outwards, rather than to work outside the onion entirely.

outputs foundation models
  1. Adaptive loop
    Sense, predict, act, update.
  2. Memory and learning
    Consolidate experience into reusable patterns.
  3. Salience and motivation
    Tag what matters: reward, threat, need, meaning.
  4. World modeling
    Simulate causes, agents, objects, futures.
  5. Abstraction
    Compress patterns into concepts, symbols, language, rules.
Foundation models sit outside the stack, modeling only the outputs the surface emits.
Contact

Get in touch

I'm always happy to talk about local learning, energy-based models, neuromorphic hardware, and continual learning. Reach me at andrew.smith.communication@gmail.com.