Highlights from IEEE CDC 2024 — Part 3: A Little Bit of Everything

I've already shared what I found interesting at the IEEE Conference on Decision and Control 2024 in two previous posts. CDC 2024 was a huge conference in an interesting venue with many parallel tracks, and I saw only a small fraction of what was on offer. In this part I gather the various talks I found interesting but didn't cover before. A big shout-out to Maria Prandini, Luca Zaccarian, and Sophie Tarbouriech, who did an amazing job as the organizing committee.

Neuroscience

There weren't many neuroscience-related works at CDC 2024, which was a bit disappointing — at least not as many as I'd hoped. Among those I attended, one was notable: "Feedforward Regulation of Interneuronal Communication," which aimed to formulate a mechanism of neurotransmitter release in a chemical synapse. It's an ambitious goal, but given how hard it is to measure neurotransmitter concentration in the synaptic cleft, I don't see how such models can currently be tested experimentally — still, it's a step in the right direction. A few other talks touched on neuroscience, but from their titles and abstracts I could tell they suffer from the issues that are unfortunately common in the mathematical modeling of biological systems: they are far removed from what neuroscientists actually care about. If you have a mathematical or technical background and want to collaborate with biologists, I highly recommend "Letters to a Young Scientist" by the late E. O. Wilson — it helps you see how biologists view the world.

Another, on computational neuroscience, caught my attention: "Approximation of Koopman Operator Using Spiking Neural Networks." I file it under neuroscience because spiking neural networks are designed to work best with neuromorphic processors — though the talk was purely theoretical, with no implementation on a neuromorphic chip, which was a letdown. If you haven't heard of neuromorphic computing, the basic idea is that, unlike conventional processors, the units communicate using "spikes" (like the action potentials our neurons use). I recommend this IEEE Control Systems Magazine article by Prof. Rodolphe Sepulchre (a preprint is findable by searching the title). I firmly believe neuromorphic computing can be a game-changer for certain models, but it needs investment from chip manufacturers — and right now few seem interested.

Data-Driven Control

Two data-driven talks stood out. "Data-Driven Architecture to Encode Information in the Kinematics of Robots and Artificial Avatars," a collaboration between groups in Naples and Hamburg, builds on the fact that people encode information in their movements — from how someone grabs a bottle you can tell whether they mean to pour from it or drink from it. Their work addressed not only intentions but emotions, e.g. detecting a person's emotion from wrist movements and carrying that into virtual reality. The other, "Towards eXplainable Data-Driven Control (XDDC): The Property-Preserving Framework," aimed to quantify how much we can trust a data-driven model or controller, bringing ideas from explainable AI into data-driven control — a concept I'm increasingly convinced deserves more attention.

Mathematical Epidemiology

Epidemiology is another topic I've worked on and enjoyed. One talk piqued my interest: "Spiking Systems in Population-Infection Dynamics," by the great Franco Blanchini. "Spiking" here refers to a phenomenon I hadn't realized could occur in epidemiological systems — the textbook case of a system that always converges to an equilibrium yet cannot be called stable. The talk mostly formulated and introduced the phenomenon in the epidemiological context; Prof. Blanchini said he's working on it now, and I'm very curious what comes out of it.

Large Language Models

Unsurprisingly, LLMs featured heavily. It began with a first-day plenary, "Ask Not What AI Can Do for Control, Ask What Control Can Do for AI," by Prof. Paolo Tabuada — I got stuck in on-site registration and missed the beginning, but it was a mathematically rigorous talk on how control theory can enhance LLMs. A whole session was dedicated to the topic; I attended two talks. "Prompt a Robot to Walk with Large Language Models" (Berkeley) fed a robot's coordinates in a virtual environment to an LLM and asked it to return low-level control actions. "REAL: Resilience and Adaptation using Large Language Models on Autonomous Aerial Robots" (MIT) claimed a scheme that uses LLM outputs to increase the resilience of the control system for scenarios it wasn't designed for. At the end of the second talk, someone asked why they used LLMs rather than the many methods control theory already offers for such cases, given that with LLMs you can claim nothing about stability. For a second I hoped for an honest answer — that they'd jumped on the LLM bandwagon because it's what everyone talks about. Instead the answer sounded a lot like what an LLM would produce if you fed it the question.

Model Reduction

A tutorial session, "Forty Plus Years of Model Reduction and Still Learning," gave a nice overview of the history of model reduction and the various approaches control theorists have proposed. I find that understanding the history of a technical concept often sheds light on its meaning — this isn't limited to model reduction; even fundamental ideas like the derivative gain new depth when you explore how they developed. Sometimes a concept's journey through history reveals insights a textbook won't.

Industry

Several companies had a notable presence. The one I liked most was Mitsubishi Electric Research Laboratories. They had a stand, and I spoke with one of their researchers who had given a talk titled "Inscribing and Separating an Ellipsoid and a Constrained Zonotope: Applications in Stochastic Control and Centering" — the title alone tells you what kind of research environment they have. They also work on autonomous vehicles and seem to be hiring; from what I gathered, it's a great place for a control engineer who wants to do research in an industrial setting.

Check out more

These were simply the topics of most interest to me. For everything that was on offer, see the full conference program.

Part 3 of 3 — originally published on LinkedIn, January 16, 2025.