Teaching

Major new special course connecting theory to AI Systems and modern tooling like NeMo, vLLM, etc. on multi-node multi-GPU machines. We secured a very generous compute donation for this course so that students can explore at scale.

Major new special course connecting theory to AI Systems and modern tooling like NeMo, vLLM, etc. on multi-node multi-GPU machines. We secured a very generous compute donation for this course so that students can explore at scale.

Major upgrades: covered Muon, muP, and a unified optimization perspective. Deeper state-space modeling and hybrid attention. Fuller RLVR treatment leading to improved RLHF treatment including DPO.

Significant upgrades: added modern state-space models and a unified treatment of DDIM and DDPM diffusion models.

A course centered on research projects in in-context learning.

Continuing upgrades to course: brought in some RLHF

Major upgrade to course: brought in GNNs and Diffusion Models

Fully stabilized the materials in a unified way

Upgrades to materials --- introduced many more computational visualizations leveraging the one-dimensional functional case, a full treatment of overparameterization, and MAML

Covid strikes! Framing update: changed the material on stable matchings to no longer use gender-based stable marriage and instead think about matching people to jobs. Easy payoff in terms of inclusion with no substantive loss in clarity. (Even if it made obsolete this silly video by former students in the class)

Consolidation of course material

Critical materials building/updating semester

Machine learning for sequential decision making under uncertainty, including regret, bandits, and reinforcement learning. Co-taught with Vidya Muthukumar, who was then my student and is now a professor at Georgia Tech.

I was just helping out Elad Alon. I gave a bunch of the lectures, but nothing else really.

Collaborative Intelligent Agents. Inspired by the DARPA Spectrum Collaboration Challenge.

This was the second semester of the course, and the first time it had been offered at scale to about 550 students. A lot of material had to be developed.

This was the second semester of the course, and the first time it had been offered at scale to about 550 students.

A fun semester. We actually managed to give voluntary oral exams to more than 500 students.

An offering in which we upgraded the material --- adding a more hands-on laboratory component using student laptops to do audio-based digital communication, as well as having much more material on the diverse uses of error-correcting codes beyond traditional point-to-point digital communication.

A special semester where we explored upgrades to the course while teaching two sections: a regular one and a special honors one with the expanded material. The expanded section moved to two discussion sections per week and also tried out having two two-hour lectures instead of two one-and-a-half-hour lectures.

Network Information Flow: a look at multiuser information theory through simplified deterministic models designed to illustrate the key effects that occur in the multiuser setting. We move from wireline to deterministic wireless to wireless as we augment our models to be closer to reality.

Focus on feedback and different non-probabilistic approaches to uncertainty: universal coding, compound channels, AVCs, and individual sequences

Focus on multiuser theory, feedback, and the intersection of control and information theory.