Rob Brekelmans


rob.brekelmans@vectorinstitute.ai

Links:

I am a Postdoctoral Fellow at the Vector Institute in Toronto, working with Alireza Makhzani and Roger Grosse. I graduated with my PhD from University of Southern California in 2022, working with Greg Ver Steeg and Aram Galstyan. I also interned with the AI Safety Analysis team at DeepMind (Blog), working with Pedro Ortega and Tim Genewein.

I'm currently on the job market for industry roles near mountains and/or ocean, feel free to reach out! (Email) (Google Scholar)

---------------------------
We won a Best Paper Award at ICML 2024 for our work "Probabilistic Inference in Language Models via Twisted Sequential Monte Carlo"!
  • Paper
  • Talk (from 1:00:30)

  • ---------------------------
    Action Matching (AM) is an implicit objective with similar applications as Flow Matching but fewer assumptions (only requires samples). Wasserstein Lagrangian Flows extends AM to solve (multi-marginal) optimal transport problems with flexible dynamical cost functions, for example in trajectory inference for single-cell RNA data.
  • WLF Paper
  • AM Paper

  • ---------------------------
    I enjoyed working on AM/WLF to study the geometry on the space of probability distributions induced by optimal transport distances. In my PhD, I studied information geometry (induced by divergence functionals), culminating in the paper:
    Variational Representations of Annealing Paths (Information Geometry journal 2024):
  • (Paper), (Tweet Thread), (Information Geometry for Data Science talk), (Slides)

  • ---------------------------
    Mutual Information Estimation ICLR 2022:
  • (Paper), (5 min Talk), (Tweet Thread)

  • ---------------------------
    PhD Thesis