Kion Fallah

headshot.jpg

San Francisco, CA 🌁
Background

Hi, welcome to my personal page! I am a research scientist at Waabi, led by Professor Raquel Urtasun, working on the next generation of autonomous driving technology. Specifically, I work on AI-based traffic-model simulation to improve and validate the autonomy stack. Before this position, I completed my Ph.D in machine learning at Georgia Institute of Technology under the supervision of Professor Chris Rozell.

Thesis Research

My research applied geometric models to deep learning to generate new views of data to improve performance with limited labels. To achieve this, I would use techniques like sparse coding, manifold learning, and self-supervised learning. I am interested in building large foundation models with sparse intermediary layers (i.e., only a few elements of each activation non-zero) for interpretability, robustness, and compute efficiency. One approach I have developed uses variational inference to reach similar performance as unrolled optimization routines, like FISTA, with a single neural network forward pass.

selected publications

  1. Manifold Contrastive Learning with Variational Lie Group Operators
    Kion Fallah, Alec Helbling, Kyle A. Johnsen, and Christopher John Rozell
    Transactions on Machine Learning Research, 2024
  2. Variational Sparse Coding with Learned Thresholding
    Kion Fallah, and Christopher J Rozell
    In Proceedings of the 39th International Conference on Machine Learning (Spotlight), 2022
  3. Learning Identity-Preserving Transformations on Data Manifolds
    Marissa Catherine Connor*Kion Fallah*, and Christopher John Rozell
    Transactions on Machine Learning Research, 2023
  4. Learning sparse codes from compressed representations with biologically plausible local wiring constraints
    Kion Fallah*, Adam Willats*, Ninghao Liu, and Christopher Rozell
    In Advances in Neural Information Processing Systems, 2020
* Equal Contribution