Rushil Anirudh

D93CCDC5-529F-4134-A5E8-909F0A40649D


Hi! I am a Computer Scientist with the Machine Intelligence Group at Lawrence Livermore National Laboratory (LLNL), a Federally Funded Research Institute. My research interests broadly include computer vision, machine learning, and high dimensional data analysis.

My CV (pdf, updated July 2022).

RESEARCH THEMES:
(a) OOD Generalization, Uncertainty Quantification: characterizing and improving ML model behavior under distribution shifts.
(b) Knowledge-driven: leveraging external knowledge to improve zero shot, few shot learning.
(c) Generative Modeling & Imaging: solving ill-posed inverse imaging problems with novel generative priors.
(d) Sim2Real with ML: Understanding the challenges of transferring, generalizing ML models trained on simulation data to the real world

OUTREACH & OTHER EFFORTS:
At LLNL, I lead the Open Data Initiative with the goal of releasing LLNL’s high impact scientific datasets for machine learning — so far, we have 10+ datasets and counting!
AI.gov recently cited this effort under “DATA RESOURCES FOR AI R&D”!

I also serve on the program committees for several ML and vision venues (NeurIPS, CVPR, ICLR, ICCV, ICML, AAAI).

Updates

  • (8/16) Improving Diversity with Adversarially Learned Transformations for Domain Generalization Accepted for publication to WACV 2023! 🏖️ [paper] [code].
    Improves domain generalization performance by expanding the space of learnable augmentations.
  • (7/18) arxiv preprint: Single Model Uncertainty Estimation via Stochastic Data Centering
    Highly accurate epistemic uncertainty estimates via a novel, trivial data centering operation
  • (7/11) arxiv preprint: Revisiting Inlier and Outlier Specification for Improved Out-of-Distribution Detection
    Studies the relative importance of inliers vs outlier exposure for effective OOD detection
  • (7/8) arxiv preprint: OOD Detection using Neural Network Anchoring
    Hetero-scedastic (per sample) temperature scaling strategy that improves OOD detection