Rushil Anirudh

Hi! I am an applied scientist at Amazon AGI, working on customizing frontier models for downstream applications. Before this, I spent ~14 mo within the Ads / Sponsored Products org working on improving ad recommendations to drive higher shopper engagement and improve Amazon's retail business. Prior to Amazon, I spent 7.5 years at Lawrence Livermore National Laboratory as a Project Lead on foundational and applied ML problems.

Career Timeline

• Applied Science @ Amazon AGI (5/2025- present)
• Applied Science @ Amazon Ads (1/2024- 5/2025)
• Principal Scientist, Lawrence Livermore National Laboratory (10/2016-12/2023)
• Postdoc, IBM Research Almaden(4/2016-10/2016)
• PhD Student, Arizona State University (8/2010-3/2016)

(9/25/2024) Recent papers accepted for publication:

• [NeurIPS 2024] spotlight! On the Use of Anchoring for Training Vision Models (arXiv)
  Explores short comings of anchoring for ViTs and fixes them to show significant improvements in ImageNet-1K OOD detection, calibration, generalization benchmarks.
• [ICML 2024] PAGER: Accurate Failure Characterization in Deep Regression Models (openreview)
  Failures in classifiers is clear (mis-classification rate) and well studied (OOD detection etc.), but very tricky in regression problems -- PAGER takes a step in formalizing failure analysis using a combination of epistemic and non-conformity scores
• [ICLR 2024] Accurate and Scalable Estimation of Epistemic Uncertainty for Graph Neural Networks (openreview)(webpage)
  Introduces G-ΔUQ (pronounced "G-DUCK"), a novel training protocol for graph neural networks that supports reliable epistemic uncertainty estimates for graph classification and node classification tasks -- lot of SoTA results!
• [IOP Machine Learning Sci & Tech] Transformer-Powered Surrogates Close the ICF Simulation-Experiment Gap with Extremely Limited Data (article)
  New training and model selection protocol for Transformer surrogates, which improve existing methods on few-shot adaption to real world ICF experiments by ~40% improvement in predictive performance.