Jay Jayaraman Thiagarajan


Portrait of  Jay Jayaraman Thiagarajan
  • Title
    Principal Machine Learning Scientist
  • Email
  • Phone
    (925) 424-2255
  • Organization

Jayaraman J. Thiagarajan (Jay) is a Principal Machine Learning Scientist in the Center for Applied Scientific Computing at Lawrence Livermore National Laboratory. He has over a decade of research experience, specializing in deep learning, AI/ML safety, generative AI (LLMs, VLMs, Diffusion, GANs), and human-centric evaluation. His research ideas have been instrumental in driving several high-impact projects across diverse domains, including sciences (notably, contribution to the breakthrough in nuclear fusion at the National Ignition Facility and drug discovery), healthcare (e.g., the Medperf platform, drug discovery), and security (robust computer vision and airport security initiatives).

Ph.D. Electrical Engineering, Arizona State University, Tempe, Arizona

For a complete list of Jay's publications, visit google scholar page

Selected Highlights

Foundation Models: Developed interfaces for LLMs (LLama, GPT, RoBERTa), VLMs (CLIP), and vision models (DINO, SimCLR) to perform model adaptation [LP+FT protocols @ ICLR’23], synthetic data generation [SiSTA @ ICML’23], OOD generalization [SPHInX @ ICML’22] and prompt generation [CREPE]

AI/ML Safety: Introduced anchored neural networks that has led to SoTA in robustness, accuracy prediction, and anomaly rejection [∆−UQ @ NeurIPS’22] [Graph ∆−UQ]. Made key advances in domain generalization [ALT @ WACV’23], test-time adaptation [CATTAn @ ACML’22], calibration [LbC @ Nature Comm.’20] and uncertainty quantification [AIP @ AAAI’20]

Inverse Problems: Designed diffusion-based methods for CT (computed tomography) imaging — 15% more accurate and 30% faster than SoTA [DOLCE @ ICCV’23]. Innovations in the use of generative priors for image restoration [MimicGAN @ IJCV 2020], audio source separation [Interspeech’20] and experimental data calibration [MaCC @ PNAS’20].

Human-Centric Evaluation: Designed novel tools for systematic auditing and exploration of ML models. Examples include bias detection in generative models [xGA @ CVPR’23], counterfactual generation [DISC @ NeurIPS’21], failure mode detection [PAGER] and understanding covariate shifts [ProFILE @ AAAI’21].

Curriculum Vitae

Here is Jay's CV

Personal Website

Jay maintains a personal website at https://jjthiagarajan.com/

This website is not affiliated with or sponsored by LLNL