A quick peak into some of my past projects.
AI/ML Safety
Improve generalization of ML models “in the wild”, e.g., covariate shifts, task shifts, sub-population shifts, unknown corruptions and semantic novelty.
Related Work:
SiSTA for generative augmentations (ICML 2023)
A closer look at model adaptation (ICLR 2023)
AGAT for improving robustness (AAAI 2021)
Uncertainty Quantification
Develop UQ tools to enable safe ML practice, e.g., reject out-of-distribution (OOD) data, delegate high-risk predictions to experts, defend against adversarial attacks, incorporating real-world priors.
Related Work:
PAGER for failure characterization (2023)
Delta-UQ (NeurIPS 2022)
Learn-by-calibrating (Nature Comm. 2020)
Knowledge-Driven AI
Leverage external sources such as knowledge graphs or pre-trained foundational models to build data-efficient and grounded AI solutions
Related Work:
CREPE for visual relation prediction (2023)
CAML for few-shot generalization (WACV 2023)
CATTAn for test-time adaptation (ACML 2022)
Scientific ML
Advances in representation learning, surrogate modeling, multi-fidelity optimization and transfer learning have led to key milestones (high-energy physics, biological sciences, epidemiology)
Related Work:
Modeling Human Cell Membrane (MLST 2022)
Multimodal representation learning (PNAS 2020)
Data-Driven Plasma Science Review (IEEE TPS)
Generative AI
Develop, adapt and leverage generative models for applications in CT reconstruction, inverse imaging, history matching and audio restoration
Related Work:
DOLCE for limited angle CT (ICCV 2023)
SPHINX for GAN inversion (ICML 2022)
MimicGAN for using GANs as priors (IJCV 2020)
Graph ML
Develop and advance graph-based ML methods, including graph neural networks, to solve node-level, edge-level and graph-level tasks.
Related Work:
Self-supervised learning in GNNs (NeurIPS 2022)
Uncertainty Matching GNNs (AAAI 2021)
GraMME for multi-layer graphs (IEEE TNNLS)
Healthcare AI
Through collaborations with clinical experts, I have worked towards solving key challenges in clinical diagnosis and patient monitoring, e.g., weak supervision, biases, multimodality, safety.
Related Work:
Medperf effort with ML Commons (Nature MI)
DDx-Net for EHR modeling (Sci Rep.)
SAnD for time-series modeling (AAAI 2018)
Explainability
I have developed a number of frameworks to – (i) introspect how a model works; (ii) support counterfactual reasoning; (iii) explain distribution shifts; and (iv) explain under shifts.
Related Work:
xGA for cross-GAN auditing (CVPR 2023)
Counterfactual reasoning (Neurips 2021)
ProFILE for feature importances (AAAI 2021)
High-Dimensional Data Analysis
Through a DOE-funded project, our team developed a new statistical mechanics framework to study ML generalization and advanced design of space-filling sampling strategies.
Related Work:
A spectral approach to DoE (JMLR 2018)
Generalization via sample design (NeurIPS 2020)
Stair Blue Noise Sampling (Siggraph Asia 2016)
jjthiagarajan.com 