Arohi Nair

Molecular structure prediction

Molecular structure prediction brings together computational chemistry and machine learning. This research benchmarks generative and co-folding models, evaluates predicted structures, and investigates BoltzGen’s input-order invariance.

Rating models for heads-up poker

Can margin of victory improve predictions of future heads-up poker outcomes? This thesis compares Elo-style models that incorporate winning margins with standard win/loss-based ratings.

Heads-up no-limit poker · Rating models · Predictive evaluation

Making medical image segmentation more adaptable

This thesis explores sequential test-time adaptation for medical image segmentation, asking whether predictions from neighboring image slices can help models adapt to unfamiliar data.

The approach combines image registration to align predictions between slices with energy-based adaptation to encourage anatomically plausible segmentations. The work examines when adaptation helps, when it doesn’t, and what explains those differences.