ML-Track
-
Interpretability-Guided Unlearning in Audio
[Undergraduate Thesis]: Interpretability-guided machine unlearning (MU)
for audio models, using SAE-driven feature mining and LEACE
concept edits, evaluated on speaker and speech recognition
datasets across English, Hindi, and Spanish. Beats SoTA concept erasure and MU methods.
-
Mechanistic Analysis of Alignment Algorithms
Mechanistic analysis of how alignment algorithms shape language
model representations. Compared DPO, GRPO, KTO, PPO, and SimPO
with activation probes, sparse autoencoders, and crosscoders.
-
"Fantastic Biases and Where to Find Them"
Mechanistic interpretation of racial and gender bias in
LLMs using linear probing, activation and attribution patching, and circuit tracing. Analyzed
GPT-2 (Medium/Large) and GPT-Neo models on StereoSet and
WinoGender datasets to identify where bias emerges in model
representations.
-
3D Scene Reconstruction
End-to-end 2D to 3D scene reconstruction using classical,
hybrid, and learning-based strategies including SfM, MVS,
GANs, Neural Radiance Fields, and Gaussian Splattings.
-
FastSim with CERN
Developing flow-based, diffeomorphic generative models for
"Fast Shower Simulation (FastSim)" in High Energy Physics.
|