Worked on developing attribution techniques using Integrated Gradients, Manifold IG, and Guided IG towards neural network interpretability.
- Currently extending these methods to sequential models and large language models (LLMs) using Granger causality and Randomized Path Integrals.
- Reproduced the results of Integrated Gradients and Manifold Integrated Gradients and tested a new methodology on MIG to improve performance.
- Future work involves computer vision applications.