Research Assistant
CurrentDecoding the Black Box: Enhancing Explainability in Transformer Models (Thesis)- Analyzed the interpretability of Large Language Models (LLMs) under different quantization schemes.Investigated the effect on dead neurons, correlation, and entropy of neurons.- Applied attribution technique Integrated Gradients to identify salient neurons contributing to model predictions.- Followed the work of Bills et al. by prompting an open-source LLM to generate human-understandable explanations of textual patterns learned by neurons.- Validated generated explanations by simulating patterns with another open-source LLM.Enabled safe deployment of LLMs in resource-constrained environments.Unveiling Patterns and Biases: Manifold Learning in NLP Embeddings-Evaluated manifold learning techniques for dimensionality reduction of NLP token embeddings.-Visualized results to identify model-learned biases and assess algorithm performance in capturing linguistic nuances.
