• Developed Describe-and-Dissect, a state-of-the-art method that generates complex natural language descriptions for the functionalities of deep neural network neurons to improve interpretability/reliability of deep vision models

• Outperformed existing methods in a 3,000-rating human evaluation study, being selected as the best label over 2× more often (54.7% vs. ≤ 23.1%)

• Published in TMLR (2025) and presented at the XAI4CV and ICML 2024 MI (spotlight) Workshops