Kunvar Thaman
Kunvar is a machine learning research engineer focused on interpretability of deep neural networks, and is currently working on addressing data contamination issues in the benchmarking of large language models. Previously, he worked as a cybersecurity engineer amd software developer at Akamai Technologies, where he specialized in securing cloud-native applications and implementing real-time solutions for detecting and mitigating sophisticated attacks. He was a Google Summer of Code Fellow during his Electrical and Electronics Engineering degree and also holds a master's in Biological Sciences.
Kunvar investigates deep NNs with a focus on decoding the complex structures and algorithms learned by trained networks, frequently referred to as mechanistic interpretability. He is also interested in structuring and curating training data, developing representation engineering techniques, and ensuring high quality evaluations. Additionally, he is exploring strategies to enhance adversarial robustness using meta-learning techniques.
Kunvar will work on attention head superposition following his hackathon project, and is also pursuing another work focused on addressing data leakage and benchmark contamination.
