Sotabase

Career

· Mentor, LASR: London AI Safety Research Labs2025–
· Research Scientist, Alignment Team, UK AI Security Institute2025–
· Mentor, AI Safety Camp2023–2023
· PhD Student, Center for Data Science, New York University2022–
· MSc Mind, Language and Embodied Cognition, University of Edinburgh2021–2022
· Bold and Basic Fellow, University of California, San Francisco. Keiser Lab2019–2020
· M1, Masters Computer Science, Data Science track, Ecole Polytechnique (Université Paris-Saclay)2018–2019
· BA Mathematics & 5 College Logic Certificate, Amherst College2013–2017

Publications (27)

Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback
Trans. Mach. Learn. Res. · 2023
738
cited
Artificial Intelligence in Dermatology: A Primer.
Journal of Investigative Dermatology · 2020
172
cited
Goal Misgeneralization in Deep Reinforcement Learning
International Conference on Machine Learning · 2021
115
cited
Stress testing reveals gaps in clinic readiness of image-based diagnostic artificial intelligence models
npj Digital Medicine · 2021
39
cited
Artificial Intelligence in Teledermatology
Current Dermatology Reports · 2019
15
cited
Global Saliency: Aggregating Saliency Maps to Assess Dataset Artefact Bias
arXiv.org · 2019
15
cited
Objective Robustness in Deep Reinforcement Learning
arXiv.org · 2021
15
cited
Robust Semantic Interpretability: Revisiting Concept Activation Vectors
arXiv.org · 2021
15
cited
383 Assessing deep learning artefact bias using global saliency
Journal of Investigative Dermatology · 2020
836 Calibration performance of deep neural networks for image classification declines on real-world, versus curated, test sets
2020
902 Assessing performance of deep neural networks used for image classification by stress testing
2020
A SMART Mnemonic Sounds like
. A SMART Mnemonic Sounds like ”Glue Tonic”: Mixing LLMs with Student Feedback to
Training Data Extraction Attack from Large Language Models in Federated Learning Through Frequent Sequence Mining
Sotabase