Closing the identifiability gap: Interpretable and reproducible scientific inference across modalities, scales and time
Department of Electrical and Computer Engineering Seminar
Speaker: Steffen Schneider
International Max Planck Research School for Intelligent Systems in Tuebingen and EPFL
Abstract: A fundamental goal in science is understanding complex relationships of biological systems across scales and time. Modern life science research is enabling data collection at a rapid pace and at increasing scale, yet our ability to understand complex systems and reason about their underlying dynamics is still limited. To this end, my work has focused on developing new machine learning tools for inferring latent structure from the data we observe. In my talk, I will primarily focus on new work developing a variant of contrastive learning suitable for scientific inference (CEBRA). As an example, I will highlight the algorithm鈥檚 ability to uncover consistent and robust neural latent dynamics. Lastly, I will discuss theoretical foundations for such models, and discuss my prior work in speech and vision, specifically regarding data efficiency and model robustness.
Biography: Steffen Schneider is a final year ELLIS PhD student in Machine Learning and Computational Neuroscience at the International Max Planck Research School for Intelligent Systems in Tuebingen and at EPFL in Geneva. His PhD work is supported by a Google PhD Fellowship and revolves around robust deployment of machine learning models and self-supervised learning for scientific data analysis. Mr. Schneider has also worked on large scale machine learning for speech, vision and language with research visits at FAIR/Meta AI in Menlo Park/NYC, and on object-centric learning at Amazon in Tuebingen. Beyond research, he co-founded the edtech/science communication startup 鈥淜I macht Schule鈥� to teach AI in schools.
