Research Seminar on AI: Path classification by linear stochastic RNNs
- Send Email
Tuesday, April 19, 2022, 4:00pm
Youness Boutaib and his team investigate the functioning of a classifying biological neural network from the perspective of statistical learning theory, modelled, in a simplified setting, as a continuous-time stochastic recurrent neural network (RNN) with identity activation function. In the purely stochastic (robust) regime, Youness Boutaib and his team give a generalisation error bound that holds with high probability, thus showing that the empirical risk minimiser is the best-in-class hypothesis. It is shown that RNNs retain a partial signature of the paths they are fed as the unique information exploited for training and classification tasks. Youness Boutaib and his team argue that these RNNs are easy to train and robust and back these observations with numerical experiments on both synthetic and real data. A trade-off phenomenon between accuracy and robustness is also exhibit.
Youness Boutaib did his PhD in mathematics at the University of Oxford after studying and working in quantitative finance in France. His research topic was the theory of rough paths, which is a theory that studies differential systems driven by highly oscillatory signals, like the ones found in stochastic analysis, but in a deterministic manner. He continued his research on this topic in his first post-doc between TU Berlin and the University of Potsdam, then in 2019 I he decided to move to the field of Machine Learning, and in particular the mathematical foundations thereof. Of particular interest are the machine learning methods that are dedicated to streaming data, which allow him to use his previous experience and knowledge in the field of dynamical systems.