Experience
Education
Bio
Research:
2015-2020: Explaining machine learning decisions
The goal of this project is to develop methods to explain the decisions (e.g. classifications) of machine learning models in terms of input variables (i.e. which input variables are relevant for the model's decision). Techniques that have been developed include the deep Taylor decomposition, that operates by performing a first-order Taylor expansion at each neuron of a deep network. These expansions are then recombined to produce a relevance propagation algorithm, where the model's decision is redistributed from layer to layer until the input is reached.
2012-2013: Machine learning in the chemical compound space
The project consists of using machine learning to dramatically accelerate the calculation of molecular properties (e.g. atomization energy or polarizability) for large collections of molecules. These properties are usually obtained from complex physics simulations, that must be performed for each molecule individually. The machine learning speedup is mainly based on exploiting similarity between molecules, and on automatically extracting a more compact task representation.