Abstract:

The interpretability of machine learning models for molecules is critical to scientific discovery, understanding, and debugging. Attribution is one approach to interpretability, which highlights parts of the input that are influential to a neural network’s prediction. With molecules, we can set up synthetic tasks such as the identification of subfragment logics to generate ground truth attributions and labels. This scenario serves as a testbed to quantitatively study attributions of molecular graphs with Graph Neural Networks (GNNs). We perform multiple experiments looking at the effect of GNN architectures, label noise, and spurious correlations in attributions. In the end, we make concrete recommendations for which attribution methods and models to use while also providing a framework for evaluating new attribution techniques.

Biography: I am a research scientist at Google Research. My research centers around using machine learning techniques to build data-driven models for the prediction of molecular properties and the generation of new molecules and materials via generative models. Applications include solar cells, solubility, drug-design, and particularly smelly molecules. I am part of a team that wants to do for olfaction, what machine learning has done for vision and speech.

I am also passionate about science education and divulgation, I am one of the founders and organizers for Clubes de Ciencia Mexico and a LatinX-centered AI conference RIIAA. In my free time, I like to run, eat ice cream and cook food.