Adam Fisch

I am a Research Scientist at Google Deepmind where I work on developing methods for efficient and reliable machine learning. Machine learning has seen tremendous progress over the last decade, yet many challenges remain. How can we best deploy existing models for use in open-domain settings? How can we mitigate the negative consequences of their errors? How can we make more efficient predictions at test time with models that have millions (or billions, or trillions) of parameters? Finding convincing answers to these questions has driven much of my past research. I am especially focused on developing rigorous tools for estimating uncertainty in order to prepare users to safely use deployed models in realistic situations where models inevitably make mistakes. At the same time, I am also interested in how to leverage uncertainty estimation to make more efficient predictions by taking the opposite approach: for easy inputs, it can pay to be less conservative, and to choose to use less expensive, simpler functions to make a prediction---but still ensure that any degradation suffered to performance is tightly bounded.

My work draws on both theoretical and empirical techniques, with the goal of creating principled and effective solutions with solid mathematical foundations. Applications of my work include language modeling, machine reading of text for automatic question answering, fact verification, lung cancer risk assessment, and in-silico screening for drug discovery.

I completed my Ph.D. student in Electrical Engineering and Computer Science at MIT, where I was advised by Professor Regina Barzilay, and also worked closely with Professor Tommi Jaakkola. Before coming to MIT, I worked as a research engineer at Meta AI Research.