Machine Learning Research by Mike

Welcome! My research spans statistical machine learning and its applications in healthcare and the sciences.

I am currently a postdoctoral fellow in computer science at Harvard SEAS, advised by Prof. Finale Doshi-Velez. I completed my Ph.D. in computer science at Brown University in May 2016, advised by Prof. Erik Sudderth.

Job Search!

I am looking for tenure-track faculty positions (2017-18). Please reach out if you have questions.

Research Summary

Recently, I've been motivated by two exciting clinical applications:

  • forecasting need for interventions in the Intensive Care Unit (ICU)
  • antidepressant recommendations for patients with major depression

These applications have inspired new contributions to core ML methods:

  • Semi-supervised learning: Our new Prediction Constrained training objective fits latent variable models so that they provide accurate predictions (e.g. drug recommendations) and interpretable generative models, even when labeled examples are rare.

  • Explainable AI: Our new Tree Regularization method lets you optimize deep neural networks so learned class boundaries are similar to decision trees.

My Ph.D. thesis work was motivated by several applied questions:

  • can we find clusters of cooccuring words that thematically organize every New York Times article from the last 20 years?
  • can we find clusters of cooccuring epigenetic modifiers that amplify or inhibit gene expression?

To answer these questions, we developed new variational inference algorithms for a broad family of Bayesian nonparametric models that include mixtures, topic models, sequential models, and relational models. Our key innovations include scaling to millions of examples and adding data-driven split/merge proposal moves to avoid poor local minima.

Please try out BNPy, our open-source Python package.