I am a Postdoctoral Research Fellow in the Department of Epidemiology at Harvard, working in the area of AI/machine learning for medicine and public health, with Andrew Beam. I received my Ph.D. in Computer Science from Harvard in 2019. My current research in AI/machine learning and natural language processing is at the intersection of metric learning and deep learning, investigating the use of "non-parametric" structures to improve the effectiveness, (data) efficiency, and interpretability of neural network models. Specifically, my current research focuses on building out the "exemplar auditing" framework for AI/ML/data analysis (Schmaltz 2019; Schmaltz and Beam 2020; slides). We are also actively working on medical QA, which is an intermediate point toward the long-term (aspirational) goal of more general automated medical reasoning.
More generally, my AI research interests can be viewed through the lens of "learning in the presence of label resolution disparity". Often we have labels at a coarse granularity, but we seek an analysis of features at a more fine-grained resolution.
Deep neural networks occupy a curious point in the epistemological space toward this end: The parameters are not identifiable and are not explainable in the same manner as, for example, low-parameter, low-variable linear regression models. However, they are very powerful pattern discoverers, and we can encode data priors in a semi-supervised manner with imputation-style losses with deep networks—to a degree not (at least readily) possible with more parsimonious approaches. We can then use methods such as exemplar auditing to introspect the data at resolutions more fine-grained than our training labels, revealing insights in very high dimensional spaces that might not otherwise be readily detectable.
As a fun aside, I am the creator of reorder.coffee, an educational game for advanced language learners.