I am a fifth-year PhD candidate in Princeton University’s Neuroscience Institute and Center for Statistics & Machine Learning, working at the intersection of Bayesian machine learning, information theory, and decision theory. My research develops statistical methods to understand how neural circuits in the brain approximate complex information and make [sub]optimal decisions.
My recent work has been developing methods for “loss-calibrated” approximate inference in hierarchical models, Bayesian neural networks, and Gaussian processes. In practice, this research explores how to tailor — or calibrate — learning in complex statistical models to an explicit decision-making problems we want to solve. Imagine a classification task in which false negatives are catastrophically costly, while false positives are nearly inconsequential: the statistical relationships in the data don't change, but what we want to learn about them should change a lot. This isn't easily expressed in the typical tools of statistical ML, and my work bridges this gap with classical tools from decision theory.
My grandest vision for these tools is to use them as normative models of neural coding in the context of human decision-making and its various cognitive biases. With perception as a guiding example, the visual/sensory world is compressed and encoded in the brain with information loss, so which information is best to lose? These statistical models formalize the hypothesis that “the optimal neural code preserves information relevant for downstream decision-making”, perhaps at odds with the efficient coding hypothesis that “the optimal neural code maximizes mutual information with the environment”.
Neural population coding in normatively optimal Bayesian models of information processing; lossy compression and "efficient'' coding of visual information
Loss-calibrated approximate inference for models with asymmetric costs and infrequent bad outcomes; developing methods that “calibrate” ML algorithms to decision-making
Latent variable models for sufficient dimension reduction on undersampled data (N < p); Riemannian optimization on matrix manifolds for subspace identification in fMRI data.