I’m a Ph.D. candidate in Machine Learning at Harvard University,
working with Professor Finale Doshi-Velez
at the Data to Actionable Knowledge Lab (DtAK).
I had the pleasure of interning with the Biomedical-ML team at Microsoft Research New England (Summer 2021).
Before joining DtAK, I received a Master’s of Music in Contemporary Improvisation from the New England Conservatory (2016) and a Bachelor’s of Arts in Computer Science from Harvard University (2015). I am currently a performing musician.
I’ve also completed summer internships at Uber (2016), Kensho (2015), and Meta (2013).
Research
My goal is to empower individuals to effectively utilize Machine Learning by making the consequences of modeling assumptions and inference decisions transparent.
Specifically, I develop deep probabilistic/Bayesian models and approximate inference methods designed for safety-critical domains, such as precision healthcare.
My work has exposed failure mechanisms in a number of popular methods, and provides ways to mitigate these failures.
My work has also explored how people (mis)understand popular techniques for explainable AI.
Teaching
My goal is to create classroom environments and mentorship experiences that (a) build student ability to independently interrogate technological systems in societal contexts, (b) that develop effective research skills (e.g. technical reading, writing, communication), and (c) help students overcome common psychological barriers to learning (e.g. imposter phenomenon) using cohort-building and discussion of academic culture. See my teaching & mentorship page for more info.
Selected Publications
For a complete list, see my publications page.
-
Mitigating the Effects of Non-Identifiability on Inference for Bayesian Neural Networks with Latent Variables
Accepted @
JMLR 2022
Previous version accepted @
ICML UDL 2019
Spotlight Talk
Bayesian Neural Networks with Latent Variables (BNN+LVs) capture predictive uncertainty by explicitly modeling model uncertainty (via priors on network weights) and environmental stochasticity (via a latent input noise variable). In this work, we first show that BNN+LV suffers from a serious form of non-identifiability: explanatory power can be transferred between the model parameters and latent variables while fitting the data equally well. We demonstrate that as a result, in the limit of infinite data, the posterior mode over the network weights and latent variables is asymptotically biased away from the ground-truth. Due to this asymptotic bias, traditional inference methods may in practice yield parameters that generalize poorly and misestimate uncertainty. Next, we develop a novel inference procedure that explicitly mitigates the effects of likelihood non-identifiability during training and yields high-quality predictions as well as uncertainty estimates. We demonstrate that our inference method improves upon benchmark methods across a range of synthetic and real data-sets.
-
Failure Modes of Variational Autoencoders and Their Effects on Downstream Tasks
In submission @ JMLR 2022
Previous version accepted @ ICML UDL 2020
Variational Auto-encoders (VAEs) are deep generative latent variable models that are widely used for a number of downstream tasks. While it has been demonstrated that VAE training can suffer from a number of pathologies, existing literature lacks characterizations of exactly when these pathologies occur and how they impact downstream task performance. In this paper, we concretely characterize conditions under which VAE training exhibits pathologies and connect these failure modes to undesirable effects on specific downstream tasks, such as learning compressed and disentangled representations, adversarial robustness, and semi-supervised learning.
-
Characterizing and Avoiding Problematic Global Optima of Variational Autoencoders
Accepted @
AABI 2019
Spotlight Talk
Additionally selected for publication @
PMLR 2019
Top 33%
Variational Auto-encoders (VAEs) are deep generative latent variable models consisting of two components: a generative model that captures a data distribution p(x) by transforming a distribution p(z) over latent space, and an inference model that infers likely latent codes for each data point. Recent work shows that traditional training methods tend to yield solutions that violate modeling desiderata: (1) the learned generative model captures the observed data distribution but does so while ignoring the latent codes, resulting in codes that do not represent the data; (2) the aggregate of the learned latent codes does not match the prior p(z). This mismatch means that the learned generative model will be unable to generate realistic data with samples from p(z). In this paper, we demonstrate that both issues stem from the fact that the global optima of the VAE training objective often correspond to undesirable solutions. Our analysis builds on two observations: (1) the generative model is unidentifiable - there exist many generative models that explain the data equally well, each with different (and potentially unwanted) properties and (2) bias in the VAE objective - the VAE objective may prefer generative models that explain the data poorly but have posteriors that are easy to approximate. We present a novel inference method, LiBI, mitigating the problems identified in our analysis. On synthetic datasets, we show that LiBI can learn generative models that capture the data distribution and inference models that better satisfy modeling assumptions when traditional methods struggle to do so.
