Nikita Dhawan

I am a PhD student at the University of Toronto and the Vector Institute, supervised by Professors Chris Maddison and Roger Grosse. I completed my Bachelors in Computer Science and Applied Math at UC Berkeley, where I enjoyed working with Professor Sergey Levine.

Email  /  LinkedIn  /  Google Scholar

I am interested in developing algorithms for reliable and trustworthy machine learning, with a particular focus on representation learning, self-supervision and continual learning.

We consider a specific case of approximating the function space distance (FSD) over the training set, i.e. the average distance between the outputs of two ReLU neural networks, based on approximating the architecture as a linear network with stochastic gating. Despite requiring only one parameter per unit of the network, our parametric approximation is competitive with state-of-the-art nonparametric approximations with larger memory requirements, when applied to continual learning and influence function estimation.

We introduce a new dataset inference defense for self-supervised models, which uses the intuition that the log-likelihood of an encoder's output representations is higher on the victim's training data than on test data if it is stolen from the victim, but not if it is independently trained. Our extensive empirical results in the vision domain demonstrate that dataset inference is a promising direction for defending self-supervised models against model stealing.

Recently, ML-as-a-Service providers have commenced offering trained self-supervised models over inference APIs, which transform user inputs into useful representations for a fee. However, the high cost involved to train these models and their exposure over APIs both make black-box extraction a realistic security threat. We explore model stealing by constructing several novel attacks and evaluating existing classes of defenses.

Machine learning systems are regularly tested under distribution shift, in real-life applications. In this work, we consider the setting where the training data are structured into groups and test time shifts correspond to changes in the group distribution. We propose to use ideas from meta-learning to learn models that are adaptable, and introduce the framework of adaptive risk minimization (ARM), a formalization of this setting.

Humans can learn from watching others, imagining how they would perform the task themselves, and then practicing on their own. Can robots do the same? We adopt a similar strategy of imagination and practice in this project to solve complex, long-horizon tasks, like operating a coffee machine or getting objects from within a closed drawer.

Hosted by Nicole Mitchell and Karolina Dziugaite.

Template