Amey Pasarkar

I am a PhD student at Princeton University, working in Vertaix on efficient data curation and evaluation algorithms for AI models. I am fortunate to be advised by Professor Adji Bousso Dieng and am generously supported by a NSF Graduate Research Fellowship.

Prior to joining Princeton, I graduated from Columbia University with a degree in Operations Research, where I had the privilege of working under Professor Itsik Pe'er on designing probabalistic graphical models for microbial dynamics. I have also recently interned at Amazon Science on the Fire TV team.

You can contact me at amey [dot] pasarkar [at] princeton [dot] edu.

CV  /  Google Scholar  /  GitHub  /  LinkedIn

Research

Datasets have exploded in size, yet our ability to measure what they actually contain remains limited. I develop scalable methods for analyzing and curating large datasets, with a focus on generative modeling, and efficient training across image, text, and scientific domains.

We introduce an efficient, diversity-based Out-of-Distribution detection algorithm that achieves state-of-the-art performance on numerous image classification datasets and model architectures.

We introduce the Vendiscope, a scalable algorithmic microscope for analyzing the contents of datasets and behaviors of models in any domain where similarity can be defined. The Vendiscope can detect outliers and duplicates, diagnose model failure modes, and reveal memorization patterns across modalities, all at scale.

We generalize the Vendi Score to an entire family of metrics called the "Vendi scores". We demonstrate the utility of these metrics in evaluating image generative models, uncovering patterns of model memorization.

We propose Vendi Sampling, a method for fast sampling from complex Boltzmann distributions that encourages diversity throughout the sampling process. We demonstrate on various molecular systems that Vendi Sampling outperforms standard Langevin Dynamics.

We develop a class of directional Gaussian Mixture models to elucidate the spatial arrangement of the gut microbiome. With no prior information, these models are able to recapitulate known spatial dynamics and propose new ones as well. Research was also presented at Probabalistic Modeling in Genomics, 2021.

We analyzed fixed broadband usage across the United States in the months following COVID-19 lockdowns to determine the effect of lockdowns on internet usage patterns and network performance. We founds signs of increased congestion following significant changes in user behavior.

We develop a multi-layer probabilistic graphical model that can deconvolve microbiome time-series data into its sources of noise. We show our model can be efficiently fit with variational inference while providing more accurate resuls than existing methods. Also presented at RECOMB 2020.