Jatin Prakash

I'm a first second year CS PhD student at New York University advised by Prof. Rajesh Ranganath. I am also part of the CILVR Lab.

I am broadly interested in designing scalable, practical and efficient architectures and training algorithms: both pre-training and post-training.

I wish to accomplish this by tackling every part of the modeling stack: from data [1, 2], to model architectures [3], to training algorithms [4], and system optimizations [5].

Previously, I spent two amazing years at Microsoft Research, where I was advised by Dr. Manik Varma and Dr. Amit Sharma. Some of my research during this time found its way into Microsoft Bing [2, 5].

Even before that, I graduated from IIT Delhi with a bachelors in CS. During my undergrad, I worked with Prof. Chetan Arora.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

tldr; We propose an architecture that provides a knob to control quality-efficiency trade-offs directly at test-time, without requiring any retraining. The proposed adaptive model outperforms efficient baselines across varying compute-memory budgets, all using a single model only.

tldr; We benchmarked recent RL algorithms on a simple star-graph task where they fail in zero reward scenarios, even those specially designed for this case. Turns out, a very simple data-centric intervention of just adding easy samples of the task helps unlock RL training. Open-sourced implementations for many RL baselines (that had no official code) for the community to build upon.

tldr; We understand diversity collapse problem in RL, and how to principally fix it in 2 lines of code. Key idea is viewing KL-regularized RL as distribution matching to a target distribution. Our work explores how to define a good target for the proposal distribution (or policy in case of RL) that avoids mode collapse.

tldr; A simple, scalable and data-centric algorithm to mitigate bad quality click-data problem in retrieval (extreme classification), that scales to real-world industry query-ads datasets containing upto 10M+ documents. This outperforms SOTA significantly, highlighting the importance of good quality dataset (that contains diverse world knowledge) for retrieval. Part of this work has been deployed in Microsoft Bing.

tldr; We unlock end-to-end training of large-scale retrieval (extreme classification) models that scales to 100M+ documents and 1B+ training examples, reducing training time from weeks to under a day. Turns out, simple end-to-end learning outperforms complicated, modular SOTA methods. This work has been deployed in Microsoft Bing.