Ayush Chopra

Current: I am a PhD student advised by Prof. Ramesh Raskar.  My goal is to build tools that can guide population-level decision making by harnessing collective intelligence. This is the setting where we leverage signal from individual agents (eg: 1 million people in a city) to make decisions that are personalized to cohorts (eg: which COVID-19 test to deploy? who gets COVID-19 vaccine first?). I investigate these problems in context of our emerging data-driven but privacy-sensitive world. 

To achieve this, I focus on advancing technologies that can allow us to realistically simulate large heterogeneous populations; and privately sample distributed data sources. Specifically, my research intersects: (differentiable) multi-agent simulations, (distributed) machine learning and (private) imaging. My projects are motivated by active collaborations with domain experts in public health (Epi), biology (immunology) and marketing. My CV is here and blurb for talks/introductions is here. 

Past: Prior to MIT, I was a researcher at Adobe where I focused on advancing computer vision and machine learning to enable interactive & personalized retail experiences in the browser. I was also the youngest recipient of the Adobe Outstanding Young Engineer Award [My talk at Adobe Marketing Summit 2020].  I have been a technical (AI) advisor at RemoteHQ where we build video platforms for distributed SaaS teams to collaborate productively [RemoteHQ voted #1 on ProductHunt]

Output : My research has been published (and received best paper awards) at several top-tier AI conferences/ journals and has resulted in 25 patents filed across four countries. My projects have been covered by various digital media platforms including TechCrunch, Reuters, Venture Beat, Weather Channel, Ad-Week, Women's Wear Daily, etc.

Collab: We have a lot exciting research and development projects going on! If you would like to know more or collaborate, please reach out to me at { firstname + lastname[0] } {at} {mit.edu}

Research Overview:

i) Agent-based Simulation Modeling

Thesis: The key focus is to invent systems for scalable, fast and differentiable simulations. While current ideas are inspired from work in computer graphics and motivated from challenges in epidemiology, methods are designed for generalizability to multiple fields with social and economic relevance (cancer-biology, marketing and financial economics). Some projects in this direction - 

- GradABM accepted at ICML 2022 Workshop on AI for Agent-based Models as Long Oral (and Best Paper Award!)! General framework for building hybrid DNN-ABM pipelines that can be trained end-to-end. Provides a unified mechanism for calibration, forecasting and policy decision making. 

- DeepABM accepted at WSC 2021 as Oral! Making large-scale agent-based models  fast and differentiable.  Potential for data-driven public policy.

- Public health policy design (collab with Mayo) published at The BMJ 2021. Study public health impact of delaying 2nd dose of mRNA-based vaccine through simulations. Propose recommendations to guide public health policy decisions. 

ii) Private and Distributed Machine Learning - Imaging

Thesis: The key focus is to invent algorithms that can jointly learn from at-home visual sensors (RGB cameras, LiDAR ) - distributed across a network while protecting the privacy of each user. Pursue two lines of investigation: i) learn from siloed data that never leaves the client (Distributed/Split ML), ii)  transform siloed data to censor sensitive information so that it can be released for arbitrary use. Specifically focus on challenges for 2D and 3D visual data. Some projects in this direction - 

-CBNS accepted at ECCV 2022. Mechanisms for conditionally private sampling of 3D point clouds to protect user privacy while preserving perception utility (eg: protect user privacy while: the roomba navigates obstacles when cleaning your house or you use face-id to login to your iPhone; etc). Potential for deploying ML systems in privacy-sensitive environments.

- Sanitizer accepted at ECCV 2022. Mechanisms for private re-sampling of images for task-agnostic data release by protecting sensitive information. 

- AdaSplit (preprint). Distributed ML (that is NOT federated learning!) that scales to low resource scenarios and can jointly learn from heterogeneous devices (phone, watch, home-assistant etc). Significantly outperforms similar federated learning w.r.t both resource utilization and converged performance.

- DISCO accepted at CVPR 2021. Conditionally private inference in deep neural networks. Potential for deploying ML systems in privacy-sensitive environments.

iii) Computer Vision and Deep Learning

Thesis: This is research from a "previous life" when I invented algorithms to make in-browser commerce more interactive and efficient. Pursued two lines of investigation - i) Multi-modal search and retrieval; ii) Image-based Virtual Try-On. A detailed list of these papers is here. Some recent projects -

-SAC accepted at WACV 2022.  Editing visual representations with text feedback. Potential for enabling the last mile of product search. Prior work also includes searching complementary item-sets by modeling catalog as graph (WACV 2020) and utilizing visual content to search by image (CVPR-W 2019, Best Paper Award)

- ZFlow accepted at ICCV 2021. Data-driven estimation of appearance flow field for transforming deformable objects. Preserve geometric integrity to enable photorealistic virtual try-on in the browser! (with a single RGB image). Prior work focused on preserving textural integrity during virtual try-on (WACV 2020)