Ayush Chopra

Research Assistant
  • Camera Culture

Current:  I am a PhD student advised by Prof. Ramesh Raskar where I study AI and complex systems. My goal is to build 'large population models' that collectively learn from millions of interacting agents to guide decision making. These agents can be humans in the physical world (eg: that spread infections), cells in the biological world (eg: that spread tumors) or even AI avatars in the digital world (eg: that spread information). I want to realize a future where we (actively and securely) aggregate insights from these individual agents (eg: 1 million people in a city) to make decisions personalized to cohorts (eg: who should we vaccinate first? what kind of COVID-19 test to deploy?). To achieve this, I build software systems, invent algorithms and work with global policy organizations to deploy solutions.

Specifically, my thesis is advancing work in agent-based modeling and collaborative machine learning to unlock the benefit of large population models. My projects are motivated by active collaborations with domain experts in epidemiology, finance and immunology. My CV is here and intro blurb 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 browser experiences. 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 advisor at RemoteHQ (acquired by Presence) 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 Modeling

Thesis: Unlock digital experimentation on real populations through scalable, differentiable and private mechanisms to simulate, calibrate and analyze agent-based models.

a) Systems

- AgentTorch accepted at AAMAS 2024 as Oral! A framework to design differentiable agent-based models across  biological, physical and digital realms. Simple Python-API, scales to millions of agents, compatible with automatic differentiation and allows seamless integration of ABMs and DNNs. [github]

b) Methods

- Private ABM accepted at AAMAS 2024 as Oral! Simulate, calibrate and analyze million-scale agent-based models WITHOUT ever accessing the attributes or interaction trace of any individual agent. Unlocks secure and active digital experimentation with real-time data streams.

 -Don't Simulate Twice accepted at AAMAS 2023 as Oral! A protocol to conduct sensitivity analyzes on GradABMs without running any simulations, using automatic differentiation.

- GradABM accepted at AAMAS 2023 as Oral! A protocol to calibrate agent-based models using automatic differentiation and by designing end-to-end DNN-ABM pipelines. Also received the Best Paper Award at ICML 2022 Workshop on AI for Agent-based Models.  Follow-up work (Bayesian-GradABM) to calibrate posteriors using normalizing flows was accepted at AI4ABM at ICLR 2023. 

DeepABM accepted at WSC 2021 as Oral! A protocol to design tensorized and differentiable agent-based models that quickly scale to million-size populations and execute efficiently on GPUs. 

c) Applications

Supply-constrained immunization  published at The BMJ 2021. GradABM to design immunization policies under supply chain and behavior constraints. Recommend delaying second dose of mRNA COVID vaccine (3 weeks -> 3 months) to maximize population benefit. Adopted globally. 

- Geography-specific immunization published at Vaccine 2023. GradABM to personalize immunization policies across different  geographic regions based on local demographic and behavioral dynamics.

- Systemic-risk analysis published at ICAIF 2023. GradABM to model systemic risk in financial markets and calibrate via forward differentiation to enable capture of million-scale retail activity.

ii) Collaborative Machine Learning

Thesis: Data is key for simulation and analysis, but is often siloed across the collaborating agents. We build algorithms that enable agents to: a) collaboratively learn from the data, while keeping it siloed, b) desensitize the siloed data so that it can be aggregated. 

AdaSplit accepted at FLSys 2023. Enables collaboration between heterogeneous agents (phone, watch, home-assistant etc) with access to variable (computation, communication) resources. Provides a metric to evaluate 'collaboration efficiency' and scales algorithms extremely low-resource scenarios.

-CBNS and Sanitizer accepted at ECCV 2022. Provides mechanisms for private sampling of structured data. CBNS focuses on 3D point clouds while Sanitizer focuses on 2D images. Securely delegate decision making to autonomous agents by making them 'conditionally blind'.

DISCO accepted at CVPR 2021. Collaborative inference in deep neural networks. Multiple autonomous agents - one owns data, other owns model, collaborate while protecting both data and model privacy.