Raman Arora

Raman Arora is an associate professor of computer science at Johns Hopkins University whose group designs machine learning algorithms with formal guarantees under constraints such as privacy, robustness and limited feedback. After earning a PhD at the University of Wisconsin–Madison in 2009, he held research posts at the Toyota Technological Institute at Chicago, Microsoft Research and the University of Washington. His early research included coauthoring Deep Canonical Correlation Analysis at ICML 2013, which was named a runner-up for the ICML 2023 Test-of-Time award, and a 2016 algorithm for training one-hidden-layer ReLU neural networks to global optimality. During 2019–2022, he was a member at the Institute for Advanced Study and a visiting scientist at the Simons Institute for the Theory of Computing, and in 2020 he received an NSF CAREER Award for work on inductive biases in machine learning. At ICML 2025, Arora and his coauthors presented theoretical analyses of backdoor attacks on language model attention mechanisms and policy regret minimization against adaptive opponents.

Now at: Johns Hopkins University
Works: 128
Citations: 4,700
h-index: 28
Badges: Wayfarer

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

2009

PhD at Wisconsin–Madison

In 2009, he earns a PhD at the University of Wisconsin–Madison. Before joining Johns Hopkins, he works as a research assistant professor and postdoctoral scholar at the Toyota Technological Institute at Chicago, a visiting researcher at Microsoft Research in Redmond, and a research associate at the University of Washington in Seattle.
2013

Deep canonical correlation analysis

In 2013, he coauthors Deep Canonical Correlation Analysis with Galen Andrew, Jeff Bilmes and Karen Livescu at the 30th International Conference on Machine Learning. The method learns complex nonlinear transformations of two views of data so that the resulting representations are highly linearly correlated.
2016

ReLU network optimality guarantees

With Amitabh Basu, Poorya Mianjy and Anirbit Mukherjee in 2016, he posts “Understanding Deep Neural Networks with Rectified Linear Units.” The work gives an algorithm that trains a one-hidden-layer ReLU deep neural network to global optimality in time polynomial in the data size, alongside depth-size gap theorems.
2019–2022

IAS member and Simons visitor

During 2019–2020, he is a member of the Department of Mathematics at the Institute for Advanced Study. He also serves as a visiting scientist at the Simons Institute for the Theory of Computing in summer 2019 for Foundations of Deep Learning, in fall 2020 for Theory of Reinforcement Learning, and in spring 2022 for Learning and Games.
2020

An NSF CAREER Award

In March 2020, he receives an NSF CAREER Award as an assistant professor of computer science at Johns Hopkins. The five-year award supports the project “Understanding Inductive Biases in Modern Machine Learning,” which develops an explanatory and prescriptive theory of deep learning closely tied to practice.
2023

ICML Test-of-Time runner-up

In 2023, Deep Canonical Correlation Analysis is named one of two runners-up for the ICML 2023 Test-of-Time award, which recognizes papers from ICML 2013. The program chairs credit the paper with a principled approach to multimodal representation learning with a deep neural network.
2025

Backdoor attacks and policy regret

At the 42nd International Conference on Machine Learning in 2025, Arora and student Yunjuan Wang show how adversaries can hide backdoor behavior in large language models by manipulating token selection in the attention mechanism, and identify the theoretical conditions under which such attacks succeed. With former postdoctoral researcher Thanh Nguyen-Tang, he also develops policy regret minimization at ICML 2025, giving the first algorithms that reason about an adaptive opponent's reactions in complex environments such as games or simulations with many states and actions.
As of October 2026

Associate professor at Johns Hopkins

He is an associate professor in the Department of Computer Science at Johns Hopkins University, affiliated with the Mathematical Institute for Data Science, the Center for Language and Speech Processing, the Institute for Data-Intensive Engineering and Science and the Data Science and AI Institute. His group designs algorithms and proves formal guarantees about their performance, asking how much computation, memory and data an intelligent system needs under real-world constraints such as privacy, robustness or limited feedback.