Yaniv Romano

Modern machine learning (ML) models produce valuable predictions across various applications, influencing people’s lives, opportunities, and scientific advancements. However, these systems can fail in unexpected ways, generating unreliable inferences and perpetuating biases present in the data. These issues are particularly troubling in high-stakes applications, where models are trained on increasingly diverse, incomplete, and noisy data and then deployed in dynamic environments—conditions that often exacerbate test-time failures.

My team and I strive to create AI systems that can be deployed safely and effectively in the real world.

Our approach is built on two key pillars. The first is the development of protective ecosystems that can seamlessly integrate with any ML model, providing formal statistical guarantees on the reliability of “black-box” data-driven inference. The second is the design of new learning paradigms grounded in fundamental statistical principles. What makes this approach to safe and powerful ML particularly appealing is that it is both highly applicable and grounded in solid theoretical foundations:

• Prediction with confidence, by developing methodologies for distribution-free uncertainty estimation that work with any complex predictive algorithm, such as neural nets and random forests. The uncertainty estimation methodology I invented together with Emmanuel Candès and Evan Patterson at Stanford, called CQR, was used by The Washington Post to estimate outstanding votes for the 2020 U.S. presidential election.

• Reproducibility in science, leveraging state-of-the-art techniques in machine learning and selective inference to make new replicable discoveries.

Other topics that I've worked on include:

• Theory and practice of neural nets, by proposing and analyzing multi-layer convolutional sparse coding techniques and studying the stability of convolutional neural networks.

• Inverse problems, developing highly effective priors and algorithms for solving general image reconstruction problems. The super-resolution technology I invented together with Peyman Milanfar, called RAISR, is being used in Google's flagship products (Pixel 2/XL Phones, Google Clips, Google+, and Motion Stills), increasing the quality of billions of images and bringing significant bandwidth savings.