July 22, 2022

Meta Research PhD Fellowship Spotlight: Making the most of data with meta-learning

By: Meta Research

Each year, PhD students from around the world apply for the Meta PhD Fellowship, a program designed to encourage and support doctoral students engaged in innovative and relevant research in areas related to computer science and engineering.

As a continuation of our Fellowship spotlight series, we’re highlighting Misha Khodak, a 2021 Meta Research PhD Fellow in machine learning.

Misha is a PhD student in computer science at Carnegie Mellon University, advised by Nina Balcan and Ameet Talwalkar. His research is centered on foundations and applications of machine learning, with a focus on the theoretical and practical understanding of meta-learning and automation. He also works on related areas such as neural architecture search, federated learning, and algorithms with predictions.

When Misha began his PhD program, machine learning was rapidly evolving. He was enthralled by topics like machine translation, image generation, and meta-learning, and even more interested in how these modern approaches work.

“Meta-learning is learning to learn,” explains Misha. “You can use data from multiple tasks to learn how to improve the performance of other tasks. Say you have a collection of datasets on which you want to run some learning algorithm, but you don’t know which one. You can treat this collection itself as a dataset and run a learning algorithm to determine which learning algorithm works best. From there, you can apply that learned algorithm to future learning tasks. That’s meta-learning. It’s a powerful way to improve performance on tasks that don’t have much data—for instance, if you want to translate rare languages, each of which has limited text.”

In addition to theory, Misha envisions practical deployments for meta-learning, such as helping small hospitals diagnose medication conditions. According to Misha, many hospitals want to improve diagnoses and treatments using artificial intelligence (AI), but they can’t share data with other hospitals because of privacy regulations. Using meta-learning, researchers could create a hub of accessible information and AI tools without sending revealing information.

Misha’s research has evolved to include not only learning tasks but also algorithms. “Numerical algorithms run at large companies like Meta can also be used for important physical simulations, which help engineers design power plant reactors, planes, and ships,” says Misha. “But running simulations is expensive and requires supercomputing clusters. Instead, we can use the data from past runs to inform future ones, which saves computation, improves runtime, and increases accuracy.”

Misha’s goal is to expand both the body of knowledge and the applicability of these ideas, which he’s already begun to do. When it comes to theory, Misha’s work has empowered research teams to prove new meta-learning guarantees in areas such as reinforcement learning, distributed computing, and combinatorial optimization. In practice, his research has led to new methods for tuning hyperparameters in federated learning, where researchers train machine learning models on a network of heterogeneous devices, like a network of cellphone users.

Both avenues inspire Misha to take his career in multiple directions. “Industry and academia both enable my goals and bring opportunities to mentor other people. I’ve enjoyed working with many master’s students as part of my PhD, and the Meta PhD Fellowship has been instrumental in putting me on researchers’ radars.”

As a Meta PhD Fellow, Misha has collaborated with Meta researchers in computer science, machine learning, and federated learning. “The Fellowship lasts two years, but I anticipate my relationships with researchers will extend beyond that. The program is all about bringing people together for mutual opportunities.”

To learn more about Misha Khodak and his research, visit his Fellowship page.