Mikhail Belkin

Mikhail Belkin is a Professor at the Halicioglu Data Science Institute and the Computer Science and Engineering Department at the University of California San Diego (UCSD), as well as an Amazon Scholar. His research focuses on foundational aspects of machine learning, particularly deep learning and large language models, emphasizing understanding phenomena like feature learning, over-parameterization, and generalization. He serves as editor-in-chief of the SIAM Journal on Mathematics of Data Science (SIMODS) and was elected an ACM Fellow in 2023.

Research Interests:

• Generalization theory in deep learning
• Optimization dynamics and convergence
• Feature learning mechanisms in neural networks
• Kernel methods and their connections to neural networks
• Double descent phenomenon and model complexity

Key Contributions:

Belkin's work has advanced understanding of interpolation in over-parameterized models, revealing how zero-training-error models can generalize well despite classical statistical concerns. His analyses of optimization landscapes and loss surfaces have clarified why gradient-based methods succeed in deep learning. He introduced the Average Gradient Outer Product (AGOP) framework to explain feature learning in neural networks, bridging classical kernel methods and modern architectures.

Publications: His recent work explores feature learning in neural networks, convolutional architectures, and the theoretical underpinnings of LLMs. Over 15 recent papers are listed, spanning topics like kernel regression, matrix completion, and the mathematical foundations of deep learning.

Awards & Grants:

• ACM Fellow (2023)
• NSF/Simons Collaboration on Theoretical Foundations of Deep Learning
• PI in the TILOS AI Institute (NSF-funded)

Advising & Teams: Belkin collaborates with interdisciplinary teams at UCSD and institutions like the Broad Institute’s Eric and Wendy Schmidt Center, focusing on AI’s societal implications and scientific applications. His work emphasizes bridging theory and practice, with algorithms like Recursive Feature Machines (RFM) and EigenPro showcasing scalable, interpretable machine learning solutions.