The Power and Limitations of Kernel Learning

Explore the capabilities and constraints of kernel learning in this 33-minute lecture by Misha Belkin from Ohio State University. Delve into topics such as shallow architectures, kernel methods for big data, eigenvalue decay, and the EigenPro algorithm. Examine the role of kernel learning in modern machine learning, compare it with state-of-the-art techniques, and gain insights into SGD, batch size optimization for parallel computation, and kernel overfitting. Investigate accelerated methods for kernels and conclude with thought-provoking reflections on the subject.

Intro
The limits and power of kernels
"Shallow"/kernel architectures
Kernel learning for modern ML
Kernel methods for big data
The limits of kernels
Eigenvalue decay
Eigenpro: practical implementation
Comparison with state-of-the-art
Understanding SGD
Batch size for parallel computation
Overfitting with kernels
Kernel overfitting/interpolation
Accelerated methods for kernels
Parting Thoughts