Co-Director, Autonomous Learning Lab
Associate Editor, Journal of Machine Learning Research
Research Areas:
Artificial Intelligence
Representation Discovery
| |
Co-Director, Autonomous Learning Lab Associate Editor, Journal of Machine Learning Research Research Areas: Artificial Intelligence Representation Discovery |
|---|
My research interests are broadly in the area of artificial intelligence. My past work has spanned a number of areas in computational models of learning and sequential decision-making. My current research is focused on the problem of representation discovery: the construction of basis functions that capture regularities like symmetries and bottlenecks in the data or state space. The basis functions constructed are useful across multiple learning and decision-making modalities, from clustering and (semi-)supervised learning to reinforcement learning.
My current research into representation discovery is based on harmonic analysis, a subfield of mathematics where spatial and temporal data is remapped into a frequency oriented coordinate system. I am exploring both global Fourier techniques based on diagonalization principles, such as eigenvector representations (e.g. Laplacian eigenfunctions), as well as multiscale wavelet methods, such as diffusion wavelet analysis. I am also exploring group representation theory for building compact basis functions on large "symmetric" spaces.
One application of representation discovery is a unified framework for the credit assignment problem, wherein agents jointly learn representation and control to solve sequential decision-making tasks. In this context, the basis functions are referred to as proto-value functions, since they are most often used to approximate task-specific value functions. Proto-value functions can either be global (using the Fourier approach) or local (using the wavelet approach). The basis functions can be task-independent or can be influenced by task-specific goals and reward information. Since the basis functions are discovered from experience, they reflect the underlying geometry of the environment. My students are exploring a range of projects including multiscale dimensionalty reduction approaches, factorization methods for compressing basis functions, and basis representations based on directed graphs that jointly model state and action spaces.
Contact InformationSridhar MahadevanDepartment of Computer Science 140 Governor's Drive University of Massachusetts Amherst, MA 01003 Administrative Assistant: Gwyn Mitchell, mitchell@cs.umass.edu VOICE: (413)545-3140
|
|---|