Learning spatio-temporal statistics from the environment in recurrent networks
Abstract
The ability to learn the spatio-temporal statistics of complex environments is essential for many real world tasks. This can in prin"ciple be accomplished by learning the parameters of recurrent neural networks. However, it has proven very difficult to develop algo"rithms to learn such parameters from examples. Yet this does not seem such a difficult task for brain circuits. We propose to develop a theoretical foundation for how brain circuits implement and learn their dynamics from the statistics of the environment. This th"eory will be biophysically realistic, yet using advanced analytical techniques, we will make it as simple as possible. Such a theory" can then be used to improve learning algorithms inartificial systems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 03, 2017
- Source ID
- N000141713004
Entities
People
- Nicolas Brunel
Organizations
- Duke University
- Office of Naval Research
- United States Navy