Learning spatio-temporal statistics from the environment in recurrent networks
Abstract
Statement of Work:This project aims to build models of recurrent neuronal networks that will be able to learn dynamical patterns of activity from the spatio-temporal statistics of their inputs. The work shall include building learning rules that are consistent with a large body of recent experimental data; to simulate large recurrent neuronal networks that implement these learning rules; and to analyze mathematically the dynamics of such networks. Objective:The objective is to train recurrent neural networks (RNN~s) using real world statistics. This has proven a difficult problem in machine learning yet the brain seems to be able to accomplish this effortlessly. The PIs propose to develop algorithms, motivated by the expanding knowledge of Neuroscience in order to accomplish this task. Approach:The PIs will use both large-scale simulations of networks of stochastic spiking neurons, and simpler continuous models that can be derived from networks of spiking neurons using methods from mean-field theory (MFT). The correspondence between the spiking models and their MFT analogs will make it possible to both use simpler tractable models, and keep in contact with experimental reality. The low dimensional continuous models derived via MFT are simple to understand and manipulate and computationally inexpensive. With a continuous lower dimensional system the standard dynamical systems methods can be used to analyze the dynamics, and this is much more difficult with the stochastic spiking networks. It is important to use stochastic networks because trained stochastic networks are likely to be more robust than deterministic networks. Furthermore, stochasticity might be a critical component that allows learning to work, as in reinforcement learning algorithms that rely on a source of stochasticity to perform stochastic gradient ascent of a reward function. The PIs have subdivided this complex problem into three tasks: 1. Understanding how networks learn spatio-temporal associations between stimulus and reward. 2. Learning the order of spatiotemporal sequences. 3. Learning both the order and the timing of spatio-temporal sequences.Overall Merit and ONR Mission/Relevance:This proposal has future relevance to the Navy because many real-world problems, encountered by the Navy, could be addressed by the type of plastic dynamical systems that will be analyzed in this proposal. Examples of such problems include, the recognition of targets that are not static but move and change over time, or self-navigation in complex and changing environments. To approximate, emulate and predict in such cases one must be able to rapidly and robustly train complex dynamical systems using real world examples. This is exactly the type of problems brains seem able tosolve easily, but that current algorithms often fail to solve, or are slow and non-robust.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 08, 2016
- Source ID
- N000141612327
Entities
People
- Nicolas Brunel
Organizations
- Office of Naval Research
- United States Navy
- University of Chicago