Spiking Neural Networks with Delay Learning (W911NF-19-S-0007)
Abstract
The goal of the project is to design and optimize a novel spiking neural network paradigm for the purposes of solving classification problems and event prediction. Spiking neural networks are biologically-inspired computational engines that are sometimes called the third generation of neural networks. In the proposed approach, a preset number of network spikes arriving approximately simultaneously at a downstream node with individual delay times cause the latter node to spike, sending additional messages downstream, resulting in sustained activity in the network. Training of the network from data occurs through adjusting the delay times with a biologically-inspired protocol called spike-timing-dependent plasticity (STDP). Classification problems are solved by presenting input data as spike volleys to the network, and a class of an input is indicated by a spike at a prearranged network node at a prearranged time instant. The spiking networks, or reservoirs, can be connected in layers for better classification results, and a variety of architectures will be tested and compared. After the neural networks are optimized for classification problems, the successful architectures will be extended to sustained dynamics, and applied to the prediction problem. Current prediction algorithms based on reservoir dynamics developed for times series over the last several years will be adapted to the spiking neural network domain, and tested on a wide range of prediction problems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 01, 2019
- Source ID
- W911NF1910492
Entities
People
- Timothy Sauer
Organizations
- Army Contracting Command
- George Mason University
- National Security Agency