GPU Cluster For Bidirectional, Biological Deep Networks
Abstract
Abstract We propose to purchase a 25-node computer cluster with GPU’s (graphical processing units) and the most recent Intel Haswell CPU’s with AVX2 vector processing, to enable us to significantly scale-up our existing biologically-based computational models of deep networks in the brain. We have received extensive funding for this research from ONR over the years, and have current funding for performing human activity recognition by developing structured representations of agent relationships (Tom McKenna, PM), and to develop large-scale models of bidirectional dynamics in the visual system (Harold Hawkins, PM). Our existing models have demonstrated considerable promise for significantly advancing progress in these domains, but we have been limited by existing hardware in the scale of models that we can productively develop. The proposed cluster will enable us to run significantly larger models on significantly larger training sets, and recent work with feedforward deep networks has shown that these models improve dramatically in performance as they are scaled up in these ways. Thus, this hardware will enable us to explore the true potential of our models, which have a number of important advantages over existing feedforward deep networks. The resulting models have the potential to revolutionize a large number of DoD relevant applications, from autonomous robotic applications to automated intelligence processing to understanding how to develop better training programs for military personnel.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 12, 2016
- Source ID
- N000141512832
Entities
People
- Randall O Reilly
Organizations
- Office of Naval Research
- Regents of the University of Colorado
- United States Navy