Distributed, Efficient Algorithms for Deep Network Training Without Pretraining
Abstract
Under this grant, PI Goldstein and will study new, alternative methods for training neural networks and performing other large-scale model fitting tasks. The objectives of this grant are (1) develop new distributed algorithms for large-scale model fitting using the alternating direction method of multipliers, (2) adapt this model-fitting paradigm to the task of training deep neural networks on many processors using distributed implementations, and (3) apply the new algorithms to domain-specific applications of interest to the Navy.The implementation built by PI Goldstein~s group will be compatible with the computing resources provided by the Navy DoD Supercomputing Resource Center (DSRC). Domain specific applications will span multiple topics, and will focus largely on image analysis and object identification problems, but may also include automated visual friend-or-foe identification, anomaly detection in autonomous systems, and artificial-or-natural underwater sounds classification.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 04, 2017
- Source ID
- N000141712078
Entities
People
- Thomas Goldstein
Organizations
- Office of Naval Research
- United States Navy
- University of Maryland