GPU/CPU Cluster to Support Research at the University of Maryland Center for Machine Learning
Abstract
Recent years have seen tremendous success in the application of machine learning to a wide rangeof practical, real-world problems. This success largely stems from the availability of large datasets, and the ability of modern machine learning algorithms to handl e this massive amount ofdigital data. However, training machine learning systems with such large data sets requiresconsiderable co mputational resources. The use of graphical processing units (GPUs), especiallyin learning with deep neural networks has helped spe ed up learning systems, but even so, currentresearch relies on the availability of large clusters of GPU-CPU computational nodes.T his proposal calls for a significant enhancement in the availability of GPU-CPU computationalresources at the University of Marylan d Center for Machine Learning. These resources will beused to speed up the development of algorithms, and to enable the Center to u ndertake newresearch initiatives.Researchers at the UMD Center for Machine Learning currently lead research on a wide varietyof t opics central to Department of Defense concerns. For example, the Center is currentlyengaging in work under three different grants to understand and defend against adversarial attacksagainst machine learning systems. Such attacks can allow malicious actors to s ubtly change theinputs to machine learning systems, causing catastrophic failure. They can potentially threatenweapon or security systems that rely on machine learning, or important systems that may be widelyused in the future, such as self-driving cars. The C enter also conducts research under several grantsthat address the problem of how robots can navigate in challenging environments an d coordinatetheir actions. Work in perception addresses robotics, threat assessment through gait analysis, andmedia forensics, aim ed at identifying faked or manipulated data. Work in the study of the theoryof games can help determine what combination of managem ent techniques and under whatconditions one can successfully create decision outcomes that favor US and Allied interests, whilethe study of online markets can lead to better methods of fairly connecting kidney donors andpatients. The proposed new computational resources will also help support new research aimed atunderstanding how to ensure that machine learning systems are fair and unbia sed, and ensureprivacy, while sustaining reliability.These new resources will also be used to support the University of Marylands educationalmission. The University of Maryland has one of the largest undergraduate computer scienceprograms in the US, with over 3300 majors. In particular, about 235 majors participate in thehonors program, which contains a significant research component. Ma ny UMD undergrads arekeenly interested in participating in research related to machine learning, on many topics relevantto DoD int erests. However, because the number of undergraduates at UMD has quadrupled in thelast few years, it has been difficult for the CS department to keep up with the high demand forcomputing resources. With the additional resources provided by this grant, the ML Ce nter will beable to step in and fill this void, providing access to our cluster to undergraduates engaged in MLrelated research wi th our faculty or staff. Resources will also be used by graduate students,allowing them to begin initiating research projects more rapidly, and supporting their workthroughout their MS or PhD.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 20, 2021
- Source ID
- N000142112860
Entities
People
- David R. Jacobs
Organizations
- Office of Naval Research
- United States Navy
- University of Maryland