Fast Sparese Coding: A New Approach to Accelerate Intelligence Cycle Computations
Abstract
In this effort, we develop mathematical theory and computational algorithms for fast sparse coding that will lead to a new generation of high-performance computer organizations with the potential of speeding up intelligence-cycle computations by orders of magnitude. Sparse coding is a powerful mathematical method in transforming raw input data into sparse feature-space representations capturing prominent patterns learned from the data. It is well known from the literature that sparse representations in the feature space offer more reliable data characterization than operating on raw data directly. Specifically, sparse feature-space transformations can mask noise and variations effectively, while characterizing high dimensional data in terms of a few salient hidden components. Such gains have been demonstrated in classification and prediction tasks in a variety of application domains including computer vision, speech recognition, text understanding, and robotics. However, sparse coding for feature vectors with hundred or more variable dimensionalities is a computational bottleneck that delays applications in automated, real-time incorporation of intelligence for intelligence, surveillance and reconnaissance (ISR). We will develop new, fast algorithms for sparse coding that will alleviate this bottleneck. We will apply the result of this research to deep learning. The proposed work will be conducted by the PI with two of his graduate students and several undergraduates at Harvard. The research grant will be used solely to support basic open research in the proposed fast sparse coding technologies. Fast sparse coding is expected to contribute fundamentally to the development of next-generation computing platforms such as neuromorphic, biologically-inspired computer architectures. Advances in such new computing technology can enhance intelligent systems for humanitarian causes, and benefit the general public in their use of computer applications. To ensure general access to research findings, all results will be published in open literature and disseminated in research communities.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 08, 2016
- Source ID
- N002441510050
Entities
People
- H.t. Kung
Organizations
- President and Fellows of Harvard College
- United States Air Force