A Future of Abundant Sparsity: Novel Use and Analysis of Sparse Coding in Machine Learning Applications
View/ Open
COMITER-SENIORTHESIS-2015.pdf (16.06Mb)
Access Status
Full text of the requested work is not available in DASH at this time ("restricted access"). For more information on restricted deposits, see our FAQ.Author
Metadata
Show full item recordCitation
Comiter, Marcus Zachary. 2015. A Future of Abundant Sparsity: Novel Use and Analysis of Sparse Coding in Machine Learning Applications. Bachelor's thesis, Harvard College.Abstract
We present novel applications and analysis of the use of sparse coding within the con- text of machine learning. We first present Sparse Coding Trees (SC-trees), a sparse coding-based framework for resolving classification conflicts, which occur when different classes are mapped to similar feature representations. More specifically, SC-trees are novel supervised hierarchical clustering trees that use node specific dictionary and classifier training to direct input images based on classification results in the feature space at each node. We validate SC-trees on image-based emotion classification, combining it with Mirrored Nonnegative Sparse Coding (MNNSC), a novel sparse coding algorithm leveraging a nonnegativity constraint and the inherent symmetry of the domain, to achieve results exceeding or competitive with the state-of-the-art. We next present SILQ, a sparse coding-based link state model that can predictively buffer packets during wireless link outages to avoid disruption to higher layer protocols such as TCP. We demonstrate empirically that SILQ increases TCP throughput by a factor of 2-4x in varied scenarios.Citable link to this page
http://nrs.harvard.edu/urn-3:HUL.InstRepos:17417575
Collections
- FAS Theses and Dissertations [6097]
Contact administrator regarding this item (to report mistakes or request changes)