 |
General: 663 |
P6: Analyzing Task Driven Learning Algorithms
Author: Mike Pekala , Advisor: Doron Levy (Mathematics and CSCAMM)
Problem Statement Presentation
Project Proposal
Abstract
The underlying notion of sparse coding is that, in many domains, data vectors can be concisely represented as a sparse linear combina- tion of basis elements or dictionary atoms. Recent results suggest that, for many tasks involving compressible data (e.g. classi cation, denois- ing), performance on that task can be improved by explicitly learning the sparsifying dictionary directly from the data [15, 16]. This is in contrast with classical approaches, such as wavelets, that use prede- ned dictionaries known to work well on a broad class of signals. Fur- thermore, results also suggest that additional performance gains are possible by jointly optimizing the dictionary for both the data and the task of interest [12]. In this project, we propose to implement the task-driven dictionary learning algorithm of [12] with a focus on the binary classi cation task. We will verify the correctness of our implementation through a combination of unit testing and comparison with published results on open data sets. Finally, we shall apply this algorithm to data sets not considered by [12].
MidYear Progress Report and
Presentation
Final Presentation ,
Final Report
