DESIGN OF COMPACT AND DISCRIMINATIVE DICTIONARIES

Wilson, S and C, Krishna Mohan (2018) DESIGN OF COMPACT AND DISCRIMINATIVE DICTIONARIES. PhD thesis, Indian Institute of Technology, Hyderabad.

Preview

Text
Thesis_Phd_CS_3817.pdf - Submitted Version
Download (4MB) | Preview

Abstract

The objective of this research work is to design compact and discriminative dictionaries for e�ective classi�cation. The motivation stems from the fact that dictionaries inherently contain redundant dictionary atoms. This is because the aim of dictionary learning is reconstruction, not classi�cation. In this thesis, we propose methods to obtain minimum number discriminative dictionary atoms for e�ective classi�cation and also reduced computational time. First, we propose a classi�cation scheme where an example is assigned to a class based on the weight assigned to both maximum projection and minimum reconstruction error. Here, the input data is learned by K-SVD dictionary learning which alternates between sparse coding and dictionary update. For sparse coding, orthogonal matching pursuit (OMP) is used and for dictionary update, singular value decomposition is used. This way of classi�cation though e�ective, still there is a scope to improve dictionary learning by removing redundant atoms because our goal is not reconstruction. In order to remove such redundant atoms, we propose two approaches based on information theory to obtain compact discriminative dictionaries. In the �rst approach, we remove redundant atoms from the dictionary while maintaining discriminative information. Speci�cally, we propose a constraint optimization problem which minimizes the mutual information between optimized dictionary and initial dictionary while maximizing mutual information between class labels and optimized dictionary. This helps to determine information loss between before and after the dictionary optimization. To compute information loss, we use Jensen-Shannon diver- gence with adaptive weights to compare class distributions of each dictionary atom. The advantage of Jensen-Shannon divergence is its computational e�ciency rather than calculating information loss from mutual information.

[error in script]

IITH Creators: