Candidate: Benyamin Ghojogh
Title: Data reduction for machine learning
Date: June 22, 2020
Time: 11:00 AM
Place: REMOTE PARTICIPATION
Supervisor(s): Crowley, Mark - Karray, Fakhri
Abstract:
This thesis is on data reduction for machine learning. Data reduction can be categorized into two main categories which are prototype selection and dimensionality reduction. In this thesis, I propose different algorithms in machine learning for data reduction. In prototype selection, different methods are proposed which are principal sample analysis, numerosity reduction using matrix decomposition, inverse curvature anomaly detection, and isolation Mondrian forest. In the dimensionality reduction section, I propose Structural Similarity Index (SSIM) kernel, image structural component analysis, locally linear image structural embedding, Roweis discriminant analysis, quantile-quantile embedding, quantized Fisher discriminant analysis (FDA), weighted FDA, backprojection for training neural nets, FDA loss for training Siamese nets, and embedding histopathology images using Siamese networks. The experiments show the effectiveness of the proposed methods for data reduction.