Open Access Thesis
Date of Award
High dimensionality and the sheer size of unlabeled data available today demand
new development in unsupervised learning of sparse representation. Despite of recent
advances in representation learning, most of the current methods are limited when
dealing with large scale unlabeled data. In this study, we propose a new unsupervised
method that is able to learn sparse representation from unlabeled data efficiently. We
derive a closed-form solution based on the sequential minimal optimization (SMO)
for training an auto encoder-decoder module, which efficiently extracts sparse and
compact features from any data set with various size. The inference process in the
proposed learning algorithm does not require any expensive Hessian computation
for solving the underlying optimization problems. Decomposition of the non-convex
optimization problem in our model enables us to solve each sub-problems analytically.
Using several image datasets including CIFAR-10, CALTECH-101 and AR
face database, we demonstrate the effectiveness in terms of computation time and
classification accuracy. Proposed method discovers dictionaries that are able to capture
low level features in larger dimensional patches in quite lower executional time
than the other alternatives. Then by detailed experimental results, we present that
our module outperforms various similar single layer state-of-the-art methods including
Sparse Filtering and K-Means clustering method.
Mahnaz, Faria, "Effective Auto Encoder For Unsupervised Sparse Representation" (2015). Wayne State University Theses. 427.