MS defense: Feature Extraction using a Hierarchical Growing Neural Gas

M.S. Thesis Defense Announcement

Feature Extraction using a Hierarchical Growing Neural Gas

Roger Guseman

12:00pm 25 April 2011, ITE 210

Unsupervised, data-driven, automatic feature extraction from image data is an interesting and difficult problem. High dimensional data, such as images, often contain less information than they do data. For an agent to better reason about this data, finding the "interesting" features in the data is helpful. A current technique, known as the Growing Neural Gas (GNG), is a neural network approach to feature extraction. There are, however, adaptations that can be made to the Growing Neural Gas in order to increase its performance.

Contributions of this work include development of a new neural network algorithm extending the Growing Neural Gas framework, known as the Hierarchical Growing Neural Gas (HGNG), identification of how the parameters of the HGNG affect feature extraction performance, and theoretical and empirical comparisons of performance between the normal GNG and the HGNG neural networks.

Thesis Committee:

• Dr. Tim Oates (chair)
• Dr. Tim Finin
• Dr. Marie desJardins