The Divergent Autoencoder (Kurtz, 2007, 2015) combines aspects of generative and discriminative methods by directly applying reconstructive learning toward a classification task. This differs from the approach of learning a task-neutral generative model to provide item representations for a discriminative classifier. The core design principle is the use of separate channels of output nodes each dedicated to autoassociative learning of the members of one class - with a hidden layer shared by all channels. Classification likelihood depends on reconstructive success. I will focus on: (1) formal and informal characterizations of divergent autoencoder behavior and performance; and (2) how the approach is extensible in terms of architecture (e.g., number of hidden layers) and task (e.g., unsupervised learning, regression, co-learning).
Last modified: Wednesday, 22-Mar-2017 11:37:16 NZDT
This page is maintained by the seminar list administrator.