In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy
will defend his dissertation
Conceptual Domain Adaptation Using Deep Learning
Domain adaptation scenarios have been successively gaining attention in practical applications of machine learning. Here, the source distribution in which the classifier is trained, differs from the target distribution in which the classifier will ultimately be applied. The discrepancy between source and target distributions often results in poor predictive performance.
Recent work shows several approaches providing a solution to alleviate the problem of distribution discrepancy. Deep neural networks extract high-level abstract representations of data, and have been used to transform the source and target data into a new common space, such that the aforementioned discrepancy is minimized.
In this document we introduce conceptual domain adaptation where semantic information hidden in high-level concepts --as opposed to information from low-level representational properties of the data-- is directly used for adaptation. We investigate current deep-learning-based domain adaptation approaches and argue that due to their reliance on representational properties of data, using them for domain adaptation is prone to failure under certain scenarios. These scenarios are investigated as cases with inherently lower-level representational discrepancy.
In this project, we introduce an adjustment approach as a solution towards conceptual domain adaptation. Accordingly, we propose a search framework to adjust high-level representation of target data along with basic supervised, graph-based and PCA-based fitness evaluation. Based on experimental results, we contend that the proposed solution is beneficial for domain adaptation problems with lower-level representational in supervised scenarios.
Date: Wednesday, July 19, 2017
Time: 10:00 AM
Place: PGH 550
Advisor: Dr. Ricardo Vilalta
Faculty, students, and the general public are invited.