International Society of Science and Applied Technologies |
|
Analysis of a Transfer Learning Application Using the Transfer Learning Test Framework | ||||
Author | Karl R. Weiss
|
|||
Co-Author(s) | Taghi M. Khoshgoftaar
|
|||
Abstract | In the field of machine learning, a transfer learning environment is when the training and testing data have different distribution properties, due to the data being drawn from different domains. A transfer learning environment is further characterized by having either very limited or no labeled test data available to build a machine learner. In a transfer learning environment, standard validation techniques (e.g. data splitting, cross-validation, etc.) may not be reliable. In previous research, a transfer learning test framework has been introduced to evaluate the robustness of algorithms to different types of distribution differences. The test framework can aid a machine learning practitioner in the evaluation of algorithms in the absence of reliable validation techniques. The transfer learning test framework stress tests machine learning algorithms in a transfer learning environment. This is accomplished by creating training and the testing data pairs that exhibit many different distribution differences that simulate conditions in a transfer learning environment. In this paper, we apply the transfer learning test framework to the source data of a transfer learning application. We then compare the test framework results to the actual results of the given transfer learning application for multiple transfer learning algorithms.
|
|||
Keywords | Transfer learning, Test framework analysis | |||
Article #: 23-151 |
August 3-5, 2017 - Chicago, Illinois, U.S.A. |