Domain Playground: Extending Deep Learning Models to Open Domain Boundaries
Date
2021Metadata
[+] Show full item recordAbstract
Deep learning models have demonstrated monumental performance in classification tasks but require extensive data and training procedures to converge. Additionally, the performance is only guaranteed when there is no domain gap (e.g., the distribution of the source and target data are similar). Domain adaptation is a specialized classification task that aims to mitigate the domain gap by transferring knowledge between domains, allowing flexible deep learning models for various tasks. However, existing approaches make strong assumptions about the data distribution, often requiring that the source and target domains have no category gap (e.g., they share label sets). In reality, this is rarely the case. There is a strong need for methods that can handle a variety of different domain and category gaps. Universal Domain Adaptation methods have recently arisen, attempting to train models for all domain and category gap situations. However, these still rely on abundant labeled data and extensive training procedures and provide no way to classify target-specific classes without retraining.
In this thesis, we propose a framework called Domain Playground (DP) that aims to extend deep learning models across multiple domains while overcoming the limitations of existing domain adaptation approaches. Based on our previous work, Class Representatives (CRs), the DP framework extracts features from pre-trained models and aggregates them to build a model dynamically in the form of discrete representations. We evaluate the DP framework's domain adaptation capabilities compared to existing approaches and demonstrate compelling performance in domain adaptation tasks. From the domain adaptation evaluation, we have confirmed that DP achieves superior accuracy for classification on the Office-31 dataset. It also shows an ability for transfer learning with diverse datasets and explainability to evaluate classification performance. Finally, we developed a web-based DP application to demonstrate our methods in an interactive web application.
Table of Contents
Introduction -- Transfer Learning -- Domain Adaptation -- Proposed Methods -- Results and Evaluations -- Demo Application -- Conclusion
Degree
M.S. (Master of Science)