Challenges in using existing domain adaptive techniques for real-world settings
Author(s)
Kartik, Deeksha
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Recent years have seen major progress in machine learning and computer vision research. However, the successes seen so far are largely driven by supervised models with the help of large-scaled annotated data relying on the i.i.d. assumption. In reality, this assumption may be violated as the training and test data might come from different distributions. This could have serious repercussions for real-world, large-scale deployments of machine learning models. Domain Adaptation (DA) is a widely recognized paradigm to combat this. DA is a sub-field of machine learning that is explicitly concerned with designing algorithms to account for the distributional differences between sets of data. It involves leveraging knowledge from an abundantly labeled source domain and transferring it to a target domain with limited or no labels.
Most DA methods in the literature are evaluated on smaller toy datasets and do not accurately capture the challenges faced in the real-world. In this thesis, we examine the applicability of existing DA techniques in the wild. To this end, we deep dive into three unique real-world scenarios. We propose tailored solutions for DA faced with the challenge of simultaneous data and label distribution shift and DA in the challenging source-free setting. We also analyze the sources of variation in pathological data and provide an exploration into tackling domain shift for pathology.
Sponsor
Date
2022-05-03
Extent
Resource Type
Text
Resource Subtype
Thesis