Towards Advanced Transfer Learning in Complex Environments

To date, due to the fast development of digital technologies, huge volumes of data with complex properties which come from various sources, such as mobile devices and internet applications, are involved in many machine learning (ML) tasks. As problems seldom exist in isolation, transfer learning (TL) has been proposed in the literature as a paradigm to leverage the rich labeled data from the source domain to improve the learning performance in the target domain where only few (or no) labeled data are available. TL is a revolutionary breakthrough that makes it possible to reuse prior work and democratize ML models (https://tdwi.org/articles/2018/08/24/adv-all-secretbehind- new-ai-spring-transfer-learning.aspx). Recently, TL has been used as one of the core techniques in Google’s Cloud AutoML which was announced in January 2018 (http://www.fast.ai/2018/07/23/auto-ml-3/). However, despite the great success TL has enjoyed so far, there is a lack of TL models designed to pursue the positive transfer performance in today’s applications with data in complex environments, such as intelligent robot training and logistic systems, where the data available for training are often noisy, imbalanced, and received incrementally. Keeping the above in mind, we aim to design and develop a new TL framework for resolving real-world applications featuring noisy, imbalanced, and incremental data, where most existing TL methods may fail to enable positive knowledge transfer and thus cannot achieve the desirable performance. In particular, by leveraging techniques such as importance sampling and Bayesian estimation, which have been successfully applied in deep learning, imbalanced learning and online learning, the research proposed in this project will be carried out in four different aspects including 1) Develop a noise module for addressing noisy data in transfer learning; 2) Develop an imbalance module for addressing the issue of imbalanced label data; 3) Develop an online module for handling incremental data; 4) Apply the developed TL framework to real-world applications, i.e., optimal motion generation for robots in complex environments and fast route planning in logistic systems.