Learning in the Open World: Techniques for Identifying and Adapting to the Unknown

Abstract

Traditional machine learning assumed a closed world scenarios with massive well-built offline training data and a similar data distribution for evaluation afterwards. While this assumption could be valid in numerous scenarios, it is frequently disregarded in open environments, where there could always be "unknown" happened. Open-world learning could be defined as learning a model that could not only perform the intended task and also identify new things that have not been learned before and then incrementally learn the new things.This thesis investigates several open-world scenarios, each with varying degree of information or knowledge provided. The first scenario involves the appearance of an ``unknown'' category during the testing phase of supervised classification, while the second centers on detecting "unknown" in multi-task learning and there will be the ``unknown'' task that crops up during the testing. The third scenario examines the Reinforcement learning setting where the environment is ``unknown'' at first and strategy are designed for better exploration. Finally, the thesis delves into the most extreme case where there is no additional knowledge about the data. A better representation learning algorithm is proposed from the perspective total correlation estimation. Extensive experiments are conducted to demonstrate the effectiveness of the proposed algorithm under each scenario.