Machine learning (ML) strives to build mathematical models that uncover and mimic the hidden decision-making process of human minds in a data-driven fashion. Despite its popularity and successes in many industrial and scientific endeavors, a fundamental difference lying between the learning environments of ML models and human-beings hinders the widespread application of ML in real tasks. To wit, we humans live in an evolving world so must keep adjusting and reshaping our mental process to accommodate the world’s dynamics and be open to new concepts. While in ML, the models are prescribed to live in a closed world, where the environment into which the models will be deployed in future are prescribed to be identical as that the models were trained with before. As a result, the more wildly the environment evolves, the more rapidly the ML model performance decays.
How to enable ML to work in an open world, to generalize to a changeable data distribution, and to adapt to newly emerging concepts and decision options? The solutions to this question are of a great importance and interest in expanding the ML application domains, and eventually, paving the path to the development of general artificial intelligence. This talk shall focus on 1) summarizing the state of the art of machine learning in open world and 2) envisioning the opportunities by presenting the research gaps and the great potential for real applications.