Lifelong Machine Learning and Sequential Decision Making for Natural Language Interfaces

点击翻译按钮获取中文摘要

What could be possible if we let a machine learn continuously over its lifetime? Objectives: The goal of this work is to develop lifelong learning algorithms with application to dialog management. At a theoretical level, this research will investigate and advance the principles by which a machine can gradually learn over a long period of time, discover new concepts and generalize concepts to new situations. This will be put in practice by developing open ended dialog systems for natural language interfaces. Personalization of such interfaces at the language level and adapting to the habits and preferences of the user will benefit tremendously from continuous learning. Methods: Since continuous learning is a sequential process, the techniques developed will be in the framework of reinforcement learning. Particular emphasis will be put on the development of non-parametric techniques that do not make a closed world assumption by allowing new concepts (e.g., new words, expressions, habits, goals) to be discovered and represented on the fly. Hierarchical representations will be employed to organize and reason about the concepts from low level language units to high level user intentions. The techniques will be deployed in speech and text interfaces for smart phone applications. Novelty and significance: Lifelong machine learning is a new research direction that remains vastly unexplored. This research will develop the theory and practice of this new paradigm, which will enable a new breed of intelligent systems. It will contribute to the next generation of natural user interfaces based on speech for smart phones, video-gaming and hands-free car consoles.