美DARPA“终身学习机器”项目研究进展综述

Progress and Trend of DARPA's Lifelong Learning Machine Project:A Survey

“终身学习机器”(L2M)项目隶属于美国国防预先研究计划局(DARPA),其目标是开发支持下一代自适应人工智能系统所需的技术,使其能够在实际环境中持续地学习并改善性能,同时保持原来预先给定的能力。对L2M项目基本情况和研究进展进行了综述。首先,介绍了L2M项目的背景、目标、研究内容和项目阶段;然后,从终身学习理论方法研究、边缘终身学习与终身学习机器的硬件实现和终身学习机器人三个方面,对L2M项目的研究进展进行介绍和分析评述,其中理论方法方面包括不确定性调节的终身学习及其应用、自动驾驶中的终身学习、本征任务终身学习框架、通过元学习的神经调节克服灾难遗忘等;最后,对未来发展趋势进行了展望。综述表明,终身机器学习亟需继续开展研究,尤其是生物机制启发的终身学习、多智能体协同终身学习和终身学习技术在现实世界复杂场景中的应用等方向。

The Lifelong Learning Machine(L2M)program is an important project supported by the Defense Advanced Research Program Agency(DARPA)of USA.The objective of the L2M program is to develop a next generation adaptive artificial intelligence(AI)system capable of continually learning and improving performance in real-world environments,while without forgetting the previously learned knowledge and abilities.This review summarizes the basic information and research progress of the L2M project.First,the background,objectives,research content and project stage of the L2M project are introduced.And then the research results of the project,such as theory and methods,lifelong learning at the edge&L2M hardware implementations,and lifelong learning robotics,are reviewed and analyzed.With respect to theory and methods,uncertainty modulated lifelong learning and application,lifelong learning in autonomous driving,the eigentask framework,and meta-learned neuromodulation are included.Finally,the development trends and insights for future are provided.This survey shows that research on lifelong machine learning is urgent to continue.Important future directions include biological mechanisms inspired lifelong learning,shared experience lifelong learning with a population of agents,and the application to reality world and complex scenarios of lifelong machine learning.