From One-off Machine Learning to Perpetual Learning: A STEP Perspective

While classical machine learning highlights on improving the performance P of a computing system for tasks T with experience E, perpetual (or lifelong) learning emphasizes on the accomplishment of such performance improvement incrementally and continuously through an open sequence of learning episodes. Thus, in addition to the aforementioned T, E, and P, perpetual learning also needs a set S of learning stimuli so as to initiate the learning episodes. We call this the STEP perpetual learning. In this paper, our focus is on elucidating the basic ideas of the STEP perpetual learning. Through two use cases, we explain how the approach can be utilized to develop real applications. The STEP perpetual learning draws its inspiration from a broad range of research fields. Finally, the take-home message is: machine learning, just like human learning, is and should be a life-long endeavor.