That is the place continuous studying (CL) jumps in. In CL, information arrives incrementally over time, and can’t be (absolutely) saved. The machine studying mannequin is educated solely on the brand new information; the problem right here is catastrophic forgetting: efficiency on previous information drops. The rationale for the efficiency drop is that the mannequin adapts its weights to the present information solely, as there isn’t a incentive to retain data gained from earlier information.
To fight forgetting and retain previous data, many strategies have been proposed. These strategies could be grouped into three central classes*: rehearsal-based, regularization-based, and architecture-based. Within the following sections, I’ll element every class and introduce choose papers to discover additional. Whereas I give attention to classification issues, all coated concepts are largely equally legitimate for, e.g…