Whether context identity must be inferred in case of domain increment？

[youtao1]

In your paper "Three types of incremental learning" domain increments do not require contextual identity to be inferred. But in the paper you cited "[17] An Efficient Domain-Incremental Learning Approach to Drive in All Weather Conditions" it is mentioned that "A requirement of DISC and the proposed domain-IL scenario is to have access to the task-ID during inference , as it is done in task-incremental learning approaches”. Is it because the task_ID can be obtained through the car sensor, so there is no need to infer the contextual identity?

[Gido van de Ven]

Thanks for your interest in this work! I can see the confusion, let me try to clarify. The problem that is described in ref [17], learning to drive in different weather conditions, is a domain-incremental learning problem because, in general, solving this problem does not require identifying the weather condition. However, as for any domain-incremental learning problem, one possible solution to this problem is through first identifying the domain (in this case the weather condition), which then allows the use of task- or domain-specific components. This is the kind of solution that is taken by DISC. So the problem that is described in [17] does not require inference of task-ID, but the solution they propose does. Hope this helps!