Free-Flow Class-Incremental Learning: Towards Robust CIL under Variable Class Arrivals
The episode discusses a paper addressing limitations in Class-Incremental Learning (CIL), where models learn new classes while retaining old ones. The hosts explain that current testing methods fail because real-world class arrivals are irregular, or 'free-flow.' The authors propose solutions, such as the Class-Wise Mean objective, to make CIL robust against this messy, unpredictable flow.