2024/12/12 by Amanda Rios, Ibrahima J. Ndiour, Rios, Amanda +8
Computer Science · Engineering · #Intelligent Tutoring Systems and Adaptive Learning #Advanced Data Processing Techniques #AI-based Problem Solving and Planning
paper · pdf · doi:10.48550/arxiv.2412.09701
AI deployed in many real-world use cases should be capable of adapting to\nnovelties encountered after deployment. Here, we consider a challenging,\nunder-explored and realistic continual adaptation problem: a deployed AI agent\nis continuously provided with unlabeled data that may contain not only unseen\nsamples of known classes but also samples from novel (unknown) classes. In such\na challenging setting, it has only a tiny labeling budget to query the most\ninformative samples to help it continuously learn. We present a comprehensive\nsolution to this complex problem with our model "CUAL" (Continual\nUncertainty-aware Active Learner). CUAL leverages an uncertainty estimation\nalgorithm to prioritize active labeling of ambiguous (uncertain) predicted\nnovel class samples while also simultaneously pseudo-labeling the most certain\npredictions of each class. Evaluations across multiple datasets, ablations,\nsettings and backbones (e.g. ViT foundation model) demonstrate our method's\neffectiveness. We will release our code upon acceptance.\n