If you're interested in understanding the logic behind how to detect novelty from determinism to PAC-Bayes Learning, we have a preprint at arxiv.org/abs/2501.09331. The reasoning is grounded in computation and information theory.

The paper is at doi.org/10.1613/jair.... the benchmarks starts with Kinetcs400, then splits Kinetics600 and 700 into 5 increments each, where new activities are introduced and to be detected & learned. We limit the feedback to see which samples the predictor prioritizes for ground truth labels.

Can your AI learn novelty? I'm presenting at @ijcai.org an incremental online learning benchmark where novel classes are introduced in each increment with prior known classes. Come talk with me at #IJCAI2025 about predictor detection & adaptation to novel phenomena! 😁