Open-World Class Incremental Learning with Adaptive Threshold

Existing class incremental learning (CIL) settings are mainly based on the closed-world assumption. Its training and testing sets only contain in-distribution (ID) samples. However, CIL methods are typically applied in open-world scenarios where out-of-distribution (OOD) samples are widely present, leading to unpredictable behavior of intelligent agents. The OOD categories may also change as the incremental tasks progress. In this paper, we focus on handling a realistic and challenging new setting called Open-World Class Incremental Learning (OWCIL). Specifically, OWCIL includes the accumulated OOD categories during testing to simulate the open-world continual learning scenario. Moreover, it does not provide any OOD samples for model training. Existing methods integrate techniques for CIL and OOD detection to address scenarios similar to the OWCIL. However, they either fail to handle evolving OOD categories or require OOD samples during training, limiting their performance under OWCIL. We propose an adaptive threshold (AT) method to handle the accumulated OOD categories. Additionally, a discriminative optimization method is introduced to determine the threshold without OOD samples. The effectiveness of our proposed method has been validated through extensive experiments on multiple benchmark datasets under OWCIL. Detailed analysis shows that both AT and discriminative optimization can clearly boost performance.