Incremental learning algorithm for fast adaptation to frequently changing environments

Abstract

In a constantly changing environment, systems that can learn from continuously incoming inputs without forgetting previously acquired knowledge are very important in terms of efficiency, time, and computational resources. The few-shot class incremental learning (FSCIL) paradigm enables algorithms to absorb new classes with limited data while maintaining the essence of existing classes. This research introduces methodologies that approach such FSCIL algorithms from various perspectives to improve performance. Firstly, we reveal that the overlap between the features of base session classes and the features of incremental session classes is significant. Thus, we design training method to reduce the overlap. Secondly, we highlight the importance of the feature extractor learned from a large number of base session training data and study how it can effectively classify images for both existing trained classes and newly introduced classes. We introduce a balanced supervised relative learning method to show balanced performance for existing trained classes and new classes. Lastly, we approach using the universally recognized Contrastive Language-Image Pre-training (CLIP) model. We present the results obtained when introducing the CLIP model to FSCIL and methodologies to utilize it more effectively. Specifically, we designed to rapidly learn class-specific prompts for each session. All approaches have demonstrated state-of-the-art (SOTA) performance on three benchmark datasets commonly used for FSCIL algorithms: CIFAR100, miniImageNet, and CUB200. Moreover, various ablation studies have been experimentally conducted to confirm the efficacy of each module involved in the design. In conclusion, our research presents methodologies that pave the way for creating more robust learning systems capable of withstanding the continuous flow of information characteristic of the modern data ecosystem.