Class Incremental Learning Across Diverse Data Paradigms

Abstract

Class Incremental Learning (CIL) is a fundamental machine learning paradigm that enables models to continuously learn new classes over time, while retaining previously acquired knowledge. Despite significant progress, existing CIL approaches often rely on restrictive assumptions, limiting their applicability in real-world scenarios where data availability, supervision levels, and class distributions evolve dynamically. In this thesis, we address these challenges by developing novel CIL frameworks tailored to diverse learning settings, including semi-supervised, long-tail, and few-shot learning, and further introduce a unified approach that generalizes across these paradigms. Wefirst investigate the problem of Generalized Semi-Supervised Class Incremental Learning (GSS-CIL), where newly introduced classes have limited labeled data, while large amounts of unlabeled data may contain a mixture of previously seen, novel, and outlier classes. To address this, we propose the Expert-Suggested Pseudo-Labeling Network (ESPN), which effectively utilizes unlabeled data to enhance learning while maintaining a balanced performance across tasks. We introduce a harmonic mean-based evaluation metric to better assess performance in this challenging setting. Next, we address the Long-Tail CIL problem, where newly introduced classes exhibit highly imbalanced distributions, making it difficult for models to learn underrepresented categories. We propose a two-stage framework that leverages global variance-based classifier alignment along with prototype-based class representations, eliminating the need for explicit re-balancing techniques. Our approach significantly improves incremental adaptation under severe class imbalance. We then focus on Few-Shot CIL (FSCIL), where models must learn new classes with only a few labeled examples per class. To mitigate overfitting and enhance generalization, we introduce the Self-Supervised Stochastic Classifier (S3C), which integrates stochastic classifier weights and self-supervised feature learning to improve learning across incremental steps. Our method achieves state-of-the-art performance in both the standard and more realistic FSCIL settings. Beyond these scenario-specific solutions, we propose Aggregated Self-Supervision (AggSS), a plug-and-play module that enhances feature representations using self-supervised auxiliary tasks such as image transformations. By integrating AggSS into different CIL models, we demonstrate consistent performance gains across multiple CIL protocols, improving the adaptability of incremental learning frameworks. Finally, we introduce a unified CIL framework designed to generalize across diverse learning settings. Under the All-Shot Class Incremental Learning (ASCIL) protocol, which spans highshot to extreme low-shot regimes, we employ distance-based losses and stability constraints on Mahalanobis distances to ensure robust learning across varied data distributions. Our unif ied approach sets a new benchmark for CIL generalization, particularly excelling in low-shot scenarios where conventional methods struggle. In summary, this thesis presents a comprehensive study on incremental learning across varying data paradigms, bridging the gap between isolated frameworks and real-world adaptability. Through extensive experiments, we demonstrate the effectiveness of our proposed methods in enhancing the scalability, generalization, and robustness of CIL models in dynamic learning environments.