Scalable Class-Incremental Learning Based on Parametric Neural Collapse

TL;DR

The paper addresses scalable class-incremental learning by reducing structural growth and mitigating feature drift during backbone expansion. It introduces SCL-PNC, a framework that combines an adapt-layer, a dynamic parametric ETF classifier, and parallel expansion with knowledge distillation to preserve feature geometry across modules and align evolving class distributions. Grounded in neural collapse, the method pushes backbone representations toward ETF vertices, enabling efficient long-term incremental growth. Empirical results on CIFAR-100 and ImageNet-100 demonstrate superior average incremental accuracy and favorable efficiency relative to state-of-the-art dynamic expansion approaches.

Abstract

Incremental learning often encounter challenges such as overfitting to new data and catastrophic forgetting of old data. Existing methods can effectively extend the model for new tasks while freezing the parameters of the old model, but ignore the necessity of structural efficiency to lead to the feature difference between modules and the class misalignment due to evolving class distributions. To address these issues, we propose scalable class-incremental learning based on parametric neural collapse (SCL-PNC) that enables demand-driven, minimal-cost backbone expansion by adapt-layer and refines the static into a dynamic parametric Equiangular Tight Frame (ETF) framework according to incremental class. This method can efficiently handle the model expansion question with the increasing number of categories in real-world scenarios. Additionally, to counteract feature drift in serial expansion models, the parallel expansion framework is presented with a knowledge distillation algorithm to align features across expansion modules. Therefore, SCL-PNC can not only design a dynamic and extensible ETF classifier to address class misalignment due to evolving class distributions, but also ensure feature consistency by an adapt-layer with knowledge distillation between extended modules. By leveraging neural collapse, SCL-PNC induces the convergence of the incremental expansion model through a structured combination of the expandable backbone, adapt-layer, and the parametric ETF classifier. Experiments on standard benchmarks demonstrate the effectiveness and efficiency of our proposed method. Our code is available at https://github.com/zhangchuangxin71-cyber/dynamic_ ETF2. Keywords: Class incremental learning; Catastrophic forgetting; Neural collapse;Knowledge distillation; Expanded model.

Figures (9)

• Figure 1: The two issues of existing model-expansion methods. (a) illustrates the inter-module feature drift that arises during module alignment with a substantial discrepancy between category features A and B. Consequently, the knowledge transferred across modules becomes discontinuous ultimately leading to catastrophic forgetting of old classes. (b) shows the misalignment due to evolving class distributions. As the number of categories increases while the ETF classifier maintains fixed parameters, a pronounced distribution bias emerges.
• Figure 2: Scalable class-incremental learning based on parametric neural collapse architecture
• Figure 3: The structure of the base-layer and the expand-layer
• Figure 4: Average Accuracy of the Class Incremental Learning Methods Under Different Experimental Strategies on the CIFAR-100 Dataset
• Figure 5: Task-stage drift on CIFAR-100 feature space via t-SNE. The visualization tracks feature space evolution from task $0 \to 1$ (a) to task $4 \to 5$ (e). Gray points show old-class features before the task; colored triangles show new-class features after the task. Colored arrows denote the minimal drift trajectory of old-class centroids, confirming representational stability.