SAMix: Calibrated and Accurate Continual Learning via Sphere-Adaptive Mixup and Neural Collapse

TL;DR

This work tackles calibration in continual learning by addressing the prediction reliability gap in Neural Collapse (NC)-based methods. It introduces Sphere-Adaptive Mixup (SAMix), which adaptively mixes inputs and fixed class prototypes on the unit sphere using Slerp, and integrates this into a two-stage plasticity/stability framework with fixed Equiangular Tight Frame (ETF) prototypes. By combining SAMix with Dot-Regression (DR) or FNC$^2$ losses, the approach improves across-task accuracy while yielding better calibration and reduced overconfidence, including favorable results without memory. The authors also propose CL-specific calibration metrics AECE and AOE and provide theoretical arguments for preferring DR over contrastive losses with mixed samples. Empirically, SAMix achieves state-of-the-art performance on Seq-Cifar-100 and Seq-Tiny-ImageNet across Class-IL and Task-IL settings, with notable calibration benefits, making neural-collapse-based continual learning more reliable for real-world deployments.

Abstract

While most continual learning methods focus on mitigating forgetting and improving accuracy, they often overlook the critical aspect of network calibration, despite its importance. Neural collapse, a phenomenon where last-layer features collapse to their class means, has demonstrated advantages in continual learning by reducing feature-classifier misalignment. Few works aim to improve the calibration of continual models for more reliable predictions. Our work goes a step further by proposing a novel method that not only enhances calibration but also improves performance by reducing overconfidence, mitigating forgetting, and increasing accuracy. We introduce Sphere-Adaptive Mixup (SAMix), an adaptive mixup strategy tailored for neural collapse-based methods. SAMix adapts the mixing process to the geometric properties of feature spaces under neural collapse, ensuring more robust regularization and alignment. Experiments show that SAMix significantly boosts performance, surpassing SOTA methods in continual learning while also improving model calibration. SAMix enhances both across-task accuracy and the broader reliability of predictions, making it a promising advancement for robust continual learning systems.

Figures (3)

• Figure 1: Illustration of our proposed method.
• Figure 2: Linear interpolation and Slerp.
• Figure 3: Results on Seq-Cifar100 with the Class-IL setting.