New Method Boosts Incremental Learning for Long-Tailed Data

Researchers have developed a new framework called Geometry-Spectral Rectification (GSR) to overcome significant challenges in Analytic Class-Incremental Learning (ACL), particularly when dealing with long-tailed data distributions. Existing ACL methods, while computationally efficient, struggle with class imbalance, where "tail" classes (those with few samples) suffer from severe spectral collapse, making them numerically indistinguishable from noise. GSR addresses this by treating long-tailed learning as a spectral regularization problem. Unlike standard Ridge Regression, which applies uniform regularization, GSR functions as an anisotropic spectral filter. It selectively inflates the collapsed eigenvalues of tail classes by constructing a structured, data-dependent spectral perturbation matrix. Theoretical analysis confirms that GSR guarantees an improved stable rank for the Gram matrix, ensuring numerical stability. Extensive experiments demonstrate that GSR achieves state-of-the-art results for analytic Class-Incremental Learning, offering a superior balance between computational efficiency and robust generalization in long-tailed settings.