CIFNet: Analytic Framework for Class-Incremental Learning
Researchers have introduced CIFNet, a neural learning framework designed for Class-Incremental Learning (CIL) that substitutes traditional iterative gradient-based optimization with deterministic, closed-form adaptations for classifiers. Utilizing frozen pre-trained representations, CIFNet modifies classifier weights through Regularised Recursive Least-Squares (RRLS), which serves as an exact ridge-regression method within a stable embedding space. This approach eliminates the need for backpropagation and iterative convergence, thereby lowering computational demands and sensitivity to hyperparameters. To mitigate structural initialization bias caused by new output neurons that lack evidence from previous classes, CIFNet features a lightweight calibration buffer in latent space with density-aware modifications. The framework is detailed in arXiv:2509.11285.
Key facts
- CIFNet is an analytic neural learning framework for Class-Incremental Learning (CIL).
- It uses frozen pre-trained representations and closed-form classifier adaptations.
- No backpropagation or iterative convergence is required.
- Classifier weights are updated via Regularised Recursive Least-Squares (RRLS).
- A calibration buffer in latent space counters structural initialisation bias.
- The method reduces computational cost and hyperparameter sensitivity.
- The paper is available on arXiv with ID 2509.11285.
- The announcement type is replace-cross.
Entities
Institutions
- arXiv