Metrics
1 Record Views
Conference proceeding
Adaptive Additive Parameter Updates of Vision Transformers for Few-Shot Continual Learning
Conference record - Asilomar Conference on Signals, Systems, & Computers, pp.680-686
Asilomar Conference on Signals, Systems, and Computers, 59th (Pacific Grove, California, USA, 10/26/2025–10/29/2025)
10/26/2025
Abstract
Integrating new class information without losing previously acquired knowledge remains a central challenge in artificial intelligence, often known as catastrophic forgetting. Few-shot class incremental learning (FSCIL) addresses this by first training a model on a robust set of base classes and then incrementally adapting it in successive sessions using few labeled examples per novel class. However, this approach is prone to overfitting on limited new data, which can compromise performance and exacerbate forgetting. In this work, we propose a simple yet effective FSCIL framework that leverages a frozen Vision Transformer (ViT) backbone augmented with parameter-efficient additive updates. Our approach freezes the pre-trained ViT parameters and selectively injects trainable weights into the self-attention modules via an additive update mechanism. This design updates a small subset of parameters to accommodate new classes without sacrificing representations learned during the base session. By fine-tuning only a few parameters, our method preserves the generalizable features in the frozen ViT while reducing overfitting risk. Furthermore, as most parameters remain fixed, the model avoids overwriting knowledge when novel data batches are introduced. Extensive experiments on benchmark datasets demonstrate that our approach yields state-of-the-art performance compared to baseline FSCIL methods. Our results confirm improvements in robustness and accuracy.
Related links
Details
- Title
- Adaptive Additive Parameter Updates of Vision Transformers for Few-Shot Continual Learning
- Publication Details
- Conference record - Asilomar Conference on Signals, Systems, & Computers, pp.680-686
- Resource Type
- Conference proceeding
- Conference
- Asilomar Conference on Signals, Systems, and Computers, 59th (Pacific Grove, California, USA, 10/26/2025–10/29/2025)
- Publisher
- IEEE
- Copyright
- © 2025, IEEE
- Identifiers
- 99381734858106600
- Academic Unit
- Institute for Human and Machine Cognition; Center for Cybersecurity and AI; Intelligent Systems and Robotics; Division of Academic Affairs; Hal Marcus College of Science and Engineering
- Language
- English