Strategic Base Representation Learning via Feature Augmentations for Few-Shot Class Incremental Learning

Parinita Nema, Vinod K. Kurmi

Abstract

Few-shot class incremental learning (FSCIL) requires models to learn new classes from limited data while retaining knowledge of base classes. A common challenge is the overlap between old and new classes in the feature space, leading to performance degradation. We propose Feature Augmentation driven Contrastive Learning (FACL), a framework that proactively enhances class separation during base training. By utilizing feature augmentations and proxy labels, FACL expands the feature space to accommodate future classes. Combined with a self-supervised contrastive loss, our method achieves state-of-the-art performance on CIFAR100, miniImageNet, and CUB200 benchmarks.

Introduction

The field of Few-Shot Class Incremental Learning (FSCIL) addresses the challenge of teaching AI models new tasks sequentially using very few training instances (e.g., 5-shot learning) without forgetting previously acquired knowledge. Standard deep learning models typically suffer from two primary issues in this setting: catastrophic forgetting, where new weight updates overwrite old knowledge, and overfitting, where the model fails to generalize from the sparse new data.

While many existing frameworks attempt to mitigate forgetting by freezing the model's backbone after the initial "base" session, this often results in a suboptimal feature space. If the base classes are not sufficiently separated, new classes will inevitably overlap with them, causing classification errors.

Conceptual motivation: FACL improves class separation in the base session to prevent overlap with future incremental classes.

Methodology: Feature Augmentation driven Contrastive Learning (FACL)

The core philosophy of FACL is "Strategic Base Representation Learning." Instead of merely learning to classify base categories, the model is trained to organize its internal representation space to be "future-ready." The framework introduces three key components:

1. Feature Augmentation (FA)

Rather than augmenting raw images alone, FACL operates on the feature level. It extracts features from two differently augmented versions of an image and blends them using a mixing coefficient $\delta$. This process generates "mixture" features that occupy regions of the feature space not strictly tied to any existing base class.

2. Proxy Label Space

These blended features are assigned "proxy labels." These labels act as placeholders for potential future classes. By treating these synthetic mixtures as distinct categories during the base session, the model is forced to reserve "gaps" in its feature space, effectively pre-allocating room for real novel classes that will appear in later sessions.

3. Self-Supervised Contrastive Loss (SSCL)

To further refine the feature space, FACL employs a self-supervised contrastive loss (based on the MoCo architecture). This loss pulls features of the same class (and its augmentations) together while pushing features from different classes—including the proxy classes—apart. This ensures high intra-class compactness and inter-class separation.

The FACL pipeline illustrating the flow from data stream through feature augmentation to the contrastive loss and integrated classifier.

Experimental Results

The FACL framework was evaluated on three standard FSCIL benchmarks: CIFAR100 (60 base, 40 incremental classes), miniImageNet (60 base, 40 incremental classes), and CUB200 (100 base, 100 incremental classes).

FACL consistently outperformed existing state-of-the-art methods:

  • CIFAR100: Achieved a 4.77% improvement in average accuracy and a 2.87% boost in final session accuracy.
  • CUB200: Demonstrated a 1.59% average accuracy increase, with final session performance rising by 2.2%.
  • miniImageNet: Showed a 0.6% average improvement.

Ablation studies highlighted that the combination of contrastive loss and proxy class generation is critical. Specifically, the model reached peak performance when the mixing coefficient $\delta$ was set to 0.5, indicating that a balanced mixture of features provides the most effective "placeholder" for future learning.

Discussion and Conclusion

The success of FACL underscores the importance of anticipatory learning. By simulating the presence of novel classes during the initial training phase, the model develops more robust and transferable features. Visualizations (such as t-SNE plots) confirm that FACL produces much tighter and better-separated clusters compared to standard cross-entropy training.

Despite these gains, challenges remain. The "Performance Drop" (PD) metric for FACL is relatively high because it starts from a very high base accuracy; this suggests a need for more nuanced evaluation metrics in incremental learning. Future research could explore "soft-freezing" techniques for the backbone or the use of more complex, domain-specific augmentations to further enhance the model's flexibility.

In summary, FACL provides a powerful strategy for FSCIL by proactively managing the feature space, setting a new benchmark for performance in few-shot incremental scenarios.

Summary updated

BibTeX
@inproceedings{nema2025strategic,
  author = {Parinita Nema and Vinod K. Kurmi},
  title = {{Strategic Base Representation Learning via Feature Augmentations for Few-Shot Class Incremental Learning}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
  doi = {10.1109/wacv61041.2025.00623},
}
Figure