Introduction
Generalized Few-Shot Learning (GFSL) presents a significant challenge: training a model to recognize novel classes from minimal examples while maintaining performance on previously learned "base" classes. Conventional approaches rely on access to base data to prevent catastrophic forgetting. However, real-world constraints—such as data privacy in healthcare or storage limitations—often make base data unavailable. This paper introduces Zero-Base GFSL, a scenario where models must adapt to new classes without any access to the original training set.
The Challenge: Mean Shifting and Boundary Collapse
The primary obstacle in Zero-Base GFSL is the inherent bias introduced during fine-tuning. When a model is updated only on novel samples, the decision boundaries for the base classes tend to collapse. The authors identify the "mean shifting phenomenon" as the root cause: novel class weights develop a significantly higher mean than base class weights because they receive frequent positive updates without the counterbalancing negative updates from base data.
As shown in the figure above, the introduction of a novel class (red triangles) without proper normalization causes its decision region to aggressively encroach upon the territories of the base classes (stars, squares, and circles), leading to poor generalization.
The MVCN Framework
To address this bias, the researchers propose Mean-Variance Classifier Normalization (MVCN). This framework consists of three key components designed to rebalance the classifier weights without requiring base images:
- Online Mean Centering: During fine-tuning, novel class weights are normalized to maintain a zero mean. This prevents them from "bullying" base weights during the softmax calculation.
- Offline Variance Balancing: After training, the standard deviation of base class weights is adjusted to match that of the novel weights, ensuring structural parity between the two groups.
- Post-Linear Optimization: A final refinement step uses two learnable parameters per class to optimize decision boundaries further.
By applying MVCN, the decision boundaries are restored to a more equitable state, as illustrated in the figure above, allowing the model to distinguish between all classes effectively.
Experimental Results
The effectiveness of MVCN was validated across standard benchmarks including mini-ImageNet, tiered-ImageNet, and CUB. On mini-ImageNet (5-shot), MVCN achieved an overall accuracy of 71.22%, significantly outperforming previous state-of-the-art methods like XtarNet (66.52%), even though XtarNet utilizes base data.
Ablation studies highlighted the importance of mean centering, which alone improved base class preservation by over 25%. These results suggest that the bias identified by the authors is a more fundamental hurdle in GFSL than previously recognized, and that sophisticated normalization can be more effective than simple data rehearsal.
Conclusion
MVCN provides a robust solution for Generalized Few-Shot Learning in data-constrained environments. By mathematically addressing the mean and variance discrepancies in classifier weights, the method achieves superior performance without the need for base data. Future work will explore extending this analysis to non-linear classifiers and non-frozen feature extractors.
Summary updated