Introduction
Few-shot learning (FSL) aims to bridge the gap between machine and human intelligence by enabling models to recognize novel categories from just a handful of labeled examples. Traditionally, the field has been dominated by meta-learning strategies, specifically episodic training, which attempts to mimic the few-shot task structure during the training phase. However, this paper challenges that paradigm, suggesting that the complexity of meta-learning may be unnecessary and that the rich information contained within "base classes"—the large-scale datasets used for initial training—is significantly underutilized.
Rethinking the Baseline: Baseline++
The authors first propose Baseline++, a robust supervised learning framework that demonstrates how a well-tuned standard model can outperform complex meta-learning algorithms. By training a ResNet-12 backbone using a global cross-entropy loss on the entire base dataset rather than episodic tasks, they create a more powerful feature extractor. This baseline is enhanced by three key technical refinements:
- Advanced Data Augmentation: Incorporating horizontal flips, random crops, and random erasing to improve generalization.
- Softmax Temperature Scaling: Adjusting the smoothness of the probability distribution to refine the decision boundaries.
- Dense Classification: Instead of relying on global average pooling, the model classifies every local feature vector in the final feature map, providing a much denser supervisory signal during training.
The Cooperative Bi-path Metric (CBM)
Building upon this strong baseline, the authors introduce the Cooperative Bi-path Metric (CBM). Most metric-based FSL methods rely solely on an "Inductive Path," which directly compares the cosine similarity between query images and novel support samples. CBM introduces a complementary Transductive Path that uses base classes as a shared coordinate system.
In this transductive approach, both query and support samples are mapped to a "similarity distribution" against all known base classes. If a query image and a support image exhibit similar patterns of proximity to these "landmark" base classes (e.g., both are highly similar to 'dog' and 'wolf' but distant from 'car'), they are likely to belong to the same novel category. To ensure precision, the authors apply Locally Linear Embedding (LLE), which restricts the contribution to the $k$-nearest neighbor base classes, effectively filtering out noise from irrelevant categories.
Experimental Results
The proposed method was evaluated on standard benchmarks, including miniImageNet and tieredImageNet. The results demonstrated that the non-parametric CBM could be seamlessly appended to any pre-trained backbone to achieve state-of-the-art (SOTA) performance without additional training.
On tieredImageNet, Baseline++ achieved a 1-shot accuracy of 71.27%, significantly surpassing previous leaders like CAN (69.89%). On miniImageNet, the full CBM approach reached 64.77% for 1-shot tasks, outperforming complex models such as TADAM and RN. These results confirm that a strong feature extractor combined with an intelligent use of base class relationships is a highly effective strategy for few-shot recognition.
Conclusion
This research provides two vital insights for the FSL community: first, that global supervised learning can be superior to episodic meta-learning for feature extraction; and second, that base classes provide a valuable "proxy" space for comparing novel samples. While CBM currently uses a handcrafted metric, future work may explore replacing the transductive path with learnable components to further increase flexibility and performance across diverse visual domains.
Summary updated