Few-Shot Learning for Multi-New-Class Scenarios: Challenges, Methodology, and Experimental Evaluation

Small‑sample (few‑shot) learning aims to mitigate the large prediction errors caused by insufficient training data, a common issue in real‑world applications. While most research focuses on few‑new‑class settings, the multi‑new‑class scenario—where many unseen categories appear with only a few samples each—has received little attention.

The multi‑new‑class setting presents three main challenges: existing benchmarks are designed for few‑new‑class tasks; many few‑shot methods incur prohibitive computational costs when the number of new classes grows; and the “transfer collapse” problem degrades performance when transferring knowledge to many novel categories.

To address these challenges, the authors construct a new dataset (ImageNet‑MNC) that satisfies the multi‑new‑class criteria, and propose a parallel training framework that distributes the workload across multiple GPUs. Their SHA‑Pipeline method consists of three components: (1) feature regularization using Z‑score normalization to reduce communication overhead, (2) a non‑parametric hierarchical clustering step to capture class relationships, and (3) representation learning that combines hierarchical distance, Euclidean distance, and a custom loss function.

Extensive experiments compare SHA‑Pipeline with ProtoNet, ProtoNet‑Fix, SimpleShot, and P> M>F on both the multi‑new‑class dataset and standard few‑shot benchmarks (CIFAR‑FS, miniImageNet, and meta‑ImageNet). Results show that SHA‑Pipeline consistently achieves the highest accuracy while maintaining reasonable computational cost. Ablation studies confirm the importance of each component, and visualizations illustrate the improved feature clustering after processing.

The work concludes by highlighting three contributions: extending few‑shot learning to open‑world, multi‑new‑class scenarios; introducing an efficient distributed training strategy; and leveraging hierarchical class semantics for fine‑tuning. Future directions include further optimizing efficiency and generalization, enhancing online adaptation, improving interpretability, and tackling multimodal and multitask learning.