Standard few-shot learning (FSL) benchmarks often feel like a toy version of reality. In a typical "5-way" evaluation, a model is asked to distinguish between five unrelated categories—perhaps a "pizza," a "bathtub," and a "panda." Because these classes are semantically distant, the model can rely on coarse, low-level features to succeed.
However, real-world applications—like industrial catalogs, rare disease diagnosis, or large-scale named entity recognition—require models to distinguish between hundreds or thousands of classes simultaneously. As Etienne Bennequin and colleagues argue in Few-Shot Image Classification Benchmarks are Too Far From Reality, current models suffer a "catastrophic performance degradation" when transitioning from 5-way to 100-way classification. This research direction, often termed Many-Way Few-Shot (MCFS) learning, focuses on closing the gap between restricted laboratory settings and high-cardinality real-world environments.
The Geometry of Crowded Feature Spaces
Why does increasing the number of classes (the "way") make the problem so much harder? In a 5-way task, the decision boundaries are wide and forgiving. But as the number of classes grows, the feature space becomes "crowded."
Lu Liu and others note in Many-Class Few-Shot Learning on Multi-Granularity Class Hierarchy that increasing the target classes exponentially expands the search space and increases the likelihood of inter-class confusion. When a model only has one or five examples per class, it struggles to define the precise boundaries needed to separate a "novel" class from dozens of semantically similar "base" classes. This is particularly evident in Zero-Shot Learning (ZSL), where Yuchen Guo demonstrates in Zero-shot Learning with Many Classes by High-rank Deep Embedding Networks that standard "low-rank" mapping functions simply lack the expressive power to distinguish between tens of thousands of fine-grained categories.
Leveraging Class Hierarchies
One of the most effective ways to navigate a crowded feature space is to use the inherent structure of the world. Humans don't learn "Golden Retriever" and "Labrador" as isolated concepts; we understand them as sub-types of "Dog."
Researchers are increasingly using class hierarchies to provide a roadmap for the model. Aoxue Li and Tao Xiang show in Large-Scale Few-Shot Learning: Knowledge Transfer With Class Hierarchy that by leveraging automatically generated taxonomies, models can learn more transferable features. Instead of treating every class as an independent island, the model learns to position new classes within an existing tree of knowledge, which significantly improves performance when scaling to thousands of categories.
From Point Comparisons to Distributional Relationships
In low-way settings, models often use simple metric learning—calculating the distance between a query image and a class "prototype." But in many-way scenarios, these point-to-point comparisons become noisy and unreliable.
To combat this, Ling Yang and colleagues introduced the Distribution Propagation Graph Network (DPGN). Rather than looking at individual samples in isolation, DPGN models the relationships between the distributions of different classes. By propagating these distributional relationships across a dual-graph architecture, the model gains a more global view of the label space, making it more robust when it has to choose between many competing categories.
Similarly, in the realm of natural language, Ruiying Geng and Xiaodan Zhu proposed Induction Networks. These networks use dynamic routing (borrowed from Capsule Networks) to "induce" a generalized class representation. This move from specific examples to abstract class-level "intent" is crucial for maintaining stability in massive-way scenarios involving hundreds of potential text categories.
Redefining the Evaluation Paradigm
The shift toward many-way learning is as much about how we test as it is about what we build. New datasets like FEW-NERD, released by Ning Ding and researchers at Tsinghua and Alibaba, provide a hierarchy of 66 fine-grained entity types to simulate the complexity of real-world information extraction.
As the field moves away from 5-way "toy" tasks, the focus is shifting toward Generalized Few-Shot Learning (G-FSL). In this setting, as explored by Zhuotao Tian in Generalized Few-shot Semantic Segmentation, the model must not only learn new classes but also continue to recognize the "base" classes it was originally trained on. This "open-world" requirement is the ultimate test for many-way systems: can they expand their vocabulary indefinitely without forgetting the past or becoming confused by the sheer number of possibilities?