Abstract:The goal of few-shot learning is to learn a classifier that generalizes well even when trained with a limited number of training instances per class. The recently introduced meta-learning approaches tackle this problem by learning a generic classifier across a large number of multiclass classification tasks and generalizing the model to a new task. Yet, even with such meta-learning, the low-data problem in the novel classification task still remains. In this paper, we propose Transductive Propagation Network (TPN), a novel meta-learning framework for transductive inference that classifies the entire test set at once to alleviate the low-data problem. Specifically, we propose to learn to propagate labels from labeled instances to unlabeled test instances, by learning a graph construction module that exploits the manifold structure in the data. TPN jointly learns both the parameters of feature embedding and the graph construction in an end-to-end manner. We validate TPN on multiple benchmark datasets, on which it largely outperforms existing few-shot learning approaches and achieves the state-of-the-art results.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to improve the generalization ability of the classifier in few - shot learning when the number of training samples is extremely limited. Specifically, traditional methods are prone to over - fitting when dealing with new categories due to insufficient training data, resulting in poor performance of the model on new data. To solve this problem, this paper proposes a new meta - learning framework - Transductive Propagation Network (TPN), which propagates labels from labeled instances to unlabeled test instances, constructs graphs using the manifold structure in the data, and makes classification predictions for the entire test set in an overall manner. ### Main Contributions 1. **Explicitly Model Transductive Inference for the First Time**: Different from previous works, this paper explicitly models the transductive inference process in few - shot learning for the first time. 2. **Learn Label Propagation Graphs**: Through the meta - learning framework, the label propagation graph is learned in each simulated test set, which is significantly superior to heuristic - based label propagation methods. 3. **Experimental Verification**: Verification has been carried out on multiple benchmark datasets, and the results show that TPN not only outperforms existing methods in few - shot learning tasks but also performs excellently in semi - supervised learning settings. ### Core Methods - **Feature Embedding**: Use a deep neural network to map the input to an embedding space. - **Graph Construction Module**: Utilize the union of the support set and the query set to capture the manifold structure of new types of classes. - **Iterative Label Propagation**: According to the constructed graph structure, propagate labels from the support set to the query set in an iterative manner and finally obtain a closed - form solution. - **Loss Calculation**: Combine the propagated scores and the true labels of the query set to calculate the cross - entropy loss to update the feature embedding and graph construction parameters. Through these innovations, TPN can classify more effectively in the case of low data volume, thereby alleviating the data scarcity problem in few - shot learning.

Learning to Propagate Labels: Transductive Propagation Network for Few-shot Learning

Adaptive multi-scale transductive information propagation for few-shot learning

Feature Transformation for Few-Shot Learning

Ensemble Transductive Propagation Network for Semi-Supervised Few-Shot Learning

Adaptive Anchor Label Propagation for Transductive Few-Shot Learning

Transductive Few-shot Learning with Prototype-based Label Propagation by Iterative Graph Refinement

Transductive Relation-Propagation With Decoupling Training for Few-Shot Learning

Node Classification on Graphs with Few-Shot Novel Labels Via Meta Transformed Network Embedding

Graph-based few-shot learning with transformed feature propagation and optimal class allocation

Transductive Episodic-Wise Adaptive Metric for Few-Shot Learning

Meta-Learning for Semi-Supervised Few-Shot Classification

Adaptive-Step Graph Meta-Learner for Few-Shot Graph Classification

Graph Prototypical Networks for Few-shot Learning on Attributed Networks

LPN: Language-guided Prototypical Network for few-shot classification

Meta-learning Siamese Network for Few-Shot Text Classification.

Task-Equivariant Graph Few-shot Learning

SSL-ProtoNet: Self-supervised Learning Prototypical Networks for few-shot learning

Inductive Linear Probing for Few-shot Node Classification

Hybrid Graph Neural Networks for Few-Shot Learning

Bidirectional Gated Edge-Labeling Graph Recurrent Neural Network for Few-shot Learning

Transductive meta-learning with enhanced feature ensemble for few-shot semantic segmentation