[Paper Review] Learning to Propagate Labels: Transductive Propagation Network for Few-shot Learning

Yanbin Liu, Juho Lee+5 more|arXiv (Cornell University)|May 25, 2018

Domain Adaptation and Few-Shot Learning486 citations

TL;DR

TPN introduces end-to-end transductive label propagation for few-shot learning by learning a graph on the union of support and query sets, achieving state-of-the-art results on mini ImageNet and tiered ImageNet.

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.

Motivation & Objective

• Motivate solving few-shot classification by exploiting transductive inference over the whole query set to alleviate data scarcity.
• Propose an end-to-end learnable framework that jointly learns feature embeddings and a task-adaptive graph for label propagation.
• Leverage episodic training to simulate test conditions and optimize the label propagation mechanism.
• Demonstrate that learning to propagate labels via a transductive graph improves accuracy over inductive/meta-learning baselines on standard benchmarks.

Proposed method

• Embed inputs with a CNN to obtain feature representations f_phi(x).
• Construct an example-wise length-scale graph parameter sigma_i via a graph construction module g_phi applied to f_phi(x_i) for all S ∪ Q.
• Compute a kNN similarity graph W using per-example scaled embeddings and normalize with a graph Laplacian S = D^(-1/2) W D^(-1/2).
• Perform label propagation F* = (I - αS)^(-1) Y to propagate labels from the support set to the query set, where Y encodes known labels.
• Train end-to-end by minimizing cross-entropy between propagated scores F* (softmaxed) and ground-truth labels on S ∪ Q, updating φ and φ with backpropagation.
• Graph construction per episode (N-way K-shot) learns to adapt to task manifolds; complexity is kept small since n ≈ N×K + T (e.g., 80–100).

Experimental results

Research questions

• RQ1 Can transductive label propagation, learned end-to-end within an episodic meta-learning framework, outperform inductive few-shot methods?
• RQ2 Does learning example-wise graph scales and distances improve propagation of labels in novel class spaces?
• RQ3 What is the impact of higher-shot or higher-way training strategies on transductive few-shot performance?
• RQ4 How does transductive propagation fare on standard benchmarks (mini ImageNet, tiered ImageNet) compared to state-of-the-art methods?

Key findings

• TPN achieves state-of-the-art results on mini ImageNet and tiered ImageNet, outperforming inductive and other transductive baselines.
• Learning an example-wise, task-adaptive graph for label propagation yields significant gains, especially in 1-shot scenarios.
• Higher-shot training (more examples per task) further improves TPN performance, with larger gains in 1-shot settings.
• Semi-supervised adaptation (TPN-semi) further boosts results over comparable semi-supervised few-shot baselines.
• Transductive methods leveraging the entire query set are particularly advantageous when labeled data is scarce, with diminishing returns as the labeled set grows.