[Paper Review] Prototype Propagation Networks (PPN) for Weakly-supervised Few-shot Learning on Category Graph

Lu Liu, Tianyi Zhou+4 more|arXiv (Cornell University)|May 10, 2019

Domain Adaptation and Few-Shot Learning4 citations

TL;DR

This paper proposes Prototype Propagation Networks (PPN), a meta-learning framework that leverages weakly-labeled data with coarse-class annotations to improve few-shot classification. By propagating prototypes across a category graph using an attention-based mechanism, PPN enhances KNN classification accuracy, achieving state-of-the-art performance on WS-ImageNet-Pure and WS-ImageNet-Mix, with up to 15% relative improvement in 1-shot settings over baselines.

ABSTRACT

A variety of machine learning applications expect to achieve rapid learning from a limited number of labeled data. However, the success of most current models is the result of heavy training on big data. Meta-learning addresses this problem by extracting common knowledge across different tasks that can be quickly adapted to new tasks. However, they do not fully explore weakly-supervised information, which is usually free or cheap to collect. In this paper, we show that weakly-labeled data can significantly improve the performance of meta-learning on few-shot classification. We propose prototype propagation network (PPN) trained on few-shot tasks together with data annotated by coarse-label. Given a category graph of the targeted fine-classes and some weakly-labeled coarse-classes, PPN learns an attention mechanism which propagates the prototype of one class to another on the graph, so that the K-nearest neighbor (KNN) classifier defined on the propagated prototypes results in high accuracy across different few-shot tasks. The training tasks are generated by subgraph sampling, and the training objective is obtained by accumulating the level-wise classification loss on the subgraph. The resulting graph of prototypes can be continually re-used and updated for new tasks and classes. We also introduce two practical test/inference settings which differ according to whether the test task can leverage any weakly-supervised information as in training. On two benchmarks, PPN significantly outperforms most recent few-shot learning methods in different settings, even when they are also allowed to train on weakly-labeled data.

Motivation & Objective

• To improve few-shot classification performance by integrating weakly-supervised, coarse-label data into meta-learning frameworks.
• To address the challenge of limited labeled data in few-shot learning by leveraging inexpensive, hierarchical category relationships.
• To develop a prototype propagation mechanism that transfers knowledge across related classes via a category graph.
• To enable continual adaptation of prototypes across new tasks and classes using a reusable, graph-structured computational model.

Proposed method

• PPN trains on few-shot tasks combined with weakly-labeled data annotated at coarse levels, using a category graph to model hierarchical class relationships.
• An attention mechanism computes edge weights for propagating prototypes from parent classes to child classes on the graph.
• The model minimizes classification loss across subgraphs sampled from the category graph, using a level-wise training objective.
• Prototypes are updated via backpropagation, allowing error signals to flow from fine classes to coarse classes and vice versa.
• The framework supports two inference settings: one with and one without access to weakly-labeled data during testing.
• A computational graph of prototypes is maintained and updated for continual use across new tasks.

Experimental results

Research questions

• RQ1 Can weakly-supervised, coarse-label data significantly improve few-shot classification performance in meta-learning?
• RQ2 How can prototype propagation across a category graph enhance generalization in low-data regimes?
• RQ3 Does the attention-based propagation mechanism effectively transfer knowledge between semantically related classes?
• RQ4 How robust is the method when weakly-labeled data are unrelated to the few-shot tasks?
• RQ5 Can the prototype propagation mechanism outperform standard meta-learning baselines even when they are also trained on weakly-labeled data?

Key findings

• PPN achieves state-of-the-art performance on WS-ImageNet-Pure, with a 15% relative improvement in 1-shot classification over other methods.
• PPN+ outperforms all baselines on WS-ImageNet-Pure, especially in 1-shot settings, demonstrating the effectiveness of weakly-labeled data integration.
• On WS-ImageNet-Mix, where weakly-labeled data are unrelated to few-shot tasks, PPN+ still achieves a 10% relative improvement in 1-shot accuracy.
• Increasing the weight of propagated prototypes (1−λ) consistently improves validation accuracy, confirming the effectiveness of the propagation mechanism.
• The average per-iteration time of PPN is only ~10% higher than Prototypical Networks, indicating efficient inference with minimal computational overhead.
• The model’s performance remains strong even when weakly-labeled data are not directly linked to the few-shot tasks, showing robustness to data distribution shifts.