[Paper Review] A Hybrid Approach with Optimization and Metric-based Meta-Learner for Few-Shot Learning

Duo Wang, Yu Cheng+3 more|arXiv (Cornell University)|Apr 4, 2019

Domain Adaptation and Few-Shot Learning35 references3 citations

TL;DR

This paper proposes Meta-Metric-Learner, a hybrid few-shot learning framework that combines optimization-based meta-learning with metric-based base learners to handle tasks with flexible class numbers and improve generalization. By using a Meta-SGD meta-learner to guide gradient updates in a task-specific metric learner, the model achieves state-of-the-art performance across both standard and realistic few-shot settings with unbalanced class distributions and multi-source data.

ABSTRACT

Few-shot learning aims to learn classifiers for new classes with only a few training examples per class. Most existing few-shot learning approaches belong to either metric-based meta-learning or optimization-based meta-learning category, both of which have achieved successes in the simplified "$k$-shot $N$-way" image classification settings. Specifically, the optimization-based approaches train a meta-learner to predict the parameters of the task-specific classifiers. The task-specific classifiers are required to be homogeneous-structured to ease the parameter prediction, so the meta-learning approaches could only handle few-shot learning problems where the tasks share a uniform number of classes. The metric-based approaches learn one task-invariant metric for all the tasks. Even though the metric-learning approaches allow different numbers of classes, they require the tasks all coming from a similar domain such that there exists a uniform metric that could work across tasks. In this work, we propose a hybrid meta-learning model called Meta-Metric-Learner which combines the merits of both optimization- and metric-based approaches. Our meta-metric-learning approach consists of two components, a task-specific metric-based learner as a base model, and a meta-learner that learns and specifies the base model. Thus our model is able to handle flexible numbers of classes as well as generate more generalized metrics for classification across tasks. We test our approach in the standard "$k$-shot $N$-way" few-shot learning setting following previous works and a new realistic few-shot setting with flexible class numbers in both single-source form and multi-source forms. Experiments show that our approach can obtain superior performance in all settings.

Motivation & Objective

• To address the limitations of existing few-shot learning methods that assume uniform class numbers across tasks.
• To overcome the domain generalization issues in metric-based approaches that rely on a single task-invariant metric.
• To enable optimization-based meta-learning to handle heterogeneous, task-specific network structures with varying numbers of classes.
• To improve generalization in few-shot learning by learning task-adaptive metrics guided by a meta-learner.
• To evaluate performance in both single-source and multi-source few-shot learning settings with realistic class distribution variations.

Proposed method

• The model uses a Meta-SGD meta-learner to predict initial parameters and guide gradient descent in a task-specific metric-based base learner.
• The base learner employs a metric-learning approach (e.g., Prototypical Network or Matching Network) that is non-parametric and robust to varying class counts.
• The meta-learner's predictions are used to parameterize the metric space, enabling task-specific metric adaptation.
• The framework supports both single-source and multi-source few-shot learning, where auxiliary data sources improve generalization.
• The model is trained in a two-stage process: meta-learner learns to predict base model parameters, and the base model is fine-tuned per task.
• The method is evaluated in both 1-shot and 5-shot settings with data augmentation via 90-degree rotations on Omniglot.

Experimental results

Research questions

• RQ1 Can a hybrid meta-learning framework effectively handle few-shot learning tasks with varying numbers of classes?
• RQ2 Can a meta-learner guide metric-based models to learn task-specific, adaptive metrics across diverse domains?
• RQ3 Does combining optimization-based meta-learning with metric-based base learners improve generalization in multi-source few-shot learning?
• RQ4 How does the proposed method perform compared to existing baselines in unbalanced and realistic few-shot settings?
• RQ5 Can the model generalize across different data sources (e.g., alphabets in Omniglot) without requiring uniform class structures?

Key findings

• In the standard 5-way 1-shot setting, Meta-Metric-Learner outperforms baseline models including Meta-SGD and Matching Network.
• On the 3 vs. 5 shot split, which is more challenging, the model still achieves superior accuracy, demonstrating robustness to class imbalance.
• In the multi-source setting on Omniglot, the model achieves 86.84% accuracy in the 1-shot, 30% class split, outperforming all baselines.
• The model converges faster and achieves higher test accuracy than baselines in unbalanced class scenarios, as shown in training curves.
• The method maintains strong performance even when meta-train and meta-test sets have unbalanced class distributions, proving its flexibility.
• The model shows consistent improvement over existing methods in both single-source and multi-source forms, especially in realistic few-shot settings.