Towards robust episodic meta-learning

Beyza Ermis, Giovanni Zappella, Cédric ArchambeauUAI 2021 · OpenReview

Abstract: Meta-learning learns across historical tasks with the goal to discover a representation from which it is easy to adapt to unseen tasks. Standard episodic meta-learning relies on randomly sampled "episodes" to train models, which can be inefficient and brittle when faced with diverse or imbalanced data. This paper introduces Active Episodic Meta-Learning, a method that uses active learning scoring rules—specifically uncertainty sampling—to strategically compose training and testing episodes. By selecting the most informative data points rather than random ones, we achieve significantly better generalization, faster convergence, and improved robustness to out-of-distribution tasks.

1. Introduction

Meta-learning, or "learning to learn," aims to create models that can adapt to new tasks with very few examples (few-shot learning). The standard approach is episodic training, where the model is trained on a series of mini-tasks (episodes) consisting of a "context set" (to adapt the model) and a "query set" (to evaluate and update it).

However, this paradigm has a weakness: episodes are typically constructed by picking data points uniformly at random. In the real world, data is often messy—classes are imbalanced, and new tasks might look very different from training tasks. Random sampling often fails to capture the most "difficult" or informative examples, leading to models that are brittle and slow to adapt. We seek to bridge the gap between meta-learning and Active Learning (AL) to make the training process more intentional and robust.

2. Methodology

We propose replacing random episode generation with a scoring-based procedure focused on Uncertainty Sampling. In this framework, the model selects data points it is most "confused" about, measured by the entropy of the predicted class probabilities. We apply this in three distinct ways:

  1. Active Context Selection (Training): The model selects the most informative samples to form the context set, forcing it to learn how to adapt using high-utility information.
  2. Active Query Selection (Training): The model selects difficult samples for the query set to provide a stronger gradient signal, pushing the model toward more robust representations.
  3. Active Context Selection (Testing): At inference time, given a pool of unlabeled data, the model actively picks which samples should be labeled to facilitate the most effective adaptation.

These techniques were integrated into three popular meta-learning frameworks: Prototypical Networks (metric-based), foMAML (optimization-based), and protoMAML (a hybrid).

3. Experimental Evaluation

We evaluated our approach on several standard few-shot learning benchmarks, including Omniglot, miniImageNet, CIFAR100, and TieredImageNet. Additionally, we tested on a custom heterogeneous mixture (MINI+OMNI+CIFAR) to evaluate performance under significant distribution shifts.

3.1 Key Results

Across all models (ProtoNets, foMAML, protoMAML), active selection consistently outperformed random selection. As shown in Figure 1, the performance gap is particularly pronounced in more complex datasets and higher N-way tasks.

Accuracy comparison across models and datasets showing the superiority of active selection over random sampling.

On the mixed dataset (MINI+OMNI+CIFAR), the gap between active and random selection widened, proving that active meta-learning is significantly more robust when the test tasks are very different from the training tasks. Furthermore, active meta-learning converges much faster; on TieredImageNet, the active approach achieved higher accuracy in 1,000 episodes than the random approach did in 2,000.

3.2 Robustness to Imbalance

In real-world scenarios where some classes have many more examples than others, random sampling can be highly biased. Active selection naturally mitigates this by focusing on informative samples regardless of class frequency. Figure 2 illustrates how active selection manages sample ratios across different task complexities.

Sample ratios across different N-way tasks, illustrating how active selection manages class distribution.

4. Discussion and Limitations

While Active Episodic Meta-Learning provides clear efficiency and accuracy gains, it introduces some computational overhead. Selecting samples based on uncertainty requires a forward pass for all candidate samples in a pool. While manageable for small few-shot pools, this could become a bottleneck for massive datasets. Future work will explore Batch Active Learning to select groups of samples simultaneously and Semi-Supervised Meta-Learning to leverage the unlabeled samples not selected for context sets.

5. Conclusion

By incorporating active learning principles into the episodic training paradigm, we have demonstrated that "harder" episodes lead to better meta-learners. Our method achieves superior accuracy with 50% to 60% fewer training episodes and shows remarkable robustness to the data imbalances and distribution shifts common in real-world applications.

Summary updated

Abstract

Meta-learning learns across historical tasks with the goal to discover a representation from which it is easy to adapt to unseen tasks. Episodic meta-learning attempts to simulate a realistic setti...
BibTeX
@inproceedings{ermis2021towards,
  author = {Beyza Ermis and Giovanni Zappella and Cédric Archambeau},
  title = {{Towards robust episodic meta-learning}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2021},
  url = {https://openreview.net/forum?id=rUX4qGQ1yQ5},
}
Figure