Few-Shot Open-Set Recognition using Meta-Learning

1. Introduction

• The authors seek to extend open-set recognition from its primarily large-scale setting to a continuum that includes few-shot recognition.They focus on the two extremes, large-scale and few-shot, because training across many settings can be complex.Source: §1 · p. 2
• Few-shot open-set recognition is framed as useful and challenging because labeled data are scarce and both tasks require decisions about unseen data.The authors expect few-shot architectures' robustness to help them outperform large-scale classifiers on open-set recognition.Source: §1 · p. 2
• The proposed meta-learning procedure samples novel classes per episode and maximizes posterior entropy for their examples.The authors say this encourages high-entropy posteriors beyond target-class regions.Source: §1 · p. 2
• PEELER reports significant open-set recognition improvements over the state of the art and gains from a Mahalanobis distance over Euclidean distance.The authors report benefits in both few-shot and large-scale settings.Source: §1 · p. 2
• The stated contributions include a few-shot open-set formulation, episodic training that combines cross-entropy with an open-set loss, and a Gaussian embedding.The formulation is also intended to improve open-set performance in large-scale recognition.Source: §1 · p. 2

Open-Set Recognition

• Prior open-set methods reduce unseen classes to one extra class, whereas this work allows a cluster per seen class and treats samples outside those clusters as unseen.The authors argue that visually different unknown categories are difficult to map into a single feature-space cluster.Source: §2 · p. 3

Out-of-Distribution

• Out-of-distribution detection concerns samples outside the training distribution, which may include perturbed seen-class examples, unlike open-set recognition's unseen classes from the same dataset.The passage notes that OOD samples are frequently easier to detect than unseen classes.Source: §2 · p. 3

Few-Shot Learning

• Few-shot methods include optimization-based and metric-based approaches, with metric methods comparing support and query features.Examples include episodic training, prototypes from support features, and learned relations between support and query features.Source: §2 · p. 3
• Whether few-shot classifiers can reject unseen samples remains unclear, and this work studies that question in large-scale and few-shot scenarios.The paper focuses on traditional few-shot learning, leaving learning without forgetting beyond its scope.Source: §2 · p. 3

3.1. Classification settings

• Open-set classification allows test classes to include unseen classes, while closed-set classification uses the same classes for training and testing.Few-shot recognition uses few training examples per class, usually fewer than twenty; the paper focuses on the largely unexplored few-shot open-set setting.Source: §3.1 · p. 4

3.2. Meta-learning

• Meta-learning trains across learning problems and maps each support/query pair (S, T) into an algorithm that uses S to solve T.At test time, the resulting model is evaluated on the query set.Source: §3.2 · p. 4
• Prototypical-network episodes sample N classes and K examples per class, randomizing the classification task used to optimize the embedding.The model uses a softmax classifier, with a Gaussian-mean estimator and back-propagation to update the embedding.Source: §3.2 · p. 4
• Because episodic randomization is intended to improve generalization to unseen data, the authors propose meta-learning for both large-scale and few-shot open-set classification.The paper argues that few-shot learning and open-set recognition both require handling samples or classes not seen during training.Source: §3.2 · p. 4

4.1. Open-Set Meta-Learning

• PEELER uses open-set episodes: the support set follows standard meta-learning, while the query set is augmented with unseen classes.Its optimization combines a seen-class classification loss with an open-set loss for unseen examples.Source: §4.1 · p. 5
• In few-shot episodes, N classes and K examples per class form the seen-class support set, and M additional unseen classes are included in the query set.The support and query sets are used in the episode's optimization.Source: §4.1 · p. 5

i. However, the query set Ts

• For large-scale open-set episodes, M classes are randomly sampled as unseen, while the remaining classes are treated as seen.The seen classes adjust the model without a support-set meta-learning step, and the loss from (8) is still applied.Source: i. However, the query set Ts · p. 5
• The open-set loss maximizes entropy over seen-class probabilities for unseen examples, so their probability mass is not concentrated on a seen class.The method uses negative entropy for this objective.Source: i. However, the query set Ts · p. 5

4.2. Gaussian Embedding

• PEELER assigns samples far from the nearest class prototype to unseen; few-shot prototypes are class means, while large-scale prototypes are learned network parameters.The method is based on the prototypical-network architecture.Source: §4.2 · p. 5
• Shared-covariance Euclidean distances can be suboptimal for open-set recognition because optimal seen/unseen boundaries may vary by class.PEELER replaces Euclidean distance with Mahalanobis distance under class-specific Gaussian covariances, assumed diagonal to keep parameters manageable.Source: §4.2 · p. 6
• For large-scale recognition, class precision matrices are learned by back-propagation; for few-shot recognition, a learnable embedding function is introduced.The supplied passage identifies this setting-dependent learning of precision matrices but does not specify the full few-shot construction.Source: §4.2 · p. 6

5.1. Large-Scale Open-Set Recognition

• Across both backbones, PEELER outperforms prior methods on open-set AUROC at comparable seen-class accuracy, with gains from Gaussian embedding and open-set loss.The large-scale comparison includes CIFAR10 and extended miniImageNet.Source: §5.1 · pp. 6–7, Table 2 · p. 7

5.2. Few-Shot Open-Set Recognition

• OpenMax and Counterfactual underperform PEELER in few-shot recognition; open-set loss, Gaussian embedding, and their combination all provide gains.An AUROC of 50% corresponds to chance performance.Source: Table 3 · p. 7, §5.2 · p. 7

5.3. Ablation Studies

• 10-way classification underperforms 5-way, but open-set performance improves significantly, indicating greater class diversity is more robust to open-set samples.Source: Table 4 · p. 7, §5.3 · p. 8
• With fixed total open-set samples, varying unseen classes per episode causes only minor performance variation relative to the baseline gain.With many training episodes, the model converges well regardless of how many open-set classes appear in an episode.Source: §5.3 · p. 7, Figure 3 · p. 8
• Varying open-query class count from 1 to 10 or the number of samples per class causes only small AUROC changes.The study keeps total open-query samples unchanged when varying class count and keeps the class count at five when varying samples per class.Source: §5.3 · pp. 7–8, Figure 3 · p. 8

5.4. Weakly Supervised Object Discovery

• With 1–3 annotated relevant frames, PEELER halves misclassified frames versus VODC without requiring annotated irrelevant frames.The evaluation uses relevant frames as support and detects irrelevant frames as open-set samples.Source: §5.4 · p. 8, Table 5 · p. 8

• PEELER combines an open-set loss with a redesigned metric-learning approach, achieves state-of-the-art few-shot and large-scale open-set recognition, and is effective on weakly supervised object discovery.Source: §6 · p. 8