Prototype network learns to reject unknown sounds with limited data

Every machine learning model that classifies sound carries a hidden assumption: that the world contains only the categories it was trained on. A network taught to recognize dog barks, sirens, rain, and footsteps will, when confronted with the whine of a drone or the crackle of an unfamiliar industrial machine, confidently force the new sound into one of its known boxes. In the laboratory this is a nuisance. In the real world, where acoustic monitoring systems listen for anomalies in factories, cities, and natural environments, it can be dangerous. A safety system that cannot say “I don’t know” is a system that cannot raise an alarm. A new study published in Cluster Computing by Yanmei Jiang and Xiong Li of Sun Yat-sen University tackles this problem directly, proposing an improved convolutional prototype network that classifies known environmental sounds and rejects unknown ones, even when trained on very little labeled data.

The challenge the researchers address is known in machine learning as open-set recognition, and it is fundamentally harder than the closed-set classification that dominates the literature. In closed-set scenarios, every test sample is guaranteed to belong to one of the training categories, so the model only needs to draw boundaries between known classes. In open-set scenarios, the model must simultaneously perform two tasks: classify samples from known classes correctly and identify samples from classes it has never seen as unknown. The scarcity of labeled sound samples makes both tasks harder, because learning discriminative representations and reliable class prototypes requires enough examples to characterize what each class sounds like and, just as importantly, where its boundaries lie.

Environmental sound classification has advanced rapidly with deep learning, with convolutional neural networks, attention-based residual networks, and transformer architectures all achieving strong results on benchmark datasets. But most of these studies assume closed-set conditions. The open-set problem in acoustic scene classification was formally identified nearly a decade ago, and subsequent work has explored rejection-based multi-class classification, CNN-based open-set audio recognition, and generative approaches. Prototype-based methods have emerged as a particularly promising family for open-set recognition because they represent each class by a point, or prototype, in feature space, and classify samples by their proximity to these points. Samples that fall far from every prototype can be flagged as unknown. The convolutional prototype network, originally developed for open-set recognition in computer vision, forms the foundation of the new approach.

The method proposed by Jiang and Li contains three complementary components, each targeting a specific weakness of prototype networks under limited data conditions. The first is a time-frequency attention mechanism that emphasizes perceptually salient components of the spectrogram, the two-dimensional representation of sound in which time runs along one axis and frequency along the other. Not all parts of a spectrogram are equally informative for distinguishing one sound from another. A dog bark may be characterized by transient broadband bursts, while a siren may be defined by narrow tonal lines that sweep across frequency. Attention mechanisms allow the network to learn which time-frequency regions carry the most discriminative information and to weight them more heavily during feature extraction, enhancing the quality of the learned representations.

The second component is a Mahalanobis distance-based complementary prototype loss, which exploits information from misclassified samples to refine prototype representations. In conventional prototype learning, prototypes are updated based on correctly classified samples, and misclassified samples are treated as noise. But misclassified samples carry valuable information: they reveal where the decision boundaries between classes are ambiguous and where prototypes are poorly positioned. The complementary loss draws on the complementary information contained in these difficult samples, using Mahalanobis distance, a metric that accounts for the covariance structure of the feature distribution rather than treating all directions in feature space equally, to reshape prototypes in ways that reduce confusion between classes. This is especially important when data is limited, because with few examples per class, every sample, including the ones the model gets wrong, matters for characterizing the class distribution.

The third component is a learnable Mahalanobis metric for adaptive similarity modeling between feature embeddings and class prototypes. In standard prototype networks, similarity is measured using Euclidean distance, which assumes that all feature dimensions contribute equally and independently to the distance calculation. The Mahalanobis metric relaxes this assumption by learning a full distance function that adapts to the geometry of the feature space. Features that are correlated, redundant, or noisy can be downweighted, while discriminative dimensions can be emphasized. Because the metric is learnable, the network can discover the most appropriate way to compare embeddings with prototypes for the specific sound classification task at hand, rather than relying on a fixed geometric assumption that may not match the structure of acoustic features.

Experiments on environmental sound datasets demonstrate that the proposed method consistently improves both known-class classification accuracy and unknown-class rejection, with the largest gains appearing in low-data open-set scenarios. This is a significant result because it addresses the two failure modes that plague prototype networks when data is scarce: prototypes that are poorly estimated from too few examples, and feature representations that are not discriminative enough to separate known classes from unknown ones. By combining attention-enhanced feature learning with prototype refinement from misclassified samples and an adaptive distance metric, the method builds a more robust geometric structure in feature space, one in which known classes form tight, well-separated clusters and unknown samples fall clearly outside all clusters.

The implications extend beyond the laboratory. Environmental sound recognition is increasingly deployed on embedded devices for applications including wildlife monitoring, urban noise assessment, industrial fault detection, and smart home systems. These deployments often face exactly the conditions the new method targets: limited labeled data, because annotating sound recordings is time-consuming and expensive, and open-set conditions, because the acoustic environment is never fully predictable. A monitoring system for a nature reserve may be trained on a handful of species calls but must reject the sounds of wind, rain, passing vehicles, and human activity that it was never trained on. An industrial acoustic fault detector may know the signatures of a few machine states but must flag novel anomalies rather than misclassifying them as normal operation.

The work also connects to a broader research program on prototype learning and few-shot recognition. Prototypical networks were originally developed for few-shot learning, where models must generalize to new classes from only a handful of examples, and they have since been applied to open-set recognition, noisy label learning, and contrastive learning. The new study builds on the convolutional prototype network for open-set recognition, which was developed in the computer vision community, and adapts it to the acoustic domain with modifications that address the specific challenges of sound data. The authors’ related work on cognitive prototype learning for few-shot open-set sound recognition and on attention-fused prototypes for few-shot open-set audio classification suggests a sustained research effort in this direction, with the current study representing a step toward more general and data-efficient acoustic recognition systems.

As deep learning continues to push the frontiers of what machines can hear and understand, the ability to say “this sound is unlike anything I know” may prove as important as the ability to name the sounds it does know. The improved convolutional prototype network proposed by Jiang and Li offers a concrete technical path toward that goal, demonstrating that with careful attention to feature learning, prototype estimation, and distance modeling, machines can learn to classify known sounds and reject unknown ones from limited data. For the growing community of researchers building acoustic monitoring systems for the real world, where the soundtrack is never fully known in advance, this work provides both a practical tool and a conceptual framework for building systems that know the limits of their own knowledge.

Subject of Research: Machine learning model for open-set environmental sound classification with limited data

Article Title: An improved convolutional prototype network for open-set sound classification with limited data

Article References: Jiang, Y., & Li, X. (2026). An improved convolutional prototype network for open-set sound classification with limited data. Cluster Computing, 29(15), Article 869. https://doi.org/10.1007/s10586-026-06724-3

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06724-3

Keywords: open-set recognition, environmental sound classification, prototype network, deep learning, few-shot learning, Mahalanobis distance, attention mechanism, acoustic monitoring, improved, convolutional, prototype, network