LC-Protonets: Multi-Label Few-Shot Learning for World Music Audio Tagging

TL;DR

LC-Protonets address multi-label few-shot audio tagging in diverse world music by constructing LC-classes from the power sets of support labels and using LC-prototypes to classify queries. The method extends Prototypical Networks to multi-label settings, enabling zero-shot and few-shot in a unified framework, and demonstrates strong performance across Western and world music datasets. A two-step learning approach further expands tag sets without fine-tuning, while pre-trained backbones improve generalization. Although scalability increases as the number of labels grows, LC-Protonets offer practical gains for under-represented music cultures and provide a public benchmark for future work.

Abstract

We introduce Label-Combination Prototypical Networks (LC-Protonets) to address the problem of multi-label few-shot classification, where a model must generalize to new classes based on only a few available examples. Extending Prototypical Networks, LC-Protonets generate one prototype per label combination, derived from the power set of labels present in the limited training items, rather than one prototype per label. Our method is applied to automatic audio tagging across diverse music datasets, covering various cultures and including both modern and traditional music, and is evaluated against existing approaches in the literature. The results demonstrate a significant performance improvement in almost all domains and training setups when using LC-Protonets for multi-label classification. In addition to training a few-shot learning model from scratch, we explore the use of a pre-trained model, obtained via supervised learning, to embed items in the feature space. Fine-tuning improves the generalization ability of all methods, yet LC-Protonets achieve high-level performance even without fine-tuning, in contrast to the comparative approaches. We finally analyze the scalability of the proposed method, providing detailed quantitative metrics from our experiments. The implementation and experimental setup are made publicly available, offering a benchmark for future research.

Figures (4)

• Figure 1: The labels of the support set consisting of four items (bottom) and the LC-classes created from them (top). The concentric circles with the dashed lines pass through LCPs with same representations in the embedding space that are graphed equidistant from the query item $q$.
• Figure 2: Depiction of the two-step learning method: in the first step, a model is trained on well-represented tags via supervised learning; in the second step, the tag set is extended by applying the LC-Protonets method on the previously trained model, which serves as the backbone.
• Figure 3: t-SNE visualization of query items (in grey) and prototype embeddings (in distinct colors) for a "12-way 5-shot" ML-FSL task on the MagnaTagATune dataset. The left panel shows prototypes generated by the "ML-PNs" method (one per class), while the right panel displays those formed using the "LC-Protonets" method, where different colors within each prototype indicate the specific label combination it represents.
• Figure 4: Scalability metrics of the proposed method, averaged across all datasets. The $x$-axis represents the number of labels, the left $y$-axis shows the number of LC-Prototypes, and the right $y$-axis indicates the inference time per item. Both $y$-axes use the same logarithmic scale.