Few-shot Class-variable Incremental Audio Classification via Prototype Adaptation and Pseudo Class-variable Training
Researchers propose a new method for few-shot class-variable incremental audio classification that handles both increasing and decreasing numbers of classes, addressing a practical gap in existing models. The approach uses prototype adaptation and pseudo class-variable training to dynamically adjust classifier structure as classes change, demonstrating improved performance on multiple datasets.
This research addresses a fundamental limitation in incremental learning systems for audio classification. While existing few-shot learning methods assume classes only increase over time, real-world applications frequently encounter scenarios where class counts fluctuate—training data becomes obsolete, services consolidate categories, or production systems must handle variable classification scopes. The proposed FCIAC method tackles this asymmetry by introducing a dynamically adaptive classifier architecture paired with a pseudo class-variable training strategy that conditions the model to handle both expansion and contraction of its classification space.
The technical contribution centers on two mechanisms: a class-variable prototype adaptation network that restructures itself based on the current class set, and a training approach that simulates class changes before deployment. This contrasts with traditional few-shot learning frameworks that treat class addition unidirectionally. The work represents meaningful incremental progress in machine learning, particularly for audio systems that must operate in non-stationary environments such as acoustic event detection, speaker identification in evolving environments, or domain-adaptive sound classification.
For practitioners developing audio ML systems, this method extends the practical applicability of incremental learning to more realistic scenarios. The research validates performance across three public datasets, suggesting the approach generalizes reasonably well. However, the impact remains primarily academic rather than commercial—the improvements benefit specific audio classification pipelines but don't fundamentally reshape the industry. The public code release enables adoption by researchers, though deployment in production systems would require integration with existing audio processing infrastructure and evaluation against domain-specific baselines.
- →Proposes first method handling both class increases and decreases in few-shot audio classification, addressing practical limitations of existing approaches
- →Uses adaptive prototype networks and pseudo class-variable training to dynamically adjust classifier structure during deployment
- →Demonstrates improved average accuracy across three public audio datasets compared to previous methods
- →Enables more realistic incremental learning scenarios where classification scopes naturally contract or expand over time
- →Code publicly available, facilitating research adoption and further development in incremental audio classification