Learning Order Matters in Class-Incremental Learning for Sound Localization and Detection

Abstract

This study investigates the impact of class learning order in Class-Incremental Learning (CIL) for Sound Event Localization and Detection (SELD) by systematically evaluating class-wise localization error (LE) and F1-score across different class-ordering scenarios. A continual learning model is trained in two stages: initially on nine classes, then incrementally extended with four additional classes that vary in acoustic complexity. The results show that strategically introducing acoustically challenging (difficult to recognize) classes in the incremental learning stage enhances overall SELD performance, leading to increased F1-scores and reduced LE compared to a baseline that learns all the classes simultaneously. Furthermore, this study compares performance across balanced and imbalanced datasets, demonstrating consistent trends and highlighting the critical influence of class order. The study offers insights for designing more robust CIL frameworks for the SELD task.