Semi-Supervised Domain Adaptation for Acoustic Scene Classification by Minimax Entropy and Self-Supervision Approaches

Abstract

Unsupervised domain adaptation (UDA) has attracted significant attention and has been widely applied in acoustic scene classification (ASC). UDA methods have lower annotation costs; however, using only unannotated target samples to improve their performance is insufficient. Therefore, this study presents few-shot semi-supervised domain adaptation (SSDA) for ASC under device mismatch conditions. Few-shot SSDA is expected to have lower annotation costs for the target device compared to supervised domain generalization methods across devices. Moreover, few-shot SSDA is expected to provide better scene classification than UDA. The Minimax entropy (MME) approach for SSDA, previously proposed for image classification, might be successfully used for device mismatch ASC to achieve a domain-invariant discriminative model. However, there are still challenges such as ambiguous scene class labels and significant discrepancies between domains depending on the device. We adopt two subsets within self-supervised learning (SSL) proposed for domain adaptation into the few-shot SSDA. The first is learning discriminative feature representations using the in-domain self-supervision (InSelf) approach. The second is the adaptive prototype-classifier update (APCU) based on domain-specific estimated prototypes. Experiments on the TAU Urban Acoustic Scenes 2020 Mobile development dataset, which is a dataset with multiple target devices, provide the following two insights: (1) Compared with standard cross-entropy loss learning, the MME approach improves target device performance. (2) The SSL approach improves discriminative performance over MME for target samples with large deviations from the source samples.