[Paper Review] Adversarial Unsupervised Domain Adaptation For Acoustic Scene Classification

Shayan Gharib, Konstantinos Drossos+3 more|arXiv (Cornell University)|Aug 17, 2018

Music and Audio Processing1 references44 citations

TL;DR

Introduces an unsupervised adversarial domain adaptation method for acoustic scene classification that adapts a source-domain model to target-domain data without target labels, improving unseen-target accuracy by about 10%.

ABSTRACT

A general problem in acoustic scene classification task is the mismatched conditions between training and testing data, which significantly reduces the performance of the developed methods on classification accuracy. As a countermeasure, we present the first method of unsupervised adversarial domain adaptation for acoustic scene classification. We employ a model pre-trained on data from one set of conditions and by using data from other set of conditions, we adapt the model in order that its output cannot be used for classifying the set of conditions that input data belong to. We use a freely available dataset from the DCASE 2018 challenge Task 1, subtask B, that contains data from mismatched recording devices. We consider the scenario where the annotations are available for the data recorded from one device, but not for the rest. Our results show that with our model agnostic method we can achieve $\sim 10\%$ increase at the accuracy on an unseen and unlabeled dataset, while keeping almost the same performance on the labeled dataset.

Motivation & Objective

• Address domain shift between training and testing in acoustic scene classification (ASC).
• Propose a model-agnostic unsupervised adversarial domain adaptation framework for ASC.
• Show that leveraging unlabeled target-domain data improves unseen-domain accuracy while preserving source-domain performance.

Proposed method

• Pre-train a source-domain model M_S and a label classifier C on labeled source data X^S.
• Use an adversarial step with a target-domain model M_T and a domain discriminator D to align distributions of M_S(X^S) and M_T(X^T).
• Keep M_S and M_T unshared (no weight tying) and initialize M_T with M_S parameters.
• Incorporate the label classifier C into the adaptation step to provide a supervised learning signal during domain alignment.
• Optimize L_D to train D to distinguish source vs. target and L_M_T to train M_T to fool D while classifying source-like targets using C.
• Test M_T on X^T with C to obtain target-domain predictions.

Experimental results

Research questions

• RQ1 Can unsupervised adversarial domain adaptation align latent representations across mismatched acoustic domains without target labels?
• RQ2 Does incorporating the label classifier C during adaptation improve target-domain accuracy without severely harming source-domain performance?
• RQ3 Is the method architecture-agnostic and effective across different ASC models?
• RQ4 How does performance on unseen target data compare before and after adaptation?

Key findings

• On two ASC models (Kaggle and DCASE baselines), target-domain accuracy improved from 20.28% to 31.67% and 19.17% to 25.28% respectively after adaptation.
• Source-domain accuracy remained essentially stable, with a slight increase for the Kaggle model (65.25% to 65.37%) and a small decrease for the DCASE model (61.71% to 61.23%).
• Adaptation achieved approximately a 10% absolute gain in unseen target-domain accuracy across models.
• Confusion matrices show substantially improved correct classifications across target-domain labels after adaptation.