Meta-Learning-Based Supervised Domain Adaptation for Melody Extraction

The task of extracting the dominant pitch from polyphonic audio is crucial in the music information retrieval field. A substantial amount of labeled audio data is required to effectively train the machine learning models to perform the task. Generally, the traditional models trained on audios of one domain, i.e., source, may not accurately extract pitch from audios of different domains, i.e., target. To boost the performance, the models are adapted on minimal labeled data from the target domain, a method known as the supervised domain adaptation. Typically, the meta-learning-based domain adaptation algorithm works well with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$N$</tex>-way <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$K$</tex>-shot classification tasks where the model i s adapted on <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$K$</tex> sam-ples for each <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$N$</tex> class in the target domain. This paper uses the meta-learning algorithm as the supervised domain adaptation method for the task of melody extraction. Further, we propose a novel weighting technique for handling class imbalance when adapting to a few audios in the target domain. The experimental findings indicate that t he proposed weighted meta-learning approach outperforms the adaptive baselines, namely fine-tuning and vanilla meta-learning.