Jordi Pons

Experimenting with musically motivated convolutional neural networks

A common criticism of deep learning relates to the difficulty in understanding the underlying relationships that the neural networks are learning, thus behaving like a black-box. In this article we explore various architectural choices of relevance for music signals classification tasks in order to start understanding what the chosen networks are learning. We first discuss how convolutional filters with different shapes can fit specific musical concepts and based on that we propose several musically motivated architectures. These architectures are then assessed by measuring the accuracy of the deep learning model in the prediction of various music classes using a known dataset of audio recordings of ballroom music. The classes in this dataset have a strong correlation with tempo, what allows assessing if the proposed architectures are learning frequency and/or time dependencies. Additionally, a black-box model is proposed as a baseline for comparison. With these experiments we have been able to understand what some deep learning based algorithms can learn from a particular set of data.

Designing efficient architectures for modeling temporal features with convolutional neural networks

Many researchers use convolutional neural networks with small rectangular filters for music (spectrograms) classification. First, we discuss why there is no reason to use this filters setup by default and second, we point that more efficient architectures could be implemented if the characteristics of the music features are considered during the design process. Specifically, we propose a novel design strategy that might promote more expressive and intuitive deep learning architectures by efficiently exploiting the representational capacity of the first layer - using different filter shapes adapted to fit musical concepts within the first layer. The proposed architectures are assessed by measuring their accuracy in predicting the classes of the Ballroom dataset. We also make available1 the used code (together with the audio-data) so that this research is fully reproducible.