Cross-cultural Transfer Learning Using Sample-level Deep Convolutional Neural Networks

Abstract

This submission aims to show how cultural differences in music affect audio feature learning. Measuring how largely music content differ in each culture is conducted via various MIREX 2017 audio classification tasks in a transfer learning setting. In this submission, two pre-trained deep neural networks are provided. One is trained with Million Song Dataset with the LastFM tag annotations and the other is learned from a music dataset from NAVER with genre annotations. Furthermore, two methods are applied to improve classification accuracy. One is using raw waveform-based sample-level deep convolutional neural networks as a feature extractor. The other is multi-level feature extraction and aggregation from the pre-trained networks to tackle various levels of abstractions in MIREX audio classification tasks. Finally, we put them into a classifier based on a support vector machine and make final predictions for each target task. Our submissions were ranked at the first in 6 out of a total of 8 Train/Test tasks of MIREX 2017.