Keunwoo Choi, George Fazekas, Mark Sandler and I have received the Best Paper Award at the 18th Annual International Society for Music Information Retrieval Conference (ISMIR). The paper is <Transfer Learning for Music Classification and Regression Tasks> which investigates different ways to exploit the knowledge captured within a deep convolutional network trained to tag a song for other relevant tasks. The main idea is to use not only the final hidden activation vector (as has been usual in computer vision) but to use the activations from all the layers, as some target tasks may require low-level details. Check it out at https://arxiv.org/abs/1703.09179 and https://github.com/keunwoochoi/transfer_learning_music.
This work started while Keunwoo was visiting me at the NYU Center for Data Science. And, of course, he’s done all the work from the beginning to the end. Congratulations and thank you, Keunwoo!
