Every month, the joint laboratory invites external speakers to take part in seminars for its partners.
Shlomo Dubnov (University of California, San Diego, USA): « Uniting multi-source control with Multi-track models «
Abstract:
Musical Multi-track foundation models capture joint statistics of simultaneous voices. This is contrasted with multi-source models that extend prompting and continuation by temporal and semantic controls.
This talk proposes an information-theoretic framework that distinguishes joint modeling from directional influence, using concepts from causality, Directed Information, and Music Information Dynamics. We show how this perspective clarifies the limits of current control mechanisms in generative models and motivates new approaches to human–machine musical interaction and machine improvisation.
Bio:
Shlomo Dubnovs is a Professor in Music and CSE departments in the University of California, San Diego. He is a founding faculty of the Halıcıoğlu Data Science Institute and served as a Director of the Center for Research in Entertainment and Learning (CREL) at UC San Diego’s Qualcomm Institute
