September 14, 2026 at 3:10 PM
Humans can acquire abstract, reusable concepts from limited experience and extend them to unfamiliar cases. Modern deep learning systems, by contrast, often learn comparable abstractions from large labeled datasets or text corpora, in which the relevant concepts have already been named and curated by humans. This thesis investigates whether machines can learn such concepts without explicit supervision, a problem we refer to as self-supervised concept learning. Our central methodology is to define an initial concept prior and refine it through self-supervised learning, enabling models to generalize beyond prototypical cases.
We use music as a testbed for studying self-supervised concept learning. First, we show that carefully designed task-specific priors can support the self-supervised learning of syntactic and semantic concepts in music. We then introduce a more general framework, music-for-music modeling, which grounds abstract concept priors directly in the music modality and provides a unified approach to modeling diverse musical concepts. Across both symbolic and audio settings, our models learn high-level musical concepts from unlabeled data and apply them to music analysis and controllable generation. They match or even surpass supervised baselines across multiple analysis and generation tasks, demonstrating the effectiveness of the proposed training paradigm.
The seminar will take place at https://visio.numerique.gouv.fr/alg-omus-ada.
Online
français
English