Public articles linked to the same research event.
arXiv The study introduces a framework that uses pitch transposition as an inductive bias to induce ordered orbits via multi-view Sparse Autoencoder (SAE) alignment: it generates pitch-shifted input pairs and aligns their SAE representations to discover structured groups of pitch-related features; experimental results show that this approach recovers orbit structures corresponding to chords, keys, and melodic patterns across two state-of-the-art music foundation models, while requiring only minimal grounding such as a few anchor examples to interpret entire concept families.
The study introduces a framework that uses pitch transposition as an inductive bias to induce ordered orbits via multi-view Sparse Autoencoder (SAE) alignment: it generates pitch-shifted input pairs and aligns their SAE representations to discover structured groups of pitch-related features; experimental results show that this approach recovers orbit structures corresponding to chords, keys, and melodic patterns across two state-of-the-art music foundation models, while requiring only minimal grounding such as a few anchor examples to interpret entire concept families.
The study introduces a framework that uses pitch transposition as an inductive bias to induce ordered orbits via multi-view Sparse Autoencoder (SAE) alignment: it generates pitch-shifted input pairs and aligns their SAE representations to discover structured groups of pitch-related features; experimental results show that this approach recovers orbit structures corresponding to chords, keys, and melodic patterns across two state-of-the-art music foundation models, while requiring only minimal grounding such as a few anchor examples to interpret entire concept families.
The study introduces a framework that uses pitch transposition as an inductive bias to induce ordered orbits via multi-view Sparse Autoencoder (SAE) alignment: it generates pitch-shifted input pairs and aligns their SAE representations to discover structured groups of pitch-related features; experimental results show that this approach recovers orbit structures corresponding to chords, keys, and melodic patterns across two state-of-the-art music foundation models, while requiring only minimal grounding such as a few anchor examples to interpret entire concept families.