Interpretability Analysis of Hierarchical Generative Audio Models (Jukebox)

2022, Yale University

​

Project Overview

This project was my first venture into interpretability research––an early exploratory study of OpenAI’s Jukebox, a hierarchical generative model for raw audio. The central question was: how do hierarchical generative audio models organize internal variables, and how do these variables relate to known musical concepts?

 

To answer this, I performed standard audio interpretability analyses. I instrumented the model to visualize intermediate activations and filters, auralized internal feature maps, and compared activation patterns before and after targeted fine-tuning on small stylistic datasets. These analyses suggested that Jukebox's encoder stabilizes waveform fragments that generally reflect texture classes (mixtures of musical sounds). And these texture classes can be analytically useful to scholars studying the soundscapes of musical artists or objects.

 

This project motivated my eventual dissertation work on architectural constraints and feature stabilization. It revealed that a production-grade generative audio model largely organized its internal structure around texture fragments, but I couldn't find solid evidence that these fragments mapped cleanly onto human-recognizable concepts like pitch (which is only implicitly present in the fragments).