Louis Bradshaw
Curriculum Vitae
/ GitHub
/ Twitter
ku.ca.lumq@wahsdarb.b.l
I'm a CS/ML PhD student at C4DM, where I specialize in Deep Learning for Music. I'm also a research lead at EleutherAI. Prior to my PhD I studied Mathematics, focusing on Algebra and Geometry. My current research interests are varied and include:
- Multimodal Foundation Models for Music: My primary research focus is building deep learning based generative models which are expressive, useful, and artistically interesting. My current research project (see the Aria project), which has attracted generous support from both StabilityAI & EleutherAI, revolves around building a multimodal foundation model, incorporating both audio, e.g., musical recordings, and MIDI (symbolic, note-level information).
- Robustness in Audio-to-Seq Models: I'm interested in improving real-world usability of audio-to-seq models, such as those used for music transcription and speech recognition. I pioneered a semi-supervised approach to training music transcription models (see Aria-AMT), enabling audio-to-sequence models to utilize additional information learned from unsupervised data (e.g., note-level tokens without corresponding audio) when transcribing audio. This approach led to state-of-the-art results on the relevant benchmarks. Going forward, I am interested in applying similar techniques to other seq-to-seq problems.
- Datasets: My research into music transcription began in an effort to increase the amount (and quality) of data available for training symbolic music models. Using our transcription model, we created a dataset of realistic, expressive piano transcriptions, comprising over 100,000 hours of music, representing a 100-fold increase over pre-existing datasets.
Outside of research, I'm extremely interested in the engineering problems surrounding ML/DL. In my non-research time, I currently dedicate a portion to learning C++/CUDA and staying up-to-date with research in ML-infra. Personally speaking, I love mathematics, physics, music, reading, and painting. The best part of doing a PhD is getting to learn from all kinds of people. If you are interested in collaborating, or just chatting about research, feel free to reach out!
Aria Project
I currently run a research project on building, scaling, and aligning transformer models for symbolic music. The project gets its codename, Aria, from the Goldberg Variations, and has attracted generous funding and compute support from both EleutherAI & StabilityAI. Other than building a powerful foundation model and personal artistic projects, my secondary aim is to build a generative compositional tool which is interesting and useful to both novices and experts alike. Although this project is still a work in progress and hasn't been publicly released, here are some early samples showcasing what it can do.
If you are interested finding out more about the Aria project, the best place is on the EleutherAI discord channel.
Misc
These essays [1, 2] and these books [3, 4, 5] had a big influence on me.