← All student projects · Living with intelligent instruments

Updating small-data music models across practice sessions

How should a musician's personal AI model be updated as they collect more data over weeks of practice?

Level: Summer research, Honours, Masters
Prerequisites: Python and some machine learning (e.g., COMP3670 or COMP4670). Experience with Keras/TensorFlow is helpful.

IMPSY models are trained on tiny datasets: a few hours of one person’s playing. In our recent experiments, a model trained on 7.4 hours of controller data from seven performances started to overfit after about 26 epochs. Musicians collect more data every time they play, so what’s the best way to update a model? Retrain from scratch? Fine-tune the old model? How much new data does it take to change how the model behaves?

In this project you’ll treat existing performance logs as a sequence of practice sessions, then compare strategies for updating an MDRNN after each one. A key part of the project is defining how to tell when a model has actually changed its musical behaviour, not just its validation loss (for example, by comparing the timing and values of gestures sampled from each model).

• Summer or one semester: a reproducible experiment pipeline using IMPSY’s dataset tools; retraining vs. fine-tuning on one instrument’s data; validation loss plus one behavioural measure.
• Two semesters or Master: add regularisation strategies and a second dataset (e.g., logs from the multi-week study below or other lab instruments), and recommend a procedure for updating models between sessions.

Tags: #machine-learning #small-data #music #IMPSY

How to apply

Contact Charles Martin with your CV, your unofficial transcript (if you are an ANU student), and a brief statement (200 words) explaining how you would approach this project. Before applying, read the Join page and our project expectations.