# An Interactive Musical Prediction System with MDRNNs Charles Martin - University of Oslo / Australian National University !(/assets/logo-bar.png)
# What is this?
## Learning to Predict Sequences !(/assets/sequence-learning.png)
## Interacting with Musical Predictions !(/assets/imps/predictive-interaction-motivation-hires.png)
## Why is this needed? | Creative Deep Learning Systems | NIMEs | |--------------------------------------------|-----------------------------------------------| | Focus on MIDI data (e.g., Magenta Studio) | Yes MIDI, but also many custom sensors | | Focus on digital audio | Focus on performer gestures | | Focus on composition/artefact generation | Focus on interaction | | Rhythm on 16th note grid | Complex or no rhythm | | Focus on categorical data | Continuous data more interesting | !(/assets/imps/imps-nimes-examples.jpg)
## IMPS: Interactive Musial Prediction System !(/assets/imps/demonic_imps.jpg) - An opinionated deep learning model for NIMEs - An environment for making NIMEs that play themselves - "Wekinator" for deep learning?
# How does it work?
## Mixture Density Recurrent Neural Network !(/assets/imps/mdn-motivation.png)
## Mixture Density RNN !(/assets/imps/mdn-examples.jpg) Good at predicting creative, continuous, multi-dimensional data: handwriting, sketches... musical gestures?
## What to do with predictions? !(/assets/imps/predictive-interactions.png) 1. Call-and-Response: Continue gestures when performer stops 2. Layered predictions: Always predict next move from current gesture 3. Duet: Two interdependent processes
# Ok, how do I _use_ it?
## Three easy steps...
Collect some data:
IMPS logs interactions automatically to build up a dataset
Train an MDRNN:
IMPS includes good presets, no need to train for days/weeks
IMPS includes three interaction modes, scope to extend in future!
# Is this even _practical_?
## Deep Learning in NIMEs?? - Is it practical for real-time use? - How do the MDRNN parameters affect time per prediction? - What are "good defaults" for training parameters? - Do you need a powerful/expensive computer?
## Test Systems ![Test computers](/assets/imps/imps-test-computers.jpg)
## Results: Time per prediction ![Time per prediction vs LSTM units](/assets/imps/prediction_time_against_units.png) Time per prediction (ms) with different sizes of LSTM layers.
## Results: Time per prediction ![Time per prediction vs MDN dimension](/assets/imps/prediction_time_against_dimension_64units.png) Time per prediction (ms) with different MDN output dimensions. (64 LSTM units)
## Results: Training Error vs Validation Set Error !(/assets/imps/training_loss_12K_lightpad_model.png) 12K sample dataset (15 minutes of performance) Takeaway: **Smallest model best for small datasets.** Don't bother training for too long.
## Results: Training Error vs Validation Set Error !(/assets/imps/training_loss_100K_lightpad_model.png) 100K sample dataset (15 minutes of performance) Takeaway: **64- and 128-unit model still best!**
Results: Exploring Generation
Takeaway: Make Gaussians
, make categorical
## Try it out!
- Available on [GitHub](https://github.com/cpmpercussion/imps) - Try with your NIMEs! - Hack if you want! - Add an issue with problems/results! Twitter: [
](https://www.twitter.com/cpmpercussion) Website: [creativeprediction.xyz/imps](https://creativeprediction.xyz/imps)