Creative Prediction with Neural Networks

Supervisors

Charles Martin. Lecturer in Computer Science, Australian National University. [email protected]

Benedikte Wallace. PhD Researcher, University of Oslo. [email protected]

Creative Predictions

Learning to Predict Sequences

Melody to Harmony in MicroJam

1. Gain an overview of DL for music generation.
2. Develop a melody to harmony sequence to sequence model
3. Train the model on matched melody/harmony sequences
4. Use MicroJam-sourced data as input and see if the generated harmonies make sense!

Seq-to-Seq Music Generation

1. Understand the Transformer architecture.
2. Implement your own Transformer (e.g., in Keras).
3. Find a musical dataset that could be trained.
4. Train your model, listen to the results and find a way to evaluate them.

Generating colour palettes from audio data

1. Gain an overview of DL for audio processing.
2. Obtain a dataset of audio and video (or colour) data.
3. Try different neural network designs and evaluate the results. (Even a simple fully-connected ANN might work well!)

Motion-to-Motion Generators

1. Gain an overview of the main DL methods used for motion generation including RNNs, MDRNNs, and world models.
2. Find a dataset of motion capture or other movement data (or capture one yourself!)
3. Train the ANN and evaluate its generative abilities.