NN trained on Zelda music (Overfit example)
NN trained on Zelda music (Overfit example)
Recurrent neural network trained on 15 Zelda songs.
The Experiment
This project used a recurrent neural network (RNN) to learn from and generate music in the style of The Legend of Zelda video game series.
The Overfitting Problem
What Happened
The sample size was too small and the network became overfit.
What is Overfitting?
Overfitting occurs when a neural network:
- Memorizes the training data instead of learning patterns
- Performs well on training data but poorly on new inputs
- Essentially "remembers" rather than "learns"
- 15 songs is not enough training data for music generation
- Overfitting is a real challenge in creative ML applications
- Finding the right amount of training is crucial
- Sometimes the "wrong" result is still interesting
The Result
The network learned to reproduce the training songs almost exactly rather than generating truly original compositions. But hey, I can make a really long Zelda song if I want!
Other Outputs
The other example songs generated were... more original... but less nice sounding. This is the trade-off with neural network music generation - too much training leads to copying, too little leads to chaos.
Lessons Learned
Technical Basis
Based on the biaxial-rnn-music-composition project:
https://github.com/hexahedria/biaxial-rnn-music-composition