20170220

LEARNING FEATURES OF MUSIC FROM SCRATCH https://arxiv.org/pdf/1611.09827.pdf

• getting started with music net (their dataset) http://homes.cs.washington.edu/~thickstn/start.html

An Experimental Analysis of the Entanglement Problem in Neural-Network-based Music Transcription Systems

• https://arxiv.org/pdf/1702.00025.pdf
• glass ceiling effect discovered in music transcription using neural networks
• makes me think that labelling with chord types eg. (minor 5th) rather than exact note values

research pipeline:

1. investigate classififyers for music

• how good
• current problems
• what they are classifying
  • intervals
  • notes
  • composers
  • styles
  • etc.

2. investigate GANs for music generation

3. use GAN for music generation to create classifyer

4. use classifyer to learn features for music style transfer

to do for next monday:

1. YG investigate the classifier and dataset in "learning features of music from scratch"

• apply to other songs see error rates
• see if we can improve architecture and get better results
• use the weights already trained as a starting point for training on other songs

2. RG investigate GANs for music generation