Deep Learning Techniques for Music Generation - Companion Mini Web Site - Recent Additional Content

• What do these 5,599,881 parameters mean? An analysis of a specific LSTM music transcription model, starting with the 70,281 parameters of its softmax layer, Bob L. Sturm.
  Some very interesting paper and experiment about interpreting the role of the parameters (weights and biases) of the architecture and their correlation to the elements of the vocabulary (notes and their durations).
  Some presentation by Estevan Barbará, a student of my course.

• Neural Networks For Music: A Journey Through Its History, Jordi Pons.
  Some very interesting paper analyzing the early stages of use of artificial neural networks for music applications (music generation, but also music retrieval).
  Some personal presentation (Slides 17-39), in my course, showing that experiments by Lewis and by Todd in 1988 were real pioneering in the input manipulation strategy and in hierarchical recurrent architectures.

• How YACHT fed their old music to the machine and got a killer new album and Video presentation (from 5:30 to 23:15).
  Some very interesting description of the precise method defined and followed by the YACHT band to use Google Magenta deep learning architectures and to produce music. Examples of rules are: use only previous band music as training examples; no music element added to the music generated, only extract, collage, transpose, assign and produce. Such a discipline is very useful to better understand what are the respective contributions of deep learning generation and of musicians.
  Some personal presentation (Slides 3-5), in my course.

• Conditional LSTM-GAN for Melody Generation from Lyrics, Yi Yu and Simon Canales.
  Some rare and interesting experiment on the generation of melody from lyrics.

Jean-Pierre.Briot, 20/12/2019.