DONGRYUL LEE


ABOUT     CATALOG      EVENTS      VIDEO      COMPUTER MUSIC      RESEARCH      SOUNDCLOUD     PRESS      POPULAR MUSIC      CONTACT


[Work in progress] On a Winter’s Night a Traveller, The Deeply Learned Clavier (2022–23)
for Solo Piano
피아노를 위한, “겨울 밤의 여행자에 대하여"

Duration: ca. 25 minutes

Program notes
As you can see here on my website, I recently replaced many images on my homepage with artificially synthesized images created by a Deep Learning application. Almost about a half of images that you see here are generated by the “discord.” The speed of development of AI and Deep Learning algorithms are truly remarkable, and the artists of the 21st century may have to compete with AI machines. Rather than compete with the technology, I decided to accept machines as my creative partners.

In 2022, I embarked on a new chapter of my research on Computer Music, in the field of Artificial Intelligence and Deep Learning. I began attending the MIT OCW’s Deep Learning course, as I did in 2010 to study differential equations which significantly contributed to my DMA research. For this, I began studying and analyzing Python and Keras programming languages and codes on the Google Colab platform to implement my own Deep Learning model. For about two months, I played with the code and could make the computer learn about 110 piano pieces written by Mozart. Then, after between one and six hours of each learning process (ca. 20-200 epochs in deep learning), the computer wrote a new piano work within its understanding of the music of Mozart. In other words, the computer wrote a pastiche piece in the style of Mozart, or even tried to counterfeit the works by Mozart.


Deep Learning generated Output 020, sequence size 25, temperature 2, 74 epochs of learning


Deep Learning generated Output 062-2, 109 learned works, sequence size 256, weight: 0.085, learning rate: 0.001, optimization: RMSProp, 200 epochs of learning, temperature 3.7


Deep Learning generated Output 063-6, 109 learned works, sequence size 256, weight: 0.085, learning rate: 0.001, optimization: RMSProp, 30 epochs of learning, temperature 3.9, seed piece: Fantasy in c minor KV366


Original handwritten score of On a Winter’s Night a Traveller, The Deeply Learned Clavier

Because of the limit of my access to the Google Colab platform, I couldn’t spend enough time on the deep-learning process of more than six hours for each output, and also the number of model data-works was not enough to generate a real Mozart-like piano work, e.g., having an Alberti bassline or crisp and concise phrase models which we can easily find in his oeuvre. However, those outputs/works generated by the computer implied various interesting features that we can often find in the tonal music repertoires, including ambiguous tonalities (in-between different keys), diatonic scales, repetitions, chords and arpeggios, oftentimes realized in awkward, crude and rudimentary, anti-contrapuntal and artificial, or even distorted and mechanic ways.

For about a month I listened to the outputs generated by the computer, and began tracing the trajectories of the computer’s thoughts. I was almost an archeologist who discovered a pile of unfinished, improvisatory, spurious sketches written by an unknown composer or had been misattributed to Mozart(‘s virtual clumsy pupil, if more accurately described.) During the discovery process, I tried to reveal the computer’s intentions, develop his/her imagination, and transformed it into my own artwork.

This project is currently a work-in-progress. Although I tried to complete it by January 2023, because of my other workload, I hope to finish the piece by the Spring or Summer of 2023. It is commissioned by the Boston-based pianist Ariel Mo, who beautifully world-premiered my Le Tombeau de Harvey in January 2022.


AI-created concept image for the piece

About the title:
Although I can't logically explain the exact relation between this process with Calvino's If on a winter's night a traveler, I was thinking that Calvino's novel is made up of two/three separate worlds, one with the protagonist (dimension α) and the other with those "inner book stories" (dimension β), which are built upon multiple episodes, written by some unproven or spurious authors (dimension γ). For my creative process, my existence could be the protagonist part, and each different machine-generated music could be regarded as a separate world/story. Or this could be interpreted oppositely: those machine-generated pieces could be the continuous story since it is really an "evolving" story, which is developed by learning the music of Mozart, and my interpolations or (hu)man-made parts could be regarded as short episodes, which I wrote in accordance with those machine made stories. In addition to these elements, there is another dimension in the original novel, the existence of the authors of the short episodes. (either Cimmerian or Cimbrian, and many others) So, for this time, Mozart could be one of those almost forgotten (true/fake) authors. As in the book, we don't know if these authors are genuine, valid, and existed. They could be dubious, misattributed, or fakes, as this deep-learned Mozart can't be real and prove if a real Mozart’s ideas can be as such.

Mozart – an “ideal” author (dimension δ)
Deeply Learned Computer – spurious authors (dimension γ)
Machine-generated pieces – nested episodes created by the spurious authors (dimension β)
Myself – The Protagonist and his world (dimension α)
or
Myself – a spurious author (dimension γ)
My creations – nested episodes (dimension β)
Machine-generated pieces – The Protagonist and the real world (dimension α)
optional [Original Mozart scores – The Platonic ideal (dimension α’-prime)]
...(Other interpretations possible)

However, please don't expect too much of Wolfgang Amadé, since because of my lack of programming skills and the limited amount of GPU processing time, there is really no Mozart in the piece you will hear...