Refined Tastes and Neural Nets

A Short Review of Korsak Ovbot: Exploring Artificial Creativity through Music Composition

One of my friends who is very interested in machine learning and who I admire very much recently wrote an article titled “Korsak Ovbot: Exploring Artificial Creativity through Music Composition” and was published in The Harvard Undergraduate Research Journal. The paper describes a way for neural networks to synthesize “creativity” as it is defined. I found the paper to be incredibly interesting and groundbreaking in its own right.

In 2015, Google’s DeepDream was successful in using deep neural networks to generate psychedelic images from static or other unrelated images. This paper claims that human artwork is created in a similar fashion. Creativity is defined to be an iterative process of modifying one’s output through experimentation and slight adjustments until the artist is satisfied with the product. DeepDream uses a similar process, however, the paper claims that humans “seem to… improve their medium to maximize an aesthetic” rather than simply changing the medium to generate images as DeepDream does. To combat this, George Moe’s paper describes a Deep Neural Network, named Korsak Ovbot, that/who is able to differentiate between what it likes (aesthetic inputs) and what it doesn’t (random noise). Based on this training, the algorithm then generates content that it likes (by greedily maximizing the aesthetic value of the output medium). This process is then iterated (through evolution).

The key idea is that instead of directly coding for acceptable and unacceptable randomly generated phrases of music, Korsak is trained on annotated music that lets Korsak decide what it finds to be “good” music. Korask is then able to compose music to that standard. Note that the training music is still composed by humans, which will slightly restrict the tastes of computer generated music. However, since the music is randomly generated (just selected out), there is still much freedom for Korsak’s creative potential. Furthermore, human tastes and styles are similarly developed and cultivated through sampling of music (perhaps to a lesser degree), not only inherent.

In the future, I think that it will be very interesting if Korsak can learn to differentiate between music of different moods or portraying different feelings. I think that it would be very impressive if Korsak is able to suggest songs not based off of clicks or genre/album similarity but of mood similarity. Perhaps then, music therapy or spotify suggestions based on moods and feels will be possible.

Works Cited

Moe, G. (2018). Korsak Ovbot: Exploring Artificial Creativity through Music Composition. The Harvard Undergraduate Research Journal, 11(2), 11–17.