Beyond the Creative Species: Making Machines That Make Art and Music by Oliver Bown

Author:Oliver Bown [Bown, Oliver]
Language: eng
Format: epub
Tags: generative; computational creativity; algorithmic automation;
ISBN: 9780262045018
Google: tSAXEAAAQBAJ
Publisher: MIT Press
Published: 2021-07-15T00:32:58.457013+00:00

In another visual art example, Eric Chu proposes an “Artistic Influence GAN” or AIGAN,45 in which the generator of a network trained on a specific style takes as input an additional vector representing influencers, thus producing outputs purporting to be Artist A influenced by Artist B: “What if Banksy had met Jackson Pollock during his formative years, or if David Hockney had missed out on the Tate Gallery’s famous 1960 Picasso exhibition? How would their subsequent art differ?”

Jean-Pierre Briot and François Pachet46 review methods of musical style transfer in musical domains, noting that much of the work in this area focuses on more explicit forms of “structure imposition,” where specific structural qualities are applied from one piece to the “musical texture” of another piece. This includes using the longer-term structure of a piece, as determined by the piece’s self-similarity (a measure of how similar sections of the piece are to other sections of the same piece).

Cross-domain synthesis deals with situations in which content in one domain is used to generate content in another domain, such as generating text from images. This can be done using supervised learning. For example, a similar approach to an autoencoder can be used, this time with the stimulus matched with the expected output. One of the more obvious and common areas of application of this is between text descriptions and media. Machine learning systems have long demonstrated success in generating text descriptions from images, but more recently, GANs have been used to generate images from text descriptions. In this case the generator learns a mapping from a description to an image, and the discriminator learns to discriminate between real and generated images, given the text description.47

Closely related to such text-to-content generation is the idea of generating content that satisfies specific aesthetic or psychological criteria, such as to make happy or sad music. This can be done in a wide variety of ways (for example, Monteith, Martinez, and Ventura use Markov models.)48 Alternatively, we might want to create sensible combinations of elements, such as animated images that go with a piece of music or vice versa.

The Creative Capacity of Machine Learning Systems Big data and deep learning have opened up impressive new avenues for the automated generation of content. Large datasets of written language, speech, music, images, and so on can be fed into these systems to train them on specific styles. In music, you might use one of the many research datasets, such as the Luxembourg folksong database, stored in machine-readable MIDI49 form, commonly used by researchers for experimental purposes. Note that usually the dataset simply dictates what is typical of a given art form, rather than necessarily what is good or bad. We train machine learning systems on Bach or Picasso with the expectation of making things that sound or look like Bach or Picasso. Thus it is perhaps easy to dismiss such systems as merely concerned with imitation and fakery.

What kind of creative potential do we expect machine learning systems to have? Given their

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.