AI Exhibition in Collaboration with the Japan Graphic Designers Association Inc.

  • Share

ABSTRACT

Between March 15 and 29, 2019, the 78th exhibition of The Japan Graphic Designers Association (JAGDA) took place at the Tokyo Midtown Design Hub. This year, Qosmo collaborated with JAGDA to explore the usage of machine learning to revolutionise the future of graphic design in the first ever “Graphic Design and Co-Creation with AI” installation. 

The event visually showcased the AI generation process and the various results of the research that were conducted, including the creation of alphabetic fonts and musical scores. Additionally, the extensive works published in the JAGDA yearbook were divided into categories by the AI based on certain characteristics, and the fascinating results were presented in an interactive display.

Exhibited Works

“Learn from Modeling – Bauhaus Building Blocks” JAGDA Internet Committee + Qosmo

“Bauhaus Building Blocks” is a collection of animated screens where 3D models of the Bauhaus Building Blocks conceived by German designer Alma Seedhof Busher are combined by the AI into countless variations. The aim was to explore new methods of the creative process and the implications of its meaning in the future.

airline
container ship
wood rabbit

This project was inspired by the short fiction, “The Library of Babel” by Argentine writer Jorge Luis Borges, which is set in a universe which takes the form of an extensive library that contains all possible books of a fixed size, using all the possible combinations of the alphabet. The librarians spend their entire existence searching for meaningful books hidden amongst the majority of the books which are intelligible.

As the AI generates texts, music, and images, it is faced with a similar quest of the librarians, seeking beautiful and compelling content in a series of random combinations.

The question arises: how does the AI understand the concept of beauty and meaning? Unlike other applications of AI such as Go or Chess-playing, we do not have a set of fixed rules to determine the quality of abstract works. As such, currently the AI evaluates a work based on the similarities of the outcome with existing works of art that we as humanity value. Following this approach, the AI can potentially create literature comparable to Shakespear or music that resemble Bach. But can AI express itself in a meaningful yet original and unique way?


This project used what is called a genetic algorithm in order to optimise the way we search for “good” combinations of the building blocks. In this application, it begins with a learned model of AI for image recognition evaluating an initial pool of combinations, each represented by a gene sequence. From this pool, the combinations that achieve the highest similarity score against the desired object (such as a rabbit, airplane, or umbrella) are selected into a more refined pool. Next, the selected genes are crossed over to create the next generation. As this process repeats, we expect to get higher similarity scores and become increasingly closer to the appearance of the target object. Thus far, some similarities can be easily recognised by the human eye – in general, the AI seems to be better with inorganic objects such as airplanes – while others appear more abstract. Nonetheless, the AI has surpassed our expectations, providing thought-provoking materials that stimulate the human imagination.

“Composing Melodies Inspired by Musical Scores ” Qosmo

Inspired by a fantastical image in which an enchanter performs a spell to make musical scores emerge on an old parchment, we created a hypnotic display in which scores continuously appear over another by using a Generative Adversarial Network (GAN).

GAN is a class of machine learning systems in which two neural networks compete against each other. Given an initial set of elements, the first network, the generator, attempts to generate new elements, while the second network, the discriminator, must determine whether it is authentic or not. In this process, the generator learns to fool the discriminator to produce false positives. Recently, it has attracted much attention as this method was used to generate hyper-realistic human faces. Here, we used images of musical scores for the initial set, and the AI generated new scores that visually reflected the characteristics of the original score despite it being musically strange at times.

We continue to pursue the goal of generating music from its visual representation.

“The JAGDA Experiment – Generating Font” JAGDA Internet Committee + Qosmo

Through the evolutionary model utlised by GAN, a gradually changing font is generated. The audience can witness intriguing moments when a strange and unfamiliar font appears between fonts that directly reference the original font and its characteristics.

CVC-MUSCIMA dataset : Alicia Fornés, Anjan Dutta, Albert Gordo, Josep Lladós. CVC-MUSCIMA: A Ground-truth of Handwritten Music Score Images for Writer Identification and Staff Removal. International Journal on Document Analysis and Recognition, Volume 15, Issue 3, pp 243-251, 2012. (DOI: 10.1007/s10032-011-0168-2).

“The JAGDA Experiment – Yearbook Publication Matrix” JAGDA Internet Committee + Qosmo

In this installation, the AI was trained to classify images of the JAGDA yearbook, “Graphic Design in Japan”, which contained over 1,500 works including posters and logos made in the five year span between 2013 and 2017. The AI is able to perform this based on five different criteria, from simpler classifications based on colour and shape, to the more complex based on the contents. This allowed the archive to be displayed in various interactive forms.


In order to show how the AI internally analyses the graphics from different angles, we took the high-dimensional features provided by the neural network, VGG16, and we displayed them in 2D and 3D formats on the screens.

As we’ve used a general model for image recognition that is not able to recognise text, it cannot understand the meaning and the context behind the works. Because of this, the project was able to reveal unexpected similarities between works that normally go unnoticed by humans as we are highly influenced by context.




LINKS


CREDITS

“Learn from Modeling – Bauhaus Building Blocks” JAGDA Internet Committee + Qosmo

  • Planning / Direction

    JAGDA Internet Committee

  • Planning / Technical Direction

    Qosmo

  • Planning / Supervision

    Nao Tokui(Qosmo)

  • Poster Design

    Robin Jungers(Qosmo), Junichiro Horikawa

  • Animation

    Shoya Dozono(Qosmo)

  • Project Management

    Sakiko Yasue(Qosmo)

  • Support

    Junichiro Horikawa

"Composing Melodies Inspired by Musical Scores " Qosmo

  • Planning / Direction

    Qosmo

  • Programming

    Nao Tokui(Qosmo)

  • Assistant

    Makoto Amano(Qosmo)

  • Project Management

    Sakiko Yasue(Qosmo)

  • Sound Support

    Akiko Oishi

  • Installation Design

    JAGDA Internet Committee

  • Instrument Support

    Yamaha Music Japan Co.,Ltd.

“The JAGDA Experiment – Generating Font” JAGDA Internet Committee + Qosmo

  • Planning / Direction

    JAGDA Internet Committee

  • Programming

    Nao Tokui(Qosmo)

"The JAGDA Experiment – Yearbook Publication Matrix" JAGDA Internet Committee + Qosmo

  • Planning / Direction

    JAGDA Internet Committee

  • Planning / Technical Direction

    Qosmo

  • Technical Direction

    Nao Tokui(Qosmo)

  • Data Analysis / Web Programming

    Robin Jungers(Qosmo)

  • Project Management

    Sakiko Yasue(Qosmo)

Get in touch with us here!

CONTACT