Dedd Pixel

Earlier this month, The Atlantic featured an article about an art exhibit in Chelsea, Faceless Portraits Transcending Time, that featured artificial intelligence (AI) generated art.  Described as a “collaboration between an artificial intelligence named AICAN and its creator, Dr. Ahmed Elgammal,” the exhibit features somewhat abstract portraits of human-like figures.  Images that are suggestive of classical portrait art with a touch of disturbing thrown in. That touch of disturbing is likely the result of Elgammal’s “creative adversarial network,” or CAN, which replaces the typical discerner element in a generative adversarial network (GAN) machine-learning method with a more creative element that introduces more novel outcomes.  So instead of analyzing images and trying to produce an image that is the same as those images, Elgammal’s application tries to make creative variants on the imagery.

This brings AI and computing a step closer to that human ideal.  As quoted by Jeff Clune in the April 2017 article in the Communications of the ACM, Computing the Arts, machine-learning produced art is desirable

“because we consider artistic expression as one of the most uniquely human traits.  If we want to produce artificial intelligence that rivals human intelligence, that should include art.”

Elgammal’s process begins to address one of my impressions with AI generated art in that merely analyzing existing images and trying to produce another image that can be construed as belonging to the original set is not much more than copying the style of existing art rather than creating something new.  There’s nothing wrong with that, of course, but how far can close approximations to existing art alone push the dialogue? We’re only talking about this art now, because it was generated by a machine. If a person endeavored to create close approximations to existing art, many would ignore it. The AI art created by the Obvious Collective at Christie’s last year was certainly more pleasant to look at than Elgammal’s but if a person today painted what appeared to be an ‘unfinished’ portrait of a possibly French clergyman, I’m not sure who would care.

In the Atlantic article, art historian John Sharp discusses how traditional portraiture included objects imbued with symbolism along with the human subjects.  The objects gave cultural context to the human figures.

“For example, men might be shown with an open book to show how they are in dialogue with that material; or a writing implement, to suggest authority; or a weapon, to evince power.”

I am certain that AI portrait artists could be trained on the cultural meanings of objects and colors across geography and time to create art that is meaningful to its target audience.  Once the machine gets good enough at generating a human portrait, it could also be set to include objects with certain cultural meanings and thus invoke reactions and feelings in the art viewers as they observe and make judgement on the subjects in the painting.

The machine could also be set to invoke a random set of symbolism.  Perhaps a portrait of a young man, wrapped in Virgin Mary blue, holding a writing implement and an apple.  People will feel compelled to ascribe meaning to the colors and objects depicted with the subject and use those symbols to guess the origin or history of him. However, the machine is devoid of any intent.

On Obvious’ webpage, they refute Picasso’s complaint about computers:

“Computers are useless. They can only give answers.” 

I still think it’s a good point. While a machine could be set to paint a portrait that depicts a person with objects that represent certain symbolic meaning, could it ever come up with novel concepts or generate art to express an idea that was not guided by operator intent or randomness?

Can an AI artist pose a novel question expressed through art?