To Err Is Divine: Deep Learning and the Art of Machine Mistakes

by Mark Thompson, October 7, 2016

As we teach AI to visually recognize things it “sees” we get some interesting imagery. In the example below, a neural net exaggerates the variation in each “version” of the picture of reality it renders, creating images that have evolved to become unrecognizable to humans. So minor variations in input appear to lead to some very wrong conclusions (Nguyen, Yosinski, Clune 2015).


Image courtesy of Evolving AI Lab

Similarly, our nervous systems make many edited versions of reality before an image is finally presented to our conscious mind. Visual information is compressed as it travels from the eye to the visual cortex (Itti L, et al, 2015). When scientists look into hallucinations, they believe it is based on something called “ectopic” vision. This is similar to floaters or swirls that any of us might see in our vision at one time or another. Scientists believe hallucinations occur when the brain receives “ectopic” input and tries to make sense of it, forming things that may not be there.*


Image courtesy of Evolving AI Lab

Do the “hallucinations” of deep neural networks mirror patterns in the development of the human brain? And did hallucinations play a major role in the development of abstract symbolism and interior life?

Archeologists agree that something new happened to the human consciousness roughly 40,000 years ago. In cave art around the world, we see the emergence of symbolic art and fantastical half-human/half-animal creatures. Interestingly, some experts claim these creatures and drawings are stylistically identical to what modern day psychedelic users have drawn. Anyone who has seen Ayahuasca art from the Amazon will agree that it is a very distinctive style. It’s an astonishing notion, that caves around the world depict art that may be some of humanity’s earliest records of altered states or hallucinations.

Graham Hancock makes the case that humanity took a major leap forward in symbolic thinking and reasoning with this artistic development. What is interesting about his case is that he views it from inside the civilizations who had the experiences. The general claim of people, hunter/gatherer or modern, who have depicted these strange creatures in cave art is that they are sentient beings — they anticipated their human visitor and they have a message to impart. Shamans in the Amazon support this notion. So do modern subjects in controlled experiments with DMT, the psychoactive compound found in many hallucinogens, what some have termed the spirit molecule. As far-fetched as this view may seem, can we completely rule out that humanity may have been “lifted up” by an outside consciousness? To deny this possibility would be to fall into a form of subtle reductionism.

All of this begs the question, do robot dreams mark the beginning of stirrings in interior consciousness? What more is consciousness than an elaborately complex network of internal and external impressions? And is an inaccurate view of reality a necessary part of discovery?

*An interesting aside: An Australian scientist was blinded in an accident and told by doctors to give up “seeing,” as lacking real input, the images from his optic nerves would drive him crazy. He refused and found that though he is legally blind, he has been able to roof his own house and can, for example, watch tennis on TV. He still makes facial expressions when speaking because his visual cortex fills in a picture of what he thinks is there.

To learn how Dialog can help your business, contact us at 512.697.9425 or

This article was originally published in design4emergence, a network science community sponsored and nurtured by Dialog and Panarchy, the world’s first network design firm.



Nguyen A, Yosinski J, Clune J. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images,” Computer Vision and Pattern Recognition, CVPR ’15, IEEE (2015).

Itti L, Rees G, Tsotsos J, Anderson CH, Van Essen DC, Olshausen BA. “Directed visual attention and the dynamic control of information flow.” Neurobiology of Attention (2015).

Comments are closed.