Logo IMG
HOME > PAST ISSUE > Article Detail


Computer Vision and Computer Hallucinations

A peek inside an artificial neural network reveals some pretty freaky images.

Brian Hayes

http://People have an amazing knack for image recognition. We can riffle through a stack of pictures and almost instantly label each one: dog, birthday cake, bicycle, teapot. What we can’t do is explain how we perform this feat. When you see a rose, certain neurons in your brain’s visual cortex light up with activity; a tulip stimulates a different set of cells. What distinguishing features of the two flowers determine this response? Experiments that might answer such questions are hard to carry out in the living brain.

What about studying image recognition in an artificial brain? Computers have lately become quite good at classifying images—so good that expert human classifiers have to work hard to match their performance. Because these computer systems are products of human design, it seems we should be able to say exactly how they work. But no: It turns out computational vision systems are almost as inscrutable as biological ones. They are “deep neural networks,” modeled on structures in the brain, and their expertise is not preprogrammed but rather learned from examples. What they “know” about images is stored in huge tables of numeric coefficients, which defy direct human comprehension.

2015-11HayesF1.jpgClick to Enlarge ImageIn the past year or two, however, neural nets have begun to yield up a few fleeting glimpses of what’s going on inside. One set of clues comes from images specially designed to fool the networks, much as optical illusions fool the biological eye and brain. Another approach runs the neural network in reverse; instead of giving it an image as input and asking for a concept as output, we specify a concept and the network generates a corresponding image. A related technique called deep dreaming burst on the scene last spring following a blog post from Google Research. Deep dreaming transforms and embellishes an image with motifs the network has learned to recognize. A mountaintop becomes a bird’s beak, a button morphs into an eye, landscapes teem with turtle-dogs, fish-lizards, and other chimeric creatures. These fanciful, grotesque images have become an Internet sensation, but they can also serve as a mirror on the computational mind, however weirdly distorted.

comments powered by Disqus


Subscribe to American Scientist