The brain detects 3D-shaped fragments (bumps, holes, axes, spheres) in the early stages of object vision – a newly discovered natural intelligence strategy that Johns Hopkins University researchers also found in artificial intelligence networks trained to recognize objects visual.
A new study details how neurons in the V4 zone, the first stage-specific way of looking at the brain object, visualize 3D-shaped fragments, not just the 2D shapes used to study V4 for the past 40 years. Johns Hopkins researchers then identified nearly identical reactions of artificial neurons, in an early stage (layer 3) of AlexNet, an advanced network of computer vision. In both natural and artificial viewing, the early discovery of the 3D form apparently helps to interpret solid, 3D objects in the real world.
“I was surprised to see clear and strong 3D signals since V4,” said Ed Connor, a professor of neuroscience and director of the Zanvyl Krieger Mind / Brain Institute. “But I would never have thought in a million years that they would see the same thing happen to AlexNet, which is only trained to translate 2D pictures into object labels.”
One of the long-standing challenges for artificial intelligence has been the replication of the human vision. Deep (multi-layered) networks like AlexNet have taken great strides in object recognition, based on High Capacity Graphics Processing Units (GPUs) developed for mass gaming and training through the explosion of online images and videos.
Connor and his team applied the same image response tests to natural and artificial neurons and discovered remarkably similar response patterns in V4 and the AlexNet 3 layer. What explains what Connor describes as a “creepy correspondence” between the brain – a product of evolution and lifelong learning – and AlexNet – designed by computer scientists and trained to tag pictures of objects?
AlexNet and similar deep networks were actually created in part based on multi-stage visual networks in the brain, Connor said. He said the close similarities they observed could indicate future opportunities to exploit correlations between natural and artificial intelligence.
“Artificial networks are the most promising current models for understanding the brain. “On the contrary, the brain is the best source of strategies for bringing artificial intelligence closer to natural intelligence.”