Understanding the human visual system: the PARIETAL project team publishes new results
The PARIETAL project team, in partnership with Telecom ParisTech has published a paper entitled: “Seeing it all: Convolutional network layers map the function of the human visual system”, which reveals that models used by computer for image recognition have much in common with the human brain.
Studies of human vision and research into computer vision are carried out in two separate, yet mutually aware, worlds. In the PARIETAL project team, Michael Eickenberg, Gaël Varoquaux and Bertrand Thirion, in partnership with Alexandre Gramfort (Telecom ParisTech), studied the relationship between the human brain and artificial-intelligence models that solve the same image interpretation problems on a computer. Research has shown that computer vision, which is artificially created, shares certain properties with the human brain.
Gaël Varoquaux explains: " We have shown that the best models developed so far in the field of image recognition by computers have a similar architecture the human brain, and that these models can be used to map the human visual system more and more effectively .”
Historically, there has always been much back and worth between studies of the human brain and those in artificial intelligence, since both objects solve the same problem and are guided by the nature of images. The results obtained by the PARIETAL team show that the more progress is made in research, the closer the resemblance between artificial intelligence technology and the mechanisms of the brain with regard to image recognition.
“We do know that artificial image recognition models lack some of the mechanisms found in the brain, in particular when it comes to understanding ambiguous images. These models are not intended to be carbon copies of the brain and, when used to study the brain, they are only approximations. We have nevertheless shown that they are good approximations , because the results capture many aspects of the brain and generalise throughout many experiments,” Gaël Varoquaux concludes.