Slashdot Mirror
Virtual Robots Fooled By Visual Illusions
Roland Piquepaille alerts us to research out of University College London in which virtual robots, trained to "see" as we do, were duped by optical illusions the same way humans are. Here's one of the illusions the software system fell for.
From TFA: "The virtual robots in this study were driven solely by the statistics of their training history and used these statistics as the basis of their correct and subsequent incorrect decisions. Similarly, we believe the human brain generates perceptions of the world in the same way, by encoding the statistical relationships between images and scenes in our past visual experience and uses this as the basis for behaving usefully and consistently towards the sources of visual images." So the robot vision was created as a model of human vision, and it succeeded at doing so. That's sort of interesting, I suppose, but what does it tell us? That we were right about the way human vision works? Seems to me that the point here is really that in some ways, human vision is 'broken' and that maybe it isn't the best apparatus for machines to use. If we want to welcome our robotic overlords, we should be improving on the vision model, not trying to give machines the same flawed framework.
I think it is important to problem small weaknesses into our robots. You know, just in case.
The program wasn't designed to detect optical illusions -it was a by-product of the training the system went through. The fact that it was tricked by a similar illusion without being programmed to do so might be taken as suggestive that our learning mechanisms are similar to the ones used by the program. From TFA: