[Tyler]

Even computers are fooled by optical illusions!
18 December 2014, 12:00 am

The SpaceX founder, along with other scientists such Stephen Hawking, are concerned by the rapid pace of progress in machine intelligence. But computers may not be as clever as we believe. Daily Mail - Sciencetech (UK)Link.

Source: AI in the News

To visit any links mentioned please view the original article, the link is at the top of this post.

[Korrelan]

Cool article.

I think this a problem because they are using simple static pattern matching.  Deep Nets use the values from grids of variously spaced pixels across the image to initially glean the contents of the image.  The values of an initial wide spaced matrix will guide the Net along a path of recognition incorrectly. So somewhere in each image the Net is picking up a average/ learned pattern of pixels incorrectly. 

This is much less of a problem once the AGI has a sense of orientation, location, time, occasion, who it's dealing with, etc.  Also... optical flow counteracts a lot of these problems.  A similar problem arises with audio illusions when an AGI is learning to recognize verbal patterns with a lot of loud background noise.  Until the patterns a sufficiently reinforced in a silent environment it doesn't stand a chance. 

I use optical illusions when testing my AGI, very useful to test it's imagination/ attention. 



We would see a triangle because the sparse optical input force reminds us of a triangle.  It's not only the visual element but the visual memory that combine to trigger recognition.  There are enough visual cues to focus it's attention, and the AGI can't help but see a triangle, even though it was told otherwise.  It's surprising how many objects look like something totally different with out s sense of context.