Of Computers and Visual Perception

Posted by Simon Raggett on 4 March 2010 | 1 Comments

Tags: computers, artificial intelligence, robotics, visual perception, brains, quantum computing

The vision researcher, Steven Lehar, challenges the computer/artificial intelligence view of the brain. He describes the problems that computers have with visual perception. Computers can detect edges in objects, and this is accepted as being one of the first steps in processing visual input into the brain. However, computers have difficulty in turning this data into useful information, because they detect too many features indiscriminately. They do not just detect relevant edges, but also much less important data referring to textures etc., without the ability of biological vision to determine the relative importance of different edges.

The three-dimensional processing of spatial structure is also argued to be a problem for computers, as evidenced by the difficulty robots have in navigating an irregular environment. Lehar traces this to the two-dimensional retinal image, with the position of objects in spatial structure being inserted by cortical processing that uses a spatial algorithm. To date, computers appear to lack this algorithm.

Famous Dalmatian and other examples:  Lehar discusses some well-known images that for him serve to demonstrate that biological vision cannot derive from the bottom-up processing of edges. His first example is the image of a Dalmatian (spotted) dog against a speckled background, which some of us saw in a recent SMN lecture. Much of the edges of the dog are missing, so local information does not allow the observer to distinguish the dog from its background. Nevertheless, when the picture is viewed as a whole, the dog is clearly distinguishable. It is argued that this indicates that perception is based on global brain activity, rather than the sequential process of individual neurons analysing individual edges.

Other examples support the Dalmatian dog argument. With the Kaniza triangle, there are only three black marks printed on a piece of paper, but the brain distinguishes a triangle that does not actually exist in terms of the analysis of edges on the paper. Another example is 'invariant perception', or the ability to identify an object as the same thing in different lights or from different angles. In terms of bottom-up analysis of edges the objects are quite different, but the brain identifies them as the same object. In support of this position, it is pointed out that where the view of a picture is restricted to a few edges, human observers cannot distinguish between edges that are important to the outline of an object, and edges that are just texture. These examples as a whole are claimed to demonstrate that a form of top-down processing underlies human visual processing.

There are a number of suggestions as to how the brain might achieve this analysis. My own conclusion is that the huge number of possible solutions that have to be explored by a top-down search engine look to be well beyond the capacity of any classical computer, and this seems suggestive of a level of quantum computing in the brain.