Computers are still far away from matching our ability to visually recognise objects

Computers have problems identifying objects in varying situations, while the brain does it easily. Photo courtesy of Photocase.com.

The progress in computer technology within the past decades has been absolutely astounding. The capabilities of our personal computers have been growing exponentially (see Moore’s law) and the computer can perform calculations much faster than the human brain. However, in many other areas the computer is no match for the human brain. Especially the function of the visual system in humans is impressive. Several computer programs have been made that can identify and recognise visual inputs and although they are useful, for instance, in recognising characters and numbers on invoices and documents, human verification is often needed, especially, when variations are introduced in for instance hand written documents. Scientists from MIT in the United States in America have looked at a related problem. That of visually recognising objects viewed from different angles with variations in background and lighting. Such a task has proved very difficult for the computer while the brain solves it effortless.

Building computer models that can solve this problem is not only important for the computer industry, but can also teach us more about how the brain works. Neuro- and computer-scientists have recently developed a standard ‘natural image test’, where advanced computer models are tested on their ability to recognise a large variety of objects from different categories (images containing planes, cars, faces etc.). Recently complicated biologically inspired computer models have been showing better and better performance when using this test. However, the MIT scientists argue that the ‘natural image test’ is not accurate enough. They show that a very simple model based on matching single cell (pixel) input to a stored template performs better in the ‘natural image test’ than the complicated biologically inspired computer models. They argue that this is because the ‘natural image test’ does not properly address the difficulties of varying backgrounds and viewing angles.

Their results show that the current testing methods are too simple and that testing for and incorporation of the natural variation is critical if significant progress is to be made in understanding and emulating our brains’ impressive ability to visually recognise objects in even very different contexts.

Source
Pinto N, Cox DD, DiCarlo JJ (2008) Why is real-world visual object recognition hard? PLoS Computational Biology 4(1): e27. doi:10.1371/journal.pcbi.0040027

Tag it:

?

Be first to comment this article

Only registered users can write comments.
Please login or register.

Powered by AkoComment Tweaked Special Edition