Re: AI: Image Recognition Finding

Nick Bostrom (bostrom@ndirect.co.uk)
Sun, 16 Aug 1998 00:57:34 +0000

Doug Bailey wrote:

> Got this off of Comline today:
>
> A team from NEC {6701} and the University of Tokyo have
> discovered that applying noise to image recognition patterns
> through such artificial intelligence (AI) systems as neural
> networks will raise the accuracy of recognition systems.
> This runs counter to the prevailing belief that an image
> must be a neat as possible for clear pickup.

Actually, it has been know in the neural networks community a long time that adding some noice often improves performance, especially when you have a small number of learning samples. The noice prevents the network from "overlearning", i.e. learning to recognize each sample pattern individually. It forces the network to discover the general regularities in the sample set.

A clear picture is still better than a blurry one. You can always add noice afterwards if you like, but you can't go in the opposite direction.



Nick Bostrom
Department of Philosophy, Logic and Scientific Method London School of Economics
n.bostrom@lse.ac.uk
http://www.hedweb.com/nickb