Image recognition
This article supports preexisting research in that it sugests that face recognition is a flexible ability based on frequency of observation. This is entirely in line with studies done in the 1980s on the somatosensory cortex of monkeys, where the area of cortex responsible for processing tactile inputs from various parts of the body is proportional to the usage of that part of the body. The interesting part of this research was that it was shown that the somatosensory cortex does not remain static but is instead a dynamic competitive system. For example, if you use one finger more than any other the area of your brain dedicated to picking up signals from that finger becomes relatively larger, and conversely if that finger were to unfortunately be chopped off then the area of brain dedicated to it shrinks as surrounding areas colonise the unused space.
This article would suggest that what's happening in the visual domain is probably no different from what happens with tactile senses. You can model this sort of system using kohonen networks, or alternatively as more elaborate schemes such as the one Gerald Edelman used in his book "Neural Darwinism". The central principle is that these are competitive systems based on a frequency of use or frequency of observation basis.
You've made a good point in that there is as much value in searching for features which are *not* important as those that are. This is something which I havn't really taken into account in the systems which I've devised so far. Most systems which attempt to find patterns look for commonalities, not differences.
Just reading the Kurzweil/Wolfram review I tried to apply a cellular automata style technique to the problem of visual recognition within the ORAC system. Here there are a number of simple computational elements running in parallel. Like Wolfram my thinking was that the actual complexity of the computational elements is largely unrelated to the complexity of the end result. You could use simple logic gates, or devise very complicated simulated neurons.
- Bob
|