Studying networks of neurons in animals and insects can provide insight into the natural world -- and inspiration for manmade networks to aid in computing and other applications.
A new model of neural networks, based on recent studies of fish and insect olfactory systems, suggests a way that neurons can be linked together to allow them to identify many more stimuli than possible with conventional networks.
Researchers from the Institute for Nonlinear Science at the University of California, San Diego propose that connections between neurons can cause one neuron to delay the firing of another neuron. As a result, a given stimulus leads to a specific time sequence of neural impulses.
In essence, the interconnected neurons include time as another dimension of sensory systems through an encoding method called Winnerless Competition (WLC).
Using a locust antenna lobe exposed to fragrances such as cherry and mint for comparison, the researchers found their model could identify roughly (N-1)! (equal to (N-1) x (N-2) x ...x 2) items with a network built of N neurons.
That means that a ten-neuron WLC network should be able to identify hundreds of thousands as many items as a conventional ten-neuron network -- and the benefits would increase as networks grow.
The WLC model helps explain how the senses of animals, insects and even humans can accurately and robustly distinguish among so many stimuli. The model is a mathematical rationale as to why a rose, by any other name, would smell as sweet -- but won't ever smell like an onion.
Ultimately, the WLC model may lead to high capacity, potent computing networks that resemble an insect antenna or a human nose more than a desktop PC.
(Reference: M. Rabinovich et al, Physical Review Letters, 6 August 2001; text at http://www.aip.org/physnews/select.)
[Contact: M. Rabinovich]