I stumbled upon a paper from P.Diehl and M.Cook with the title "Unsupervised learning of digit recognition using spike-timing-dependent plasticity" and I'm trying to understand the logic behind the network connection they made.
The network is as follows. The inputs (size of an image 28x28) are connected to the usual all-to-all fashion with positive weights to an NxN layer of neurons. They are decoded using poisson random distribution in which the frequency of spikes of a pixel is set accordingly to the pixel value. The NxN layer is connected 1 on 1 to an NxN layer of inhibitory neurons. These neurons inhibit all other neurons except the one they are connected to. Thus they are connected all-to-all with an exception.
According to the paper this provides competition among neurons. I cannot understand how, in this particular connection, competition is provided. How can different neurons inherit different properties? To me it seems that all neurons will inherit the same properties, thus no differences in weights will be made in the training session among all neurons. For example, if the input 5 is passed to the network all weights of all neurons will try to adjust according to 5. Then if input 7 is passed next, all the weights will be updated according to the new number (7). It is expected, though, that some weights will keep the previous adjustment ie that some weights will have the properties of 5 and the others the properties of 7.