| |
Abstract:
Artificial Neural Networks (ANNs) are used extensively to
model cognitive processes. The relation between ANNs and biological
networks of spiking neurons is not very clear, however. Some
consider ANNs to be biologically implausible. Others simply
identify the activity of Artificial Neurons (ANs) with the fire
rate of neurons in the cortex. To investigate the biological
plausibility of ANNs, we present a model of form processing in the
visual cortex. We show that biological plausibility of ANNs entails
an 'activation carries information' principle, that requires ANs to
have zero activity if no stimulus is present in their receptive
field. It turns out that this principle can best be maintained if a
symmetric activation interval for ANs is chosen around 0, e.g.,
[-1,1]. This means, however, that a single AN can not code for the
fire rate of a single (population of) neurons(s). Instead, we
propose that the activity of an AN codes for the state of a
cross-inhibitory circuit, that consists of two populations. We use
Wilson-Cowan dynamics to describe the circuit mathematically and
show that the ordinary ANN paradigm emerges from this
interpretation for a suitable choice of spike response functions.
Using spike response functions that are motivated by
neurophysiology, rather than by the desire to reproduce the ANN
paradigm, leads to a modification of the classical AN.
|
