Study of self-inhibited analogue neural networks using the self-consistent signal-to-noise analysis
Tomoki Fukai, Masatoshi Shiino · Journal of Physics A Mathematical and General · 1992
The authors study an analogue neural network model whose response function is linear and consequently has no asymptotes in order to examine a possible new mechanism for regulating neuron activities by means of neural feedback circuits. Inhibitory self-coupling is introduced as an example of such feedback mechanisms, which for simplicity of the analysis is assumed to be piecewise linear. Their self-consistent signal-to-noise analysis is applied to explore the equilibrium properties of the model neural network. The analysis revealed that there exists a finite optimal value for the linear analogue gain that maximizes storage capacity for a given value of the self-coupling. The results of the analysis for such equilibrium properties as storage capacity are quite consistent with the results of computer simulations.