Neural Networks with Low Local Firing Rates
Naama Rubin, Haim Sompolinsky · Europhysics Letters (EPL) · 1989
P ACS. 64.60C- Order-disorder and statistical mechanics of model systems. P ACS. 05.90- Other topics in statistical physics and thermodynamics. P ACS. 89.70- Information science. Abstract.- Neural network models of associative memory with uniform inhibition are studied. It is shown that for sufficiently strong inhibition the system orders only partially even at low temperatures. A fraction of the neurons freezes in a quiescent state while the activities of the rest of the neurons fluctuate in time around an average level that is small compared to the saturation level. These models may help understanding the origin of the low neuronal firing rates observed in cortical recordings. Several computational functions of neural assemblies have been modelled by simplified neural networks consisting of highly connected, nonlinear elements [1-3]. The state of the network is defined in terms of the instantaneous activities of its elements, which represent the firing rates of biological neurons. It has been proposed that certain computations, and in particular retrieval of memory, are performed by convergence of the dynamic flow of the networ'k to the appropriate attractor. The outcome of the computation is represented by the