Ordering Chaos in A Neural Network with Linear Feedback
LP Wang
2000-01-01
Abstract:There have been extensive, multi-discipline research interests in ordering chaos recently; however, most work has been done on systems with few degrees of freedom. In the case of neural networks, numerical simulations are usually relied upon to study networks with a small number of neurons. Based on our earlier work [Wang et al, Proc. Natl. Acad. Sci. USA, vol.87, 9467, 1990], we present ordering of chaos in an exactly solvable neural network with arbitrarily large number of neurons using a simple linear feedback.In this network, N binary McCulloch-Pitts neurons are connected with both first-order and second-order Hebbian synapses, which are disconnected at random in the spirit of the sparse and asymmetric connectivities found in the real brain. In the absence of the linear feedback, depending on the network parameters, such as the number of stored patterns and the relative strength of first- and second-order interactions, the network can exhibit a variety of dynamics such as static fixed points, periodic oscillations, and chaos.When the output of each neuron, multiplied by a positive constant, is fedback to this neuron itself, we show that chaos and oscillations in this modified Hopfield neural network are suppressed or even eliminated.When a negative feedback is used, "crisis" can occur: all fixed points, oscillations, and chaos with positive overlaps between the network state and the initial attracting pattern disappear.