A NEURAL NETWORK MODEL OF AVERSIVE BEHAVIOR

摘要

A neural network model of aversive behavior is proposed. It has one neuron for each possible response of the animal (R), and from neurobiological bases we assumed that there is a prediction of the unconditioned stimulus (US). It is computed by another artificial neuron. The inputs of response neurons are: the prediction, the short-term memory of the conditioned stimuli (CSs) and of the US. Based on Hebbian or anti-Hebbian learning, depending on the prediction of the US, the synaptic weights of the response neurons are calculated. The short-term memories of CSs, of the US and of the Rs are inputs of the neuron that computes prediction. The synaptic weights of this neuron are calculated based on the delta rule. Finally, animals execute any response higher than a threshold. While the neurobiology underlying aversive behavior may be very complex, the dynamic process that governs aversive behavior is simple, responses are associated with the input stimuli by the Hebbian rule if the prediction of the unconditioned stimulus is lower than a threshold, and by the anti-Hebbian rule if it is higher. The simulations of the model and the experimental data are compared. This network can be applied to control autonomous systems.