Many papers have been devoted to discussions of the mechanism of conditioned reflex, which is the most fundamental process in animal learning and pattern recognition. Their discussions are mainly focussed on the mechanism of discriminating different patterns. Thus their models are poor in generalization which is the most essential feature of conditioned reflex and animal learning. Besides, learning process of these models is quite different from that of conditioned reflex as was reported by Pavlov.
In this paper the author presents a model of conditioned reflex which is capable of simulating the following essential features of conditioned reflex : derivation of conditioned response by reinforcement, extinction of the formerly established reflex, delayed reaction, external inhibition and disinhibition, generalization and differentiation, etc.
The model is composed of four types of neurons, two excitatory (S & R) and two inhibitory (T & V). Input and output of these cells are non-negative continuous quantity. Each cell has threshold and gives output proportional to suprathreshold input.
Conditioned stimuli are sent to sensory cell S and unconditioned stimuli are sent to response cell R. Output of cells S is sent to cell R and forward inhibitory cell T, which may inhibit cell R. Output of cell R may evoke certain response of uncoditioned reflex, in the meantime inhibiting cells S and T via backward inhibitory cell V.
In this model, as was experimentally proven by Pavlov, extinction of a conditioned reflex is not due to decay of once established connections but to development of new connections which are to overcome the effect of the former ones. Inhibitory cell T works for development of delayed reaction and extinction, while backward inhibitory cell V works for external inhibition of conditioned reflex and disinhibition.
The learning of the model is done by changing the weights of connections of neurons. Mathematical reinforcement rule is formulated from physiological reinforcement conditioning. Some examples are given to show the learning process of the model.
View full abstract