ABSTRACT
Classical conditioning is used as a means to investigate unsupervised learning in a single neuron. The double advantage of being extensively studied and rich in terms of responses to multiple temporally related stimuli makes for a good benchmark against which learning laws can be tested. This chapter relates the mathematical developments in the area of real time single-neuron learning by looking at the Hebbian, Rescorla-Wagner, Sutton-Barto, Tesauro, Gelperin-Hopfield-Tank, Klopf (drive-reinforcement) and Widrow-Hoff rules for synaptic adaptations. The learning rules are compared based on four classical conditioning simulations: CS duration, blocking, reacquisition, and second-order conditioning. The results of the examination have shown both the strengths and weaknesses of the individual models. A comparison between the Hebbian and drive-reinforcement models brings into question the use of a single learning law as the basis for the optimal development of a multiple neuron system.
