[top of page 8 of handout]
CLASSIFICATION OF LEARNING MECHANISMS

"Paradigms of Learning"pp 159-161 (Rumelhart et al ,1986)

Auto Associator set of patterns repeatedly presented and system “stores” it ... “a pattern is associated with itself”. Then when part is given, the whole original is retrieved == Habituation or perceptual learning

Pattern Associator pairs repeated. System learns that when one member is presented it is “supposed” to produce the other. “In this paradigm one seeks a mechanism in which an essentially abitrary set of input patterns can be paired with an abitrary set of output patterns. == Pavlovian conditioning, or stimulus-response motor habits.

Classification paradigm this is a variant on the previous 2. There is a fixed set of categories into which the stimulus patterns are to be classified. There is a training session in which the system is presented with the stimulus patterns along with the categories to which each stimulus belongs. This is where the “perceptron convergence theorem is proved” == Discrimination Learning

Regularity detector For a population of stimulus patterns each pattern S(k) is presented with a probability p(k). The system is supposed to discover statistically salient features of the input population. (cf Rescorla's conditional probability suggestions) and experiments on probability learning in animals. == probability learning and complex Pavlovian conditioning; discrimination of reward rates.