Slide 18 of 37
Notes:
The perception of an object provides evidence for one or more hypotheses, each of which is a possible interpretation or categorization for that object. Some of the features of the perceived object are more consistent with one hypothesis than with another. The system oscillates among the alternative hypothesis until it finds the one that best satisfies the constraints provided by the evidence and the mutual relations among the hypotheses (see also McClelland, 1986). The process by which such oscillations and settling on a conclusion is called relaxation. Waltz (1975) showed how such techniques could be applied to object recognition. Hopfield and Tank (1985) showed how a connectionist system could be used to implement a relaxation constraint-satisfaction system. The constraints can be described like rules, but they can also be implemented using other kinds of more neurobiologically consistent devices. For example, in a neural network, each of the hypotheses and each object feature to which the system responds can be represented by a small group of perhaps distributed neurons. Connections between these neurons represent the constraints and associations among the hypotheses and inputs. Connection strengths among the neurons represent the degree to which features and categories are related to one another (See Roitblat & von Fersen, 1992).
