The roots of the BCP's general interest in artificial neural networks can be found in the study of the motion correspondence problem. This problem is solved whenever a system tracks moving objects. Object tracking is problematic because typically the visual information represented in the proximal stimulus (i.e., in retinal stimulation) is impoverished (e.g., Ullman, 1979). This means that there are many different object-tracking interpretations that are consistent with the proximal stimulus: In general, if a system is tracking N different objects, then there are N! different ways of tracking objects, only one of which will be correct.

Theoretical and experimental research (e.g., Dawson, 1987, 1988, 1989, 1990a; Dawson & Pylyshyn, 1988; Dawson & Wright, 1989) has indicated that solutions to the motion correspondence problem can be achieved by "filtering" proximal stimulus information with three additional constraining properties: the nearest neighbour principle (all things being equal, assign short motion correspondence matches), the relative velocity principle (all things being equal, assign neighbouring elements similar motion correspondence matches), and the element integrity principle (all things being equal, assign matches in such a way that elements do not split apart or fuse together). In general, these results suggests that the visual system should track identities in such a way that (1) the tracking is consistent with the proximal stimulus, and (2) the tracking is also as consistent as possible with the three constraining principles.

Yellow Area Shows Where MCP Is Solved In Owl Monkey Brain

Dawson (e.g., 1991) was able to show how these principles could be used as "soft" constraints on the motion correspondence problem by incorporating them into a connectionist network. His model, the "brainstate-in-a-sphere", was a variant of Anderson, Silverstein, Ritz and Jones' (1977) autoassociative connectionist network. Processing units in the model represented potential identity matches between elements seen at different positions and at different times. Connection weights were set to represent the three constraining principles. Dawson was able to show that the model generated the same solutions as does the human visual system to a variety of apparent motion displays, which require object identities to be tracked before the illusory motion can be filled in by the visual system. More recent research (Dawson & Wright, 1994; Dawson, Nevin-Meadows & Wright, 1994) has extended Dawson's original model by building in temporal sensitivity and a preference to assign matches that have the same contrast polarity.

However, Dawson's (1991) motion correspondence model is atypical of most modern ANNs, insofar as it is not plastic: the network's connections are "hardwired"; it does not learn. A natural extension of this approach to a problem in computational vision was to consider whether a modifiable network could, through learning, extract different kinds of natrual constraints for solving an ill-posed problem. To date, we have not yet addressed this particular question. This is because when our focus was broadened to include trainable artificial neural networks, we made some discoveries that have drawn us away from the original work on motion correspondence. These discoveries are described in other sections in this document.

References