Page 69 - Foundations of Cognitive Psychology : Core Readings
P. 69

68   Jay L. McClelland, David E.Rumelhart, and Geoffrey E.Hinton






















                Figure 4.5
                A sequence of configurations assumed by the stick ‘‘person’’ performing the reaching task described
                in the text, from Hinton (1984).The small circle represents the center of gravity of the whole stick-
                figure, and the cross represents the goal to be reached.The configuration is shown on every second
                iteration.


                  Though the simulation was able to perform the task, eventually satisfying
                both goals at once, it had a number of inadequacies stemming from the fact that
                each joint processor attempted to achieve a solution in ignorance of what the
                other joints were attempting to do.This problem was overcome by using addi-
                tional processors responsible for setting combinations of joint angles.Thus, a
                processor for flexion and extension of the leg would adjust the knee, hip, and
                ankle joints synergistically, while a processor for flexion and extension of the
                arm would adjust the shoulder and elbow together.With the addition of pro-
                cessors of this form, the number of iterations required to reach a solution was
                greatly reduced, and the form of the approach to the solution looked very nat-
                ural.The sequence of configurations attained in one processing run is shown in
                figure 4.5.
                  Explicit attempts to program a robot to cope with the problem of maintaining
                balanceasitreaches foradesiredtargethaverevealedthe difficultyofderiving
                explicitly the right combinations of actions for each possible starting state and
                goal state.This simple model illustrates that we may be wrong to seek such an
                explicit solution.We see here that a solution to the problem can emerge from
                the action of a number of simple processors each attempting to honor the con-
                straints independently.

                Perception
                Stereoscopic Vision  One early model using parallel distributed processing was
                the model of stereoscopic depth perception proposed by Marr and Poggio
                (1976).Their theory proposed to explain the perception of depth in random-dot
                stereograms (Julesz, 1971; see figure 4.6) in terms of a simple distributed pro-
                cessing mechanism.
                  Julesz’s random-dot stereograms present interesting challenges to mecha-
                nisms of depth perception.A stereogram consists of two random-dot patterns.
   64   65   66   67   68   69   70   71   72   73   74