ion.) In this model retinotopic (spatial) units are connected directly to muscle control units. Each retinotopic unit can if saturated cause the appropriate contraction so that the new eye position is centered on that unit. When several retinotopic units saturate, each enables a muscle control unit independently and the muscle itself contracts an averap amount. Figure 20 shows the idea for a one-dimensional retina. For example, with units at positions 2. 4, 5, and 6 saturated, the net result is that the muscle is centered at 17 /4 or 4.25. (This idea can be extended to the case where the retinotopic units have overlappina fields.) This kind or organization could be extended to more complex movement models such as that of the organization of the superior colliculus in the monkey (Wurtz & Albano, 1980). Notice that each retinotopic unit is capable of enabling different muscle control units. The appropriate one is determined by the enabled x-origin unit which inhibits commands to the inappropriate control units via modifiers. One problem with this simple network arises when disparate groups ofretinotopic units are saturated. The present configuration can send the eye to an average position if the features are truly identical. The newtork can be modified with additional connections so that only a single connected component of saturated units is enabled by using additional object primitives. A version of this WTA motor control idea has already been used in a computer model of the frog tectum (Didday, 1976). There are still many details to be worked out before this could be considered a realistic model of vergence control. but it does illustrate the basic idea: local spatially separate sensors have distinct, active connections which could be averaged at the muscle for fine motor control or be fed to some intermediate network for the control of more complex behaviors. Convfftina Space to Time Consider the problem of controllina a simple physical motion. such as throwina a ball. It .is not hard to imagine that in a skilled motor performance unit-aroups fire each other in a fixed succession, leading to the motor sequence. The computational problem is that there is a unique set of effector units (say at the spinal level) that must receive input from each aroup at the rignt time. Figure 2ta depicts a simple case in which there are two effector units (e 1 , e2 ) that must be activated alternatively. The circles marked 1-4 represent units (or groups of units) which activate their successor and inhibit their predecessor (cf. Delcomyn, 1980). The main point is that a succession of outputs to a single effector set can be modeled as a sequence of time-exclusive groups representing instantaneous coordinate signals. Moving from one time step to the next could be controlled by pure timing for ballistic movements, or by a proprioceptive feedback signal. There is. of course. an enormous amount more than this to motor control and realistic models would have to model force.control: ballistic movements, gravity compensation. etc. The second part of Figure 21 depicts a somewhat fanciful notion or how a variety of output sequences could share a collection of lower level response units. The network shown has a single "Dixie" unit which can start a sequence and which joins in conjunctive connections with each note to specify its successor. At each time step, a WT A network decides what note gets sounded. One can imagine adding the rhythm network and transposition networks to other keys and to other modalities of output -----------------------------------------------