VISUALLY REACTIVE, MULTIMODAL, INTENTIONAL, AND PREDICTIVE MOVEMENTS: A SYNTHESIS 11.1. Avoiding Infinite Regress: Planned Movements Share Reactive Movement Parameters

In this Chapter, we tie together the design principles and mechanisms ~hich we have introduced in the preceding chapters. This synthesis quantIfies a developmental sequence in which obligatory saccadic reactions to flashing or moving lights on the retina are supplemented by attentionally mediated movements towards motivationally interesting or intermodal sensory cues. These movements are supplemented once again by predictive saccades which form part of planned sequences of complex movement synergies that can totally ignore the sensory substrate on which they are built. The analysis of such a complex system would be' hopeless in the absence of sufficient design constraints. The previous chapters illustrate our contention that an analysis of development and learning is essential to discover sufficiently many constraints. This chapter continues such an analysis until it discloses a small number of closely related global network designs that are capable of simultaneously satisfying all of the constraints. The primacy of learning constraints is perhaps most urgently felt when one squarely faces the problem of infinite regress. On what firm computational foundation can adaptive calibrations be based? If parameters in several subsystems can all be changing due to different types of learning signals, then what prevents them from confounding each other's learning by causing a global inconsistency within the system? An analysis of how the problem of infinite regress may be solved shows how to hierarchically organize motor learning circuits so that each circuit can benefit from and build upon the learning of a more primitive learning circuit. This hierarchical organization of learning problems can be profitably compared to the hierarchical organization whereby simple motor synergies in the spinal cord can be temporally organized by descending co~tical commands (Grillner, 1975; Grossberg, 1969d, 1974). The need to solve the problem of infinite regress is evident in our method of calibrating a target position command in muscle coordinates at the head-muscle interface, or HMI (Chapter 4). We have suggested that an active target position map (TPM) population will sample whatever eye position corollary discharges are read-into the HMI after the eye comes to rest. These eye position signals mayor may not correspond to the target position that is active within the TPM. The TPM will eventually sample the correct eye position only if an independent learning mechanism enables the eye to accurately foveate retinal lights. We have suggested that this independent learning mechanism forms part of the RCN, or retinotopic command network (Chapter 3), whereas the HMI forms part of the VCN, or vector command network (Chapter 4).