A network representation of low-level sensorimotor transformations using back propagation

The goal of this project was to develop a neuromuscular skeletal model that represented the transformations between different sensorimotor parameters associated with the control of posture and the generation of movement of the upper limb. The fundamental issue was one of trying to understand what type of information processing the spinal cord circuitry is capable of performing, as the lowest level of the motor hierarchy. This dissertation posited that a neural network model, trained using generalized back propagation, could mimic some of the gross systemic functions of the spinal circuitry.

The musculoskeletal system consisted of a two segment model arm with non-linear muscle-like actuators and the artificial neural system consisted of a multilayered network of highly interconnected processing elements. The muscle actuators were designed to reproduce the physiological length-tension characteristics of gross muscle tissue. The primary function of the artificial neural network was to convert goal oriented and postural command-like signals into muscle tension values (under conditions of static equilibrium). Command signals encoded the following parameters: limb end-point position; level of antagonist co-activation; limb end-point force; and rotational offset between local- and world-based coordinate systems. It was determined that network models trained using back propagation of errors could accurately internalize these highly non-linear sensorimotor transformations.

In addition, a length-sensitive neural controller was developed that consisted of two separately trained neural networks. A network that represented transformations between muscle lengths and end-point position was used as the feedback circuit to the neural controller, and a network that represented transformations between end-point position and muscle tensions was used as the feedforward circuit. The neural controller was able to maintain a stable limb end-point position following the application of muscle tension perturbations that varied both in magnitude and orientation. Error corrections were based on length encoded signals from all musculoskeletal actuators.