Neural network models of eye movement control, object recognition, and robot navigation

摘要

This dissertation develops neural circuits within which two kinds of command sources are combined to provide adaptability and goal oriented control for eye movements, pattern recognition, and robot navigation. Two broad categories of command sources for control of action can be identified. Reactive commands are initiated by bottom-up signals arising from the sensors and/or preprocessing stages. Reactive commands resemble reflex commands but their generators may be much more complicated than a classical reflex arc. Planned commands are generated by top-down or feedback signals from higher level processing stages. The superior colliculus appears to be part of a final common path used for both reactive and planned saccadic eye movements. Current data on superior colliculus activity during eye movements show two distinct patterns. A neural network model is presented which shows how these patterns may emerge as a consequence of a neural design that enables reactive and planned commands to compete for movement resources. The model also shows how learning enables planned and reactive commands to interact within a consistent coordinate frame. Typical approaches to pattern learning and recognition are fundamentally reactive in that they assume that the input is preprocessed in its entirety before activating the recognition system. A neural network model is presented which utilizes planned commands to sequentially extract only those features that improve the discriminability between candidate recognition categories. Simulations show how a sequence of choices is made and processing halted when the confidence level for a category reaches a desired value. In order to maintain sensitivity to novelty, planned commands are supplemented by reactive commands that allow salient characteristics of the input to influence the next extraction operation. Common algorithmic approaches to robot navigation within dynamic environments utilize a single high level planning stage which must be reengaged following each change in the environment. A neural network trajectory generator is introduced that combines planned commands with reactive commands in order to adapt to changes in the environment. Simulations of robot navigation through a cluttered 2-D environment show that the trajectory generator is a viable alternative to approaches with similar performance but much higher computational costs.