Self-organizing neural networks for visual navigation and adaptive control

摘要

This dissertation describes two self-organizing neural networks that learn to convert perceptual information into representations useful for planning and performing action. The first network converts optic flow into heading estimation, scene depth estimation, and moving object localization. Network weights are trained during an action-perception cycle in which self-generated eye and body movements produce optic flow. The confounding effect of eye movement during translation is suppressed by learning the relationship between eye movement outflow commands and the optic flow signals they induce. The remaining optic flow field is due only to observer translation and independent motion of objects in the scene. A self-organizing feature map encodes heading by categorizing normalized translational flow patterns. Cells in the map that respond maximally to movement along a certain heading simultaneously learn the average translational flow signals induced by that motion. Comparing these learned averages to instantaneous flow fields yields a relative depth estimate across the entire visual field. Active heading map cells also learn expected optic flow directions. These expected patterns are subtracted from normalized flow patterns to detect objects moving independently of the observer. All learning processes take place concurrently and require no external teachers. The network will automatically adapt to the sensor geometry and other opto-mechanical properties of robotic vision systems. Simulations verify its visual navigation performance using both noise-free and noisy optic flow information. The second neural architecture is a general-purpose adaptive controller based on the DIRECT model of movement control. The function approximation technique used in DIRECT is replaced with a more efficient adaptive radial basis function (RBF) network. Specifically, the network is a variant of a Gaussian normalized RBF network with linear mapping coefficients at each basis. Learning laws are derived that allow the widths and centers of the bases to adapt to unknown mappings. The result is a network that is more memory efficient and can learn inverse control problems with many fewer training movements than the original DIRECT model. Simulations illustrate kinematic control of a planar 3-joint arm and a 7-dimensional articulatory speech synthesizer.