Sensory mapping, development and their applications to feature and attention selection.

摘要

Rich literature has been generated on computer vision systems. However, an implementable computational model of immediate vision for general and unknown environments is still illusive. Motivated by the autonomous development process of humans, we are interested in building a robot vision system that automatically develops its visual cognitive skills through realtime interactions with the environment. Developmental vision is required by the demands of general-purpose vision systems for complex human environments.; Based upon the above requirement, in this dissertation we propose a general architecture called Staggered Hierarchical Mapping (SHM) that performs feature derivation for a set of receptive fields and attention selection. The work reported here is motivated by the structure of the early visual pathway. We use several layers of staggered receptive fields to model the needed units of local analysis. From sequentially sensed video frames the proposed algorithm develops a hierarchy of filters, whose outputs are maximally uncorrelated within each layer, and contains an increasing scale of receptive fields that range from low to high layers. We also show how this general architecture can be applied to occluded face recognition, which demonstrates a case of attention selection. Besides this general neural network architecture, we develop several sensory mapping learning rules for deriving feature detectors, including a fast incremental independent component analysis (ICA) method called Lobe Component Analysis (LCA), which derives independent components for many natural cases. A mathematical analysis of the algorithm has been done and which shows the advantages of the LCA method.; We push the research further by investigating the LCA method in a overcomplete setting, which means the number of basis functions is greater than dimension of the observation. A new incremental method is developed to solve the equivalent regression problem with sparse regulization term, i.e. the LASSO regression. We show that the LCA method is a special case when noise is not present. It makes LCA's principal applicable to a large variety of applications, e.g. classification, regression, and feature selection.