Physiologically-based vision modeling applications and gradient descent-based parameter adaptation of pulse coupled neural networks.

摘要

Pulse coupled neural networks (PCNN) are analyzed and evaluated for use in primate vision modeling, and an adaptive PCNN is developed that automatically sets near-optimal parameter values to achieve a desired output. Biological vision processing principles, such as spatial frequency filtering, competitive feature selection, multiple processing paths, and state dependent modulation are analysed and implemented to create a PCNN based feature extraction network. This network extracts luminance, orientation, pitch, wavelength, and motion, and can be cascaded to extract texture, acceleration and other higher order visual features. Cortical information linking schemes, such as state dependent modulation and temporal synchronization, are used to develop a PCNN-based visual information fusion network which is used to fuse the results of several object detection systems. Next, the first fully adaptive PCNN is developed. Given only an input and a desired output, the adaptive PCNN finds all parameter values necessary to approximate the desired output. Gradient descent is applied to the PCNN to derive parameter adaptation equations (training rules) for all parameters. Implementing these equations forms a fully adaptive PCNN that minimizes squared error between the actual and desired output. All equations can be applied external to an existing PCNN.