Study On The Self - Development Network Model Of Visual Cortical Visual Cortex

In this paper, some works have been done on the autonomous development mechanism of the cortex-like networks--Where-What Networks (WWNs), with the goal of application (such as for the third generation intelligent robot that will enter the people’s daily life). Theoretically, WWNs are a paradigm of DN in the sense that this category of networks roughly models the dorsal and ventral pathways of the biological visual system. The ventral pathway is mainly responsible for the shape and color information of objects. The dorsal pathway mainly processes information the about motion and position.The main contribution in this paper can be divided into three parts. The first part is to verify the practical application potential of the WWN. Via parallelizing the various stages of network training with CUDA, which is a GPU parallel computing framework provided by NVIDIA, the training time was shortened. The hierarchical parallelization technique has achieved a speedup of16times compared to the C program and the processing speed could reach26frames per second, which is generally satisfied with the requirement of real-time processing. The performance of the WWN after engineering optimization indicated the strong probability of practical application of WWNs in the aspect of processing speed.Secondly, the autonomous development mechanism of the WWNs was greatly improved to meet the requirement of application in real environments. That is, a skull-closed autonomous developmental network WWN-6was proposed based on WWN-3, which preliminary realizes a complete autonomous development function. Specifically, some new biological mechanisms were applied in WWN-6, which include the synapse maintenance to realize dynamical adjustment of the receptive fields; a design of two states and the cell re-genesis mechanism for neurons to realize the autonomous establishment and cutoff mechanism of the connections between the neurons in different layers so that the neuronal resources can be regulated dynamically; a neuron release rule which preliminarily solve the problem of over-occupation of the neuronal resources for earlier objects in the incremental learning. In the experiments of the performance verification for the above new mechanisms, some continuous video clips with objects to be learned were taken in the natural environments as the original data. The results of massive experiments proved the effectiveness of all these mechanisms and preliminarily shown that WWN-6can be used to autonomously learn and recognize the objects in practical natural environments.Finally, in order to let the network be closer to the practical application, a new WWN-7, which was able to deal with objects of multiple scales, was proposed based on the works on WWN-6. With truly skull-closed, the WWN-7was studied to clarify the possibility of multiple-scale object processing under complicated background. The results shown that WWN-7can basically realize three concepts (type, location and scale) learning for an object at the same time and therefore in potential, WWN-7can learn objects more naturally and perfectly in the real environments.