采用Reichardt运动检测器和Boltzmann Machine神经网络的运动计算

建立了一个探讨灵长类视皮层从Ｖ１区到ＭＴ区的运动信息加工原理的计算模型,这个过程的突出特征是视觉运动信息经过了从局部检测进步到整体感知。模型的第一层由用于抽提运动模式的局部速度以及结构性质的Ｒｅｉｃｈａｒｄｔ运动检测器组成,进一步的加工是通过Ｂｏｌｔｚｍａｎｎ Ｍａｃｈｉｎｅ神经网络来实现的。这种网络的学习算法具有局部更新的显著性质,在学习阶段,网络不断地修改联结权重以形成对于记录在网络的显单元上

MOTION COMPUTATION WITH REICHARDT'S MOTION DETECTORS AND BOLTZMANN MACHINE NEURAL NETWORK

A computational model is constrcted to explore the principles of visual motion processing from area V1 to area MT in primate visual cortex, where a significant transition from local motion detection to global integration of pattern motion takes place. The first layer of the model consists of Reichardt's elementary motion detectors (EMD), which extracts local velocity as well as structural properties of the moving pattern. Further processing is realized by the Boltzmann Machine neural network, whose learing algorithm inherits the distinctive property of local updaing. During the learning phase, the network continually revises the connection strengths to form the internal representation of the environmental structures, which have been coded on the "visible units" of the network. The structures here combine the various two dimensinal local vector fields presented and the corresponding true directions of pattern motion. Training of the network is influenced by a few netwok parameters. Generally, the behavior of the network improves with the increase of the number of presentations, showing that it is able to give the true direction of pattern motion regardless of different orientations of the moveing pattern.