Use of Top-Down Information for Visual Information Processing

当我们看着一个物体时，我们不会被动地接受我们视野内的全部信息，而是主动地只收集必要的信息。我们把注意力集中在吸引我们兴趣的地方。我们从那里捕获信息，并有选择地对其进行处理。我们经常试图利用周围地区的信息来预测一个模式，并通过确认最初的预测是否正确来识别它。自上而下的信息在这种主动的信息加工中起着重要的作用。最近，关于包括自上而下加工在内的高级大脑功能的各种神经生理学和心理学实验结果都有报道。我们试图从信息处理的角度对这些结果进行系统的分析，并进行建模研究，以获得新一代信息处理器的新设计原则。也就是说，我们首先提出了一个新的更高大脑功能的模型，并对该模型进行了改进，使其行为方式与生物大脑相似。同时，我们还对这些模型在实际问题中的实际应用做了一些实验。我们得到了以下结果：(1)一个能够识别和恢复部分遮挡模式的模型(2)提高了新认知器的识别率(一个稳健的视觉模式识别模型)(3)一种适合于多层神经网络的增量学习方法(4)利用模糊进行稳健的图像处理-一个从视觉模式中提取对称轴的神经网络模型(5)光流提取：MT和MST细胞的神经网络模型(6)揭示图形-背景分离、轮廓整合、LPD刺激知觉、整体运动整合和透明运动之间的关系：心理学实验和计算模型

When we are looking at an object, we do not passively accept whole information within our visual field, but actively gather necessary information only. We focus our attention to the places that attract our interest. We capture information from there and process it selectively. We often try to predict a pattern using information from surrounding areas, and recognize it by confirming whether the initial prediction was correct. Top-down information plays an important role for such active processing of information.Varieties of neurophysiological and psychological experimental results on higher brain functions, including top-down processing, have recently been reported. We tried to analyze these results systematically from the stand-point of information processing, and made modeling research to obtain new design principles for information processors of a new generation. Namely, we first propose a new model for a higher brain function, and improved the model so as to behave in a similar way as the biological brain. At the same time, we also made several experiments for practical implementation of the models to real-world problems.As a result of these researches, we have obtained the following results :(1)A model capable of recognizing and restoring partly occluded patterns(2)Improving the recognition rate of the neocognitron (a model for robust visual pattern recognition)(3)A new method for incremental learning appropriate for multi-layered neural network(4)Use of blur for robust image processing --- a neural network model that extracts axes of symmetry from visual patterns(5)Extraction of optic flow : A model of neural network for MT and MST cells(6)Psychological experiments and models revealing relations among the mechanisms of figure-ground separation, contour integration, and motion integration.(7)Relations among the perception for LPD stimuli, global motion integration, and transparent motion : psychological experiments and a computational model