A research into fine-grained object systems for computer vision and neural networks

(a)计算机视觉与神经网络的大规模并行架构本研究项目基于面向对象系统的概念，对大规模并行计算机视觉系统进行了研究。这是因为面向对象系统为我们提供了一个很好的框架来表示并行进程，以及表示复杂的系统，如计算机视觉系统。神经网络在计算机视觉中非常有用，也从大规模并行的角度进行了研究。我们采用的架构是基于一个简单的静态数据流模型和二维网格通信网络，我们设计了一个称为AMP的大规模并行机。关键问题是提供轻量级的消息处理机制，这在细粒度并行处理(例如：对象系统)执行。此外，我们还通过软件仿真证明了AMP在计算机视觉应用中的高效率。为了简化大规模神经网络的描述，我们引入了面向对象的概念，对AMP的基本函数式语言Valid进行了扩充。增强语言为我们提供了一种封装神经元和神经元网络功能的机制，也为我们提供了一个神经网络类库。(c)基于神经网络的大规模并行计算机视觉系统我们提出了一种基于多层、分层神经网络的大规模并行计算机视觉系统ICE(Image centric) system。在这个系统中，图像理解的结果是用最高级的激活单元的一系列组合来表示的：每个单元对应一个单词的含义。

(a)Massively parallel architecture for computer vision and neural networks We have made a research into a massively parallel computer vision system based on the concept of object-oriented system in this research project. This is because object-oriented systems provide us a good framework for representing parallel processes, as well as for representing complex systems such as computer vision systems. Neural networks, which are quite useful in computer vision, have been also considered from the viewpoint of massive parallelism. The architecture which we have adopted is based on a simple, static dataflow model with a two dimensional mesh communication network, and we have designed a massively parallel machine called AMP. The key issue is to provide a light-weight message handling mechanism, which is inevitable in fine-grained parallel process(i.e.,object system)execution. In addition, we have shown the high efficiency of AMP in computer vision applications by software simulation.(b)Programming language for neural networks We have augmented a programming language Valid, which is a basic functional language for AMP, by introducing the object-oriented concept in order to simplify the description of large-scaled neural networks. The augmented language provides us a mechanism to encapsulate functions of neurons and neuron networks, and also provides us a class library of neural networks.(c)Massively parallel computer vision system based on a neural network We have proposed a massively parallel computer vision system called ICE(Image CEntered)System, which is based on a multi-layered, hierarchical neural network. In this system a result of image understanding is represented in a sequence of combinations of activated units in the highest layer: each of the units corresponds to a word meaning.