反馈神经网络统一模型临界动力学研究及其在类脑计算机研制中的应用

项目名称： 反馈神经网络统一模型临界动力学研究及其在类脑计算机研制中的应用

项目编号： No.11471006

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 数理科学和化学

项目作者： 乔琛

作者单位： 西安交通大学

项目金额： 65万元

中文摘要： 由于对非线性动力系统具有良好的逼近能力，反馈神经网络已被广泛应用于众多工程技术领域中。本项目针对当前存在大量反馈神经网络模型个体以及相应数量庞大的动力学结果，而不同模型及其动力学结论之间难以直接进行比较，因此带来大量交叉与重复这一事实，通过发掘非线性激活映射所具有的本质特征，获得一个能最大程度上包括现有模型个体的反馈神经网络统一模型，并研究该统一模型在临界条件（稳定与不稳定的本质界限）下的动力学性态。所获反馈神经网络统一理论框架不仅能归并现有绝大多数反馈神经网络模型个体，而且能够消除相应模型个体动力学分析所存在的大量冗余内容、进而在最广泛的意义上概括并延伸其相应动力学结果。这一工作不仅对于神经网络研究具有重要的理论意义，而且也具有一系列实用价值。本项目将所获结论应用于IBM公司蓝脑计划之类脑计算机的研制中，以期获得一个能在最大限度下正确、稳定、高效工作的类脑计算机基本模型。

中文关键词： 反馈神经网络；统一理论；动力学；类脑计算机

英文摘要： Since recurrent neural networks have excellent approximation ability to nonlinear dynamic systems, they have become one of the powerful tools to study and solve many complicated problems in science and engineering applications. It should be pointed out that during the past 30 years, for the purpose of different requirements in applications, there have generated lots of recurrent neural networks models and for each ones, there have existed huge numbers of dynamics analysis. It's due to the fact that for different models, the equations which describe the models or the activation mappings of the models are not the same, so it's quite difficult to compare the models as well as their corresponding dynamics results, thus more and more redundancies arise. All these facts force us to find some effective solutions to deduce such nonsensical repetitions. In this project, we devote to discover some essential properties owned by most of the existing activation mappings, and based on which, we can achieve one unified continuous-time recurrent neural network model as well as one unified discrete-time recurrent neural network model. For both kinds of unified models, they can jointly cover almost all of the known model individuals belonging to continuous-time or discrete-time recurrent neural networks respectively. Under the critical condition, which is the intrinsic bounded line between stability and unstability, we further develop some convergence and stability theories for such two unified recurrent neural network models. The obtained unified theory framework for recurrent neural networks not only can merge the existing models, but also can eliminate the numerous overlaps of various dynamics results among these models. It can be shown that the research approach of unified theory adopted here is powerful, particularly in the sense of unifying, simplifying and extending the currently existing various models and the dynamics results of them. Furthermore, the obtained unified models for recurrent neural networks and the critical dynamics theories can be applied directly to the development of brain-like computers, which is a crucial part of the blue brain project of IBM. We can expect that the unified theory for recurrent neural networks will provide a basic guidance to establish a mechanism that is more similar to the human beings' brains, and ensure that such a mechanism can work correctly, stably and efficiently to the largest extent.

英文关键词： recurrent neural networks;unified theory;dynamics;brain-like computers