论文部分内容阅读
从样本复杂性、结构复杂性、学习策略和建模技术4个方面对基于领域知识的神经网络泛化性能研究进展进行了评述,指出了目前基于领域知识神经网络泛化性能研究存在的主要问题是只是利用研究对象的单调性、凸性、对称性和增益等一些简单非线性特征来虚拟训练样本、形成非监督学习算法约束条件、构造节点作用函数等方面.利用关于研究复杂对象部分已知的物理机制或动力学特性来建立有一定物理基础的神经网络模型,从而有效控制网络训练存在的过学习问题是今后神经网络泛化理论与方法研究的主要发展趋势.
From the aspects of sample complexity, structure complexity, learning strategies and modeling techniques, the progress of the research on generalization of neural network based on domain knowledge is reviewed. The main problems in the current research on the generalization performance of neural network based on domain knowledge are pointed out Just use some simple non-linear features such as monotonicity, convexity, symmetry and gain of the research object to virtual training samples, form the constraints of unsupervised learning algorithm and construct the function of nodes, etc. Using the part about the study of complex objects known Neural network model with a certain physical foundation based on physical or dynamical characteristics of the neural network can effectively control the over-learning problem existing in network training. This is the main trend in the research of neural network generalization theory and method in the future.