针对海洋矿物分类问题,提出了改进后的单输出切比雪夫多项式神经网络(single-output Chebyshev-polynomial neural network with general solution,SOCPNN-G)。该模型利用伪逆的通解来求参数,扩大解空间,能获得泛化性能更加优良的权重...针对海洋矿物分类问题,提出了改进后的单输出切比雪夫多项式神经网络(single-output Chebyshev-polynomial neural network with general solution,SOCPNN-G)。该模型利用伪逆的通解来求参数,扩大解空间,能获得泛化性能更加优良的权重。在该模型中,子集方法用于确定神经元的初始数量和获得交叉验证的最佳重数。最后将改进的SOCPNN-G模型用于海洋矿物数据集中进行实验,结果表明,该模型训练准确率和测试准确率分别达到90.96%和83.33%,且对计算性能要求较低。这些优越性表明该模型在海洋矿物的实际应用中具有很好的前景。展开更多
With the best trigonometric polynomial approximation as a metric, the rate of approxi- mation of the one-hidden-layer feedforward neural networks to approximate an integrable function is estimated by using a construct...With the best trigonometric polynomial approximation as a metric, the rate of approxi- mation of the one-hidden-layer feedforward neural networks to approximate an integrable function is estimated by using a constructive approach in this paper. The obtained result shows that for any 2π-periodic integrable function, a neural networks with sigmoidal hidden neuron can be constructed to approximate the function, and that the rate of approximation do not exceed the double of the best trigonometric polynomial approximation of function.展开更多
文摘针对海洋矿物分类问题,提出了改进后的单输出切比雪夫多项式神经网络(single-output Chebyshev-polynomial neural network with general solution,SOCPNN-G)。该模型利用伪逆的通解来求参数,扩大解空间,能获得泛化性能更加优良的权重。在该模型中,子集方法用于确定神经元的初始数量和获得交叉验证的最佳重数。最后将改进的SOCPNN-G模型用于海洋矿物数据集中进行实验,结果表明,该模型训练准确率和测试准确率分别达到90.96%和83.33%,且对计算性能要求较低。这些优越性表明该模型在海洋矿物的实际应用中具有很好的前景。
基金This paper was supported by the National Basic Research Program of China （973 Program） under Grant No. 2007CB311000, the Natural Science Foundation of China under Grant Nos. 11001227, 60972155, 10701062, the Key Project of Chinese Ministry of Education under Grant No. 108176, Natural Science Foundation Project of CQ CSTC Nos. CSTC 2009BB2306, CSTC2009BB2305, the Fundamental Research Funds for the Central Universities under Grant No. XDJK2010B005, XDJK2010C023.
文摘With the best trigonometric polynomial approximation as a metric, the rate of approxi- mation of the one-hidden-layer feedforward neural networks to approximate an integrable function is estimated by using a constructive approach in this paper. The obtained result shows that for any 2π-periodic integrable function, a neural networks with sigmoidal hidden neuron can be constructed to approximate the function, and that the rate of approximation do not exceed the double of the best trigonometric polynomial approximation of function.
