论文部分内容阅读
对多层前传网络的过拟合问题进行了探讨。定义了逼近误差和逼近度作为人工神经网络（ＡＮＮ）的建模评价指标。通过应用于多元非线性校正的数值模拟和实际药物光度分析数据解析，表明该指标意义明确，便于掌握，且能较好地定量表述ＡＮＮ逼近规律的程度。
The problem of over-fitting of multi-layer prequel networks is discussed. The approximation error and the approximation degree are defined as the modeling evaluation index of artificial neural network (ANN). Through the numerical simulation applied to multivariate nonlinearity correction and the analysis of the actual drug photometric data, it shows that the index has a clear meaning and is easy to master, and can well quantify the degree of ANN approximation.