【摘要】 研究了支持向量机(SVMs)与前馈人工神经网络(FANNs)的关系.首先说明了两者在网络结构上的相似性,即输入函数均可表示为基函数的线性组合形式.然后指出了两者之间存在的关键差异:一是优化目标不同,FANNs只需要达到经验风险最小化,而SVMs寻求结构风险最小化;二是隐含层含义不同,SVMs的隐含层节点表现为支持向量,而FANNs则需要预先设定;三是模型复杂性不同,FANNs的模型复杂性由隐含层神经元数量来控制,而SVMs的复杂性独立于维度.最后,给出了两者在UCI标准数据集上的对比实验,用以评估它们在分类中的性能.更多还原

【Abstract】 This paper presents the relationship between support vector machines( SVMs) and feed-forward artificial neural networks( FANNs). It first illustrates the similarity in terms of network structure,i. e.,their input function can be expressed as a linear combination of basis functions. Then,some key differences between SVMs and FANNs are pointed out. One is the difference of optimization goal. FANNs only needs to reach empirical risk minimization,but for SVMs,it aims at minimizing the structural risk. Another is the difference of hidden layer. In SVMs,the nodes of hidden layer are exactly support vectors. While in FANNs,the nodes of hidden layer can be empirically determined in advance. Moreover,the difference of model complexity is also concerned.For FANNs,its complexity is controlled by the number of hidden nodes. But the complexity of SVMs is independent of data dimensionality. At the end,the comparative experiments on UCI benchmark datasets are provided,for evaluating the performance of SVMs and FANNs in classification.更多还原