高级检索

    高维数据中鲁棒激活函数的极端学习机及线性降维

    Robust Activation Function of Extreme Learning Machine and Linear Dimensionality Reduction in High-Dimensional Data

    • 摘要: 极端学习机(extreme learning machine, ELM)训练速度快、分类率高,已经广泛应用于人脸识别等实际问题中,并取得了较好的效果.但实际问题中的数据往往维数较高,且经常带有噪声及离群点,降低了ELM算法的分类率.这主要是由于:1)输入样本维数过高;2)激活函数选取不当.以上两点使激活函数的输出值趋于零,最终降低了ELM算法的性能.针对第1个问题,提出一种鲁棒的线性降维方法(RAF-global embedding, RAF-GE)预处理高维数据,再通过ELM算法对数据进行分类;而对第2个问题,深入分析不同激活函数的性质,提出一种鲁棒激活函数(robust activation function, RAF),该激活函数可尽量避免激活函数的输出值趋于零,提升RAF-GE及ELM算法的性能.实验证实人脸识别方法的性能普遍优于使用其他激活函数的对比方法.

       

      Abstract: Extreme learning machine (ELM), with the advantage of fast training and high classification accuracy, has been widely used in practical applications (for example, face recognition) and got good result. While ELM algorithm is often severely affected by noise and outliers in high-dimensional real word datasets, which will reduce the accuracy rate of ELM. This should be attributed to the following two reasons: 1) the high dimensionalities of input samples; 2) improper selection of activation function. The two reasons above result in that the outputs of activation functions are approaching zero, which finally reduce the performance of ELM. As for the first problem, we propose a robust linear dimensionality reduction method, RAF-Global Embedding (RAF-GE), to preprocess the high dimensional data and then classify the data with ELM. While for the second one, we give an in-depth analysis of different activation function and propose a robust activation function (RAF) which can avoid the outputs of activation function approaching zero, thus it can improve the performance of RAF-GE and ELM as well. The experiment results show that the performance of face recognition method in this paper generally outperforms the comparing methods with other activation function.

       

    /

    返回文章
    返回