基于角度优化的鲁棒极端学习机算法

Algorithm of Robust Extreme Learning Machine Based on Angle Optimization

摘要:

极端学习机因其学习速度快、泛化性能强等优点,在当今模式识别领域中已经成为了主流的研究方向;但是,由于该算法稳定性差,往往易受数据集中噪声的干扰,在实际应用中导致得到的分类效果不是很显著;因此,为了提高极端学习机分类的准确性,针对数据集样本中带有噪声和离群点问题,提出了一种基于角度优化的鲁棒极端学习机算法;该方法利用鲁棒激活函数角度优化的原则,首先降低了离群点对分类算法的影响,从而保持数据样本的全局结构信息,达到更好的去噪效果;其次,有效的避免隐层节点输出矩阵求解不准的问题,进一步增强极端学习机的泛化性能;通过应用在普遍图像数据库上的实验结果表明,这种提出的算法与其他算法相比具有更强的鲁棒性和较高的识别率。

Abstract:

Due to its fast learning and generalization performance, etc, extreme learning machine has become the mainstream of research in today's field of pattern recognition. However, owing to the poor stability of the algorithm, the data set often vulnerable to noise, causing the classification results are not very significant in the practical application. Therefore, in order to improve the accuracy of classification of extreme learning machine, aiming at to solve the problem of noise and outliers in the data set samples, a robust extreme learning machine algorithm is presented based on angle optimization. This method using the principle of the robust activation function of angle optimization, firstly, reduces the impact of outliers on the classification algorithm to maintain the overall structure information of data set samples to achieve better denoising effect. Secondly, it can also effectively avoid the question which is the inaccurate solving of hidden nodes output matrix, and further enhance the generalization performance of extreme learning machine. The experimental results of the application of universal image database show that the proposed algorithm compared with other algorithms has better robustness and higher recognition rate.