Research On Extreme Learning Machine And Its Application In Classification

As a new learning framework,Extreme Learning Machine(ELM)has attracted a great deal of research attention in pattern recognition and machine learning.Compared to traditional neural networks,ELM randomly selects input weights,and does not require any iteration to determine the output weight by least squares.Therefore,the extreme learning machine has the characteristics of fast and efficient in the classification task.With these remarkable advantages of the ELM algorithm,we have conducted profound research.The generalization performance of the extreme learning machine is limited due to the high complexity of the data samples and the limited tagged data in supervised learning.We have studied this issue in view of its significance for research and then improved the ELM algorithm as the following two aspects:1)Based on the idea of manifold learning,this paper proposes an extreme learning machine based on locality information preserving(LPKELM).Manifold learning method reveals the inherent geometric structure of the data points.From the perspective of manifold learning,the data samples can be considered to be from the same marginal distribution.If the data points are close together in the high-dimensional space,they should have similarities,and belong to the same category.Therefore,by mining the geometric structure between sample points,it is possible to provide effective information for pattern classification.LPKELM introduces the geometric structure information and discriminant information of the data sample into the ELM model,which can remedy the problem that the ELM algorithm is not adequately studied in supervised learning.2)Using LPKELM algorithm,a new multi-mapping kernel extreme learning machine is proposed(GKELM).The role of the kernel function implies a mapping from a low-dimensional space to a high-dimensional space.This mapping can realize linearly separability of two types points that are linearly inseparable in a low-dimensional space.However,a single kernel may not be sufficient to represent all data cases,which will lead tothe ill-posed problem of the hidden layer output matrix,causing the output weight to over-fitting.Hence,the kernel function plays a crucial role in the classification of the extremelearning machine,and an effective kernel function can better obtain the distribution information of the data and avoid data redundancy.In order to make full use of the spatial distribution information of data,we use the idea of kernel function to map hyperspectral remote sensing data into multi-layer kernel function.Therefore,using the idea of kernel function,the hyperspectral remote sensing data is mapped into the multi-layer kernel function,which makes full use of the spatial distribution information of the data and improves the classification performance.Finally,the two algorithms are applied to the hyperspectral remote sensing image dataset for experiment,and compared with other classification algorithms.It can be seen from the experimental results that the proposed algorithm significantly improves classification accuracy.