An Extreme Learning Machine based on Quantum Particle Swarm Optimization and its application in handwritten numeral recognition

The Extreme Learning Machine algorithm was proposed by Prof. Guangbin Huang in 2004. It is a single hidden layer feedforward neural network. It has attracted extensive research of many scholars because of its fast speed, simple implementation and good generalization performance. In this paper, Quantum Particle Swarm Optimization was introduced to extreme learning machine to solve the problem of complex network structure which is caused by random assignments to the input weights and biases of hidden nodes. The QPSO is used in the process to select the input weights and biases instead of random assignment. Then extreme learning machine uses the result produced by QPSO to train the network. Thus can improve the prediction accuracy and response speed to unknown data and gain a more compact network structure. The proposed method is used in handwritten numeral recognition application in the end. And it gets an approving performance.

I Introduce

The Extreme Learning Machine is a single hidden layer feedforward neural network which was proposed by Prof. Guangbin Huang in Nanyang Technological University in 2004[1]. Compared to the traditional gradient-based learning algorithms, it has faster learning speed, better generalization performance and simpler implementation. It also can avoid trapping in the local optimal value. These all because that it randomly assign values to the input weights and biases of hidden nodes, calculates the output weights by using the Moore Processor generalized inverse of hidden output weights. It needs computation only once. ELM is widely used in classification and regression. It is found that the extreme learning machine have a more complex network structure and slower response speed than BP algorithm because it assigns the values of input weights and biases randomly [2].