Study On Learning Algorithms For Augmented Quaternion Extreme Learning Machine

Machine learning is one of the hot topics in the field of artificial intelligence,and neural network is one of the core research topic of machine learning.Neural networks have been widely used in many fields such as pattern recognition,data mining and computer vision,and further benefited many aspects of our daily life.Due to extremely fast learning speed and good generalization ability,the ELM model becomes an important neural network model.The latest research of statistical signal processing reveals that the learning ability of many neural network models can be further improved by representing the three-dimensional or four-dimensional signals with quaternion variables.However,the research on extreme learning machine is mainly limited to real field and complex field.When it comes to deal with three-dimensional or four-dimensional data,there are some limitations for traditional ELM models.To this end,this paper extends the extreme learning machine to quaternion domain,and proposes several learning algorithm for quaternion extreme learning machine based on the latest research development in quaternion signal processing.The main work of this thesis can be listed as the following aspects:(1)In order to effectively deal with three-dimensional and four-dimensional data,the quaternion extreme learning machine(QELM)model is proposed by extending the extreme learning machine from the real domain,where two computation methods for generalized inverse matrix are used.Moreover,in order to avoid the possible over-fitting phenomenon,a regularized QELM model is also proposed.(2)In order to fully capture the second-order statistical characteristics of the input signal and thus effectively process the non-circular signals,the involution information of quaternion input signals and hidden layer are respectively incorporated to propose four augmented quaternion extreme learning machine models.At the same time,the approximation capabilities of the proposed models are theoretically analyzed.(3)In order to meet the demands of real-time data processing,an online sequential algorithm for the augmented quaternion extreme learning machine model is proposed.(4)The above proposed learning algorithm are applied to several practical problems such as chaotic time series prediction,wind prediction and face recognition to compare their performance with the traditional real-valued extreme learning machine.The superiority of the proposed algorithms is verified.