Research On Novel HTM Based On Dynamic Optimization

Hierarchical Temporal Memory(HTM)is a neural network model that mimics the structure and function of the neocortex of the human brain.It enables learning and prediction of sequential data by simulating the properties of the cerebral cortex,such as hierarchical structure,spatial and temporal representations,and pattern memory.Compared with other artificial neural network models,HTM can effectively handle sparse and high-dimensional nonlinear data with high fault tolerance and robustness,and has achieved remarkable results in various fields such as data prediction,anomaly detection,and pattern recognition.This thesis analyzes the current status of HTM research at home and abroad in recent years,and finds that some problems still exist in sequence learning and prediction of HTM: when HTM learns data spatial features,it generally sets fixed model parameters,with large training overhead and unstable spatial representation;when learning sequence timing dependence,it does not fully exploit the characteristics of timing data,and the prediction accuracy is low;in addition,the existing HTM decoders require additional neural network learning,which has a large time overhead.To address these problems,this thesis proposes a new HTM model based on dynamic optimization,improves the three core modules of HTM spatial pooler,temporal memory and decoder respectively,designs a data-driven dynamic spatial pooler algorithm,a dynamic temporal memory algorithm integrating increment and error,and an overlap valuebased HTM decoding method,and optimizes the training of the model using the training sample distribution and long-term trend of the sequence,so that the model The model is able to use training resources in a dynamic and balanced manner,exploit the sequence data itself more fully to explore the temporal correlation,and reduce the model training time overhead while improving the prediction performance.The main contents of this thesis include:1.Aiming at the problems of fixed number of columns,poor expand-ability and unstable data representation in the training process of HTM spatial pooler,this thesis proposes a datadriven dynamic spatial pooler algorithm.By introducing a data-driven dynamic column expansion strategy and a load-bearing degree based spatial pooler algorithm,the spatial pooler can adaptively adjust the number of learned columns according to the training sample distribution,and effectively select and activate columns according to their characterization ability.In this thesis,experimental tests are conducted on synthetic sine,combined sine,logistic mapping data and real-world power load sequence data,and the results show that the algorithm can effectively improve the stability of the Spatial Pooler column representation of HTM and improve the performance of the model.2.Aiming at the problem that HTM does not fully consider the characteristics of the data itself when dealing with long-term and complex sequences leading to poor prediction performance,a dynamic time memory algorithm that incorporates incremental and error is proposed.On the basis of learning serial temporal correlations,the global trend changes of serial data are further considered,and the trend changes are encoded as incremental information,while synaptic connections between neurons and input sample encoding and incremental encoding are established,so that neurons learn more accurate serial temporal characteristics based on the information of these two dimensions;in addition,drawing on the idea of supervised learning error back propagation,the existing neuron is improved Hebbian learning rules,use the error information of the prediction results to optimize the training process of the Temporary Memory,and dynamically adjust the learning step of the synaptic persistence value of neurons according to the prediction error.This thesis is tested on three artificial datasets and one real-world datasets,and the results show that the algorithm can improve the prediction accuracy of HTM.3.Aiming at the problem that existing HTM decoders require a large time overhead for network training,this thesis proposes an efficient HTM decoding method based on overlapping values.The method can decode the prediction results quickly and accurately by comparing and matching the learned information of the spatial pooler and the temporal memory for output decoding.The method first uses the historical active column information of the HTM spatial pooler to quickly filter the set of columns containing predictive state neurons based on the column overlap values;then,further designs a synaptic connectionbased HTM decoding algorithm based on the synaptic persistence values of the temporal memory neurons to accurately determine the prediction results in the filtered set of columns.In this thesis,tests are conducted on synthetic and real scene datasets,and the results show that the method can effectively improve the decoding performance of HTM.