Research On Novel Efficient Logical Layered HTM

Hierarchical Temporal Memory(HTM)is a neural network model based on the structure and functionality of the human brain cortex,capable of learning spatial and temporal features of data,making it suitable for modeling and predicting time series data.The HTM model consists of two core algorithms: the Spatial Pooler(SP)and the Temporal Memory(TM).The Spatial Pooler extracts spatial features from the data and transforms them into Sparse Distributed Representations(SDR),while the Temporal Memory learns temporal dependencies between sequential data and makes predictions about the future.But HTM still has challenges for predicting complex time series data,such as high computational cost of the Spatial Pooler and unstable spatial features generated,as well as low accuracy of the Temporal Memory in prediction.In this thesis,we propose a novel and efficient logical-layered HTM model to improve the efficiency and reduce the prediction error of the HTM model.The main contents of this thesis are as follows:1.Analyzing the existing structure and operation mechanism of HTM,and based on the current research status in domestic and foreign literature,summarizing the key areas for improvement in HTM: Firstly,the training efficiency of HTM Spatial Pooler is low,as it requires a significant amount of comparison and multiple iterative calculations to generate spatial SDR from data,resulting in high time cost,and the inhibition rule and boosting mechanism in HTM may lead to unstable SDR;Secondly,during the learning of temporal SDR,HTM adopts a pattern matching approach,which requires strict matching of the current sequence context with the historical sequence information in memory.When data is disturbed,the prediction accuracy of the model is relatively low.To address these issues,this thesis proposes a novel and efficient logical-layered HTM model structure.2.To address the issues of low training efficiency and unstable representation results in HTM Spatial Pooler,a new algorithm called SP＿HHG(A Spatial Pooler Algorithm Based on Heterogeneous Hash Group)is proposed.In order to reduce the time overhead of Spatial Pooler while meeting the basic requirements of SDR representation,three hash function groups are designed to generate spatial SDR from input data.The first two hash function groups are based on Locality Sensitive Hashing(LSH)functions with different sensitive fields,while the third hash function group uses random hashing.The combined values of the three hash function groups are used to activate the corresponding columns in the Spatial Pooler.Experimental results on the Weather Humidity Air(WHA),Solar Power(So P),and Sinusoidal(Sin)datasets show that compared to traditional HTM,SP＿HHG reduces the time overhead by 98.7%,98.8%,and 98.4% on the three datasets respectively,and SP＿HHG generates stable internal representations.This indicates that SP＿HHG effectively reduces the time overhead of HTM Spatial Pooler and improves the stability of data representation.3.To address the issue of low prediction accuracy in HTM Temporal Memory when data is perturbed and current temporal patterns cannot be effectively matched with learned historical information,a new algorithm called TM＿LL(A Temporal Memory Algorithm Based on Logical Layer)is proposed.In order to improve the matching capability of temporal patterns in data,a logical-layered abstraction approach for temporal context is designed.The active columns representing spatial features of data are divided into logical layers,and different layers of historical active cells are connected based on the layer of columns they belong to,in order to capture logical-layered abstraction contextual information and minimize the temporal and spatial costs of logical-layered model training.The algorithm is tested on real datasets such as ETTm1,ETTm2,as well as artificial datasets with perturbations.The results show that compared to traditional HTM model and current mainstream time series prediction models,TM＿LL algorithm significantly reduces the prediction error rate and achieves similar prediction accuracy as the double-layer HTM algorithm with less time overhead.4.Efficient logical-layered HTM prototype for time series prediction(ELLHTM-TSP)is designed and implemented based on the content described above.The system’s various functionalities are tested on datasets including So P,WHA,NAB,and NYC-Taxi.The experimental results show that in the application of time series prediction tasks,ELLHTM-TSP significantly reduces the time overhead compared to traditional HTM,and the prediction accuracy is improved by 3%,3%,17%,and 28% respectively for the So P,WHA,NAB,and NYC-Taxi datasets.These results indicate that the logical-layered HTM system based on the improved spatial pooler and temporal memory algorithms can effectively enhance the prediction efficiency and accuracy of time series data.