Strengthen The Application Research Of Hierarchical Sequential Memory Algorithm In Tax Forecast Model

The main purpose of intelligent tax data prediction research is to promote the further development of tax prediction in practical applications by standardizing and normalizing the processing of tax data and modeling the structure and computational characteristics of the neocortex.Based on the analysis of the current research status,the research on tax data prediction has achieved some results,but it is still insufficient in terms of the accuracy of the prediction results and the time cost of the prediction model.This thesis mainly studies the application of Hierarchical Temporal Memory(HTM)in the intelligent prediction of tax data.By improving the spatial pooler and temporal memory algorithm of HTM,it optimizes the HTM's accuracy and training efficiency in the task of tax data prediction..This thesis analyzes and compares various tax forecasting methods,uses real tax data in a certain city to train HTM,replaces HTM neurons with gated recurrent units,and uses multi-threading technology to improve the HTM spatial pooler algorithm.The corresponding forecasting model of the tax forecasting system is implemented.The research content includes:(1)A data prediction method based on HTM is proposed to build a model of tax data prediction.This method studies the tax data of a city and establishes the forecast model of the tax data.The experiment shows that the average accuracy of the prediction results constructed by this method is 75%,and the average training time of the algorithm is 22.1s.(2)HTM's own structure is improved,HTM neurons are modified,combined with gated circulation unit or its variants,and cyclic feedback mechanism is introduced to construct gated circulation unit neurons to replace the neurons in HTM.On this basis,new HTM structure is obtained.Compared with the existing HTM prediction model,the average accuracy of this improved method is increased by 6.39%.(3)The HTM spatial pooler training method is improved.Aiming at the low training efficiency of the HTM spatial pooler algorithm,the lack of parallel computing and the big data problems in the HTM spatial pooler.Based on the independent concurrent working mechanism in the biological neuron,this thesis design and realize the HTM spatial pooler using multi-threading.It reduces the time overhead of the spatial pooler in the HTM,improve the balance of the distribution of the active microcolumns in the HTM in the spatial pooler,and enhance the adaptability of the algorithm.The average training time is 16.6s,a decrease of 24.88%.The average accuracy of the improved method is 80.6%,an increase of7.47%.(4)Combined with the algorithm in this thesis,the tax prediction system is designed and implemented.The system can predict the future tax change trend according to external factors such as policies and give corresponding measures according to different results while satisfying the forecast tax result.