Research And Implementation Of Machine Learning Algorithm Based On Multi-Model Fusion

Relying on the sufficient collection and learning of the associated information among multiple tasks,as well as the expansion of the model induction and inference information carried by the negative labels generated by the multi-model processing data,the multi-task learning model generated by the fusion of multiple models has good generalization ability.The network structure of the target multi-task model generated by multi-model fusion and the training mechanism adopted in the multi-model fusion process are of great significance to the reachable accuracy of the generated multi-task model and the total amount of spatiotemporal resources consumed in the fusion process.In the setting of network parameters in the multi-task model structure with hard parameter sharing mode,due to ignoring the judgment of multiple models on the affinity of multiple tasks,excessive setting of shared parameters leads to the introduction of extra noise,which causes the reduction of model inference accuracy,or excessive setting of task-specific parameters leads to the multiplication of the total number of model parameters,which causes excessive resource consumption.At the same time,the training objectives and training datasets in the training mechanism are usually fixed,so the influence of different training contents on the training effect of models with different degrees of maturity is ignored,resulting in a longer training time and slightly effective phenomenon.The whole thesis is roughly divided into three parts.Firstly,the multitask learning technology used in the process of multi-model fusion and the application project background are briefly introduced.Secondly,the structure decision method of multi-task model generated by multi-model fusion and the training strategy adopted in the process of multi-model fusion are analyzed and expounded.In the structure decision part,based on the inference and understanding of multiple models for the sample datasets of multiple tasks,the evaluation rules of inter-task affinity are defined,so as to design and realize the demarcation method of the target multi-task model generated by multi-model fusion to determine the network structure of the model.In the training strategy part,the distillation mechanism of"teacher-student" structure is adopted and proposes a phased progressive training strategy to reduce resource consumption by utilizing the advantages of multiple models guiding multi-task learning.Meanwhile,combining with the inference results of the trained model in different training rounds,a dynamic updating method of samples participating in training is proposed and its effectiveness is verified.Finally,the target model structure’s decision method and training strategy proposed in this paper for multi-model fusion are applied to the automatic evaluation project of student sports test video.According to multiple sports test models on the server side,the network structure of multi-task model generated by multi-model fusion is decided.The multiple models on the server side are distilled into a lightweight multi-task model by using the training strategy.And realize the automatic evaluation task of student sports test video on mobile terminal to verify the feasibility of the proposed algorithm in practical application scenarios.