Design And Implementation Of Automated Machine Learning System Based On NAS

With the development of machine learning in recent years,its complexity is increasing.If we rely on human beings to design rule algorithm to make the computer run according to the rules,it will consume a lot of human resources.The concept of automation is also developing with the times.The former definition of automation is to let machines replace people to operate.With the development of information technology,the concept of automation is also considered to replace not only human physical work,but also mental work.And the automatic machine learning system can well solve the above requirements.Most of the existing automatic machine learning systems are based on traditional machine learning methods,and different models are selected by Bayesian algorithm or evolutionary algorithm.There is a lack of one-stop automatic machine learning system based on deep learning.At the same time,the existing neural network architecture search algorithm ignores the support of incremental learning.In view of the above problems,this thesis proposes an incremental neural network architecture search algorithm based on knowledge distillation,which fills the blank of neural network search algorithm in this field,and solves the problem of forgetting the old sample data in the process of learning new samples.The model can avoid "catastrophic forgetting" of old sample data.The algorithm model proposed in this thesis can perform well on the old data without the old training data.This thesis designs and implements an automatic machine learning system based on neural network architecture search.The system includes four main modules:Data Management Center,model center,storage and automatic machine learning algorithm.Users access the service through the website to manage the model and data.When the user submits the training request,the algorithm model is trained asynchronously.After the training is completed,it will be concluded the results are returned to the front end and displayed to the user.Storage module provides stable and reliable support for user training data.Through the comparative test and system test,the incremental neural network architecture search algorithm based on knowledge distillation proposed in this thesis achieves the effect.The result is slightly inferior to the joint training of new and old data,and is obviously better than the ordinary model which only trains new data.The system has complete functions,high availability and robustness.