Research Of Optimizing Neural Architecture Search In Automated Machine Learning Systems

AutoML is a method to construct a neural network automatically,which simplified human to use artificial intelligence.Neural architecture search does a main part of AutoML,which determines the layer of network use,and in which function them connected.Neural architecture search can help human to find a more efficient architecture easily.Each layer in neural network can be seen as blocks,and the searching process can be seen as stacking blocks,which similar to the process the vocabulary stacking in natural language processing.The searching process can use a seq-GAN to generate sequence with words in it,which may optimize the process by consuming less calculation resource,and make search more efficiency.Different to generating sentences,the result of NAS is a directed acyclic graph,which needs a process from sequence to tree topology and a process from tree topology to a directed acyclic graph,by mounting leaves to other nodes.After experiment,using dataset CIFAR-10,model which searched by NAS does better in accuracy than VGG-16,which is real data of seqGAN.For enhance variety of layers to trial,and reduce dependency to a trained sequence generator,searching strategy can be trained by reinforcement learning.A model can do neural architecture search networks by merging sequence generate process and tree-topology search process.After experiment,using dataset CIFAR-10,the result model does similar accuracy to the model mentioned above.