Automatically Learn Cost-constrained Convolutional Neural Network Architectures With Reinforcement Learning

In recent years,deep neural networks have achieved subversive success in congnitive tasks(such as speech recognition,image recognition and machine translation)that are difficult to solve by traditional methods.The architecture of the network has a decisive influence on the performance.In recent years,researches in the field of computer vision have focused on the designing of network architecture for different tasks.At present,the designing of the neural network architecture is mainly done manually.It requires the user to have relevant professional knowledge and the designing process is very time consuming.Recently,some automated network architecture design methods have been proposed.However,these methods only consider the prediction accuracy of the designed network architecture.In practical applications,in addition to predicting accuracy,other types of cost are also very important,such as parameter size,for example,for embedded devices with limited memory,the size of the network parameters that can be tolerated is often limited.Therefore,it is of great significance to study the automated network architecture designing method under budgeted cost constraints.There is a method called BSN(Budgeted Super Network)for automated convolutional network architecture designing under budgeted cost constraints,but this method cannot optimize the cost of training time which can only be known after training the sampled network architecture.In view of the shortcoming of BSN,On the basis of current reinforcement learning based automated network architecture design methods,we propose a method called B-ENAS(Budgeted-Efficient Neural Architecture Search).We select three representative types of cost,parameter size,inference time and training time to do experiments on the CIFAR10 data set.In the experiments under the parameter size and the inference time our method B-ENAS can learn network architectures that satisfying different cost constraints like the previous method BSN,and the prediction accuracy of learned network architectures in our methods is higher.In the experiment under the training time,it is proved that the B-ENAS can optimize the training time,which cannot be optimized by BSN.