Neural architecture search for convolution neural networks

摘要

A method, system and apparatus comprising computer programs encoded on a computer storage medium for determining a neural network architecture.One method involves using a controller neural network having controller parameters to generate a batch of output sequences in accordance with the current values of the controller parameters.This wayOutput sequence in batchGenerating an instance of a child convolution neural network (CNN) including a plurality of instances of a first convolutional cell having an architecture defined by an output sequence;Training an instance of child CNN to perform an image processing task andAssessing the performance of child CNS trained instances on tasksDetermining the performance metrics for trained instances of child CNN andIncludes adjusting the current value of the controller parameters of the controller neural network using the performance metrics for the trained CNN trained instances.