Architecture Selection Method Based on Densely Connected Neural Architecture Search

Designing the network structure in deep learning is time-consuming and trial-and-error. The purpose of neural architecture search is to let the machine search for a suitable network structure by itself. Currently, neural architecture search relaxes the search space to a continuous differentiable space, so we can use stochastic gradient descent for optimization. It significantly reduces the search time and dramatically improves the application value of neural architecture search. A densely connected neural architecture search is an improvement in neural architecture search. The input of each layer of its network consists of the outputs of all the layers before it. Moreover, the results of each layer of the network will be the output to all subsequent layers. Therefore, it can search out the subnet structure with better performance. We propose improvements to the densely connected neural architecture search algorithm, addressing the deficiencies in architecture selection and search strategies. We verify the effectiveness of the proposed algorithm through experiments.