Training-free neural architecture search: A review

The goal of neural architecture search (NAS) is to either downsize the neural architecture and model of a deep neural network (DNN), adjust a neural architecture to improve its end result, or even speed up the whole training process. Such improvements make it possible to generate or install the model of a DNN on a small device, such as a device of internet of things or wireless sensor network. Because most NAS algorithms are time-consuming, finding out a way to reduce their computation costs has now become a critical research issue. The training-free method (also called the zero-shot learning) provides an alternative way to estimate how good a neural architecture is more efficiently during the process of NAS by using a lightweight score function instead of a general training process to avoid incurring heavy costs. This paper starts with a brief discussion of DNN and NAS, followed by a brief review of both model-dependent and model-independent training-free score functions. A brief introduction to the search algorithms and benchmarks that were widely used in a training-free NAS will also be given in this paper. The changes, potential, open issues, and future trends of this research topic are then addressed in the end of this paper.