Zero-cost agent evaluation method and system for neural network architecture

Abstract

The invention discloses a zero-cost agent evaluation method and system for a neural network architecture, and belongs to the field of neural network architecture search and deep learning, and the method comprises the steps: defining a path starting point to comprise an input node and a directly connected branch node and a terminal point to be an output node based on an information amount transmission step-by-step decreasing rule; performing exponential decline quantization coding on the path, and quantifying the information utilization rate of each operation in the network; setting a total architecture path information amount, a path information difference amount, a total output node receiving information amount and an output node receiving information difference amount based on the path information amount and the operation information amount, and constructing a proxy model through normalization processing and weighted summation. According to the method, architecture evaluation can be completed without actual training, direct numerical calculation of the neural network topological structure is realized, and the architecture design cost is remarkably reduced.

Images