A Novel Training Protocol for Performance Predictors of Evolutionary Neural Architecture Search Algorithms

Evolutionary neural architecture search (ENAS) can automatically design the architectures of deep neural networks (DNNs) using evolutionary computation algorithms. However, most ENAS algorithms require an intensive computational resource, which is not necessarily available to the users interested. Performance predictors are a type of regression models which can assist to accomplish the search, while without exerting much computational resource. Despite various performance predictors have been designed, they employ the same training protocol to build the regression models: 1) sampling a set of DNNs with performance as the training dataset; 2) training the model with the mean square error criterion; and 3) predicting the performance of DNNs newly generated during the ENAS. In this article, we point out that the three steps constituting the training protocol are not well thought-out through intuitive and illustrative examples. Furthermore, we propose a new training protocol to address these issues, consisting of designing a pairwise ranking indicator to construct the training target, proposing to use the logistic regression to fit the training samples, and developing a differential method to build the training instances. To verify the effectiveness of the proposed training protocol, four widely used regression models in the field of machine learning have been chosen to perform the comparisons on two benchmark datasets. The experimental results of all the comparisons demonstrate that the proposed training protocol can significantly improve the performance prediction accuracy against the traditional training protocols.

I. Introduction

Deep neural networks (DNNs) are becoming the dominant algorithm of machine learning [1], largely owing to their superiority in solving challenging real-world applications [2], [3]. Generally, the performance of DNNs relies on two deciding factors: 1) the architectures of the DNNs and 2) the weights associated with the architecture. The performance of a DNN in solving the corresponding problem can be promising, only when its architecture and the weights achieve the optimum combination simultaneously. Commonly, when the architecture of a DNN is determined, the optimal weights can be obtained through formulizing the loss as a continuous function, and then the exact optimization algorithms are employed for solving. In practice, the gradient-based optimization algorithms are the most popular ones in addressing the loss function, although they cannot theoretically guarantee the global optimum [4]. On the other hand, obtaining the optimal architectures is not a trivial task because the architectures cannot be directly optimized as the weights do. In practice, most, if not all, prevalent state-of-the-art DNN architectures are manually designed based on extensive human expertise, including ResNet [5], DenseNet [6], and among others.