Abstract:
A key challenge in neural architecture search (NAS) is quickly inferring the predictive performance of a broad spectrum of neural networks to discover statistically accurate and computationally efficient ones. We refer to this task as model performance inference (MPI). The current practice for efficient MPI is gradient-based methods that leverage the gradients of a network at initialization to infer its performance. However, existing gradient-based methods rely only on heuristic metrics and lack the necessary theoretical foundations to consolidate their designs. We propose GradSign, an accurate, simple, and flexible metric for model performance inference with theoretical insights. A key idea behind GradSign is a quantity Psi to analyze the sample-wise optimization landscape of different networks. Theoretically, we show that Psi is an upper bound for both the training and true population losses of a neural network under reasonable assumptions. However, it is computationally prohibitive to directly calculate Psi for modern neural networks. To address this challenge, we design GradSign, an accurate and simple approximation of Psi using the gradients of a network evaluated at a random initialization state. Evaluation on seven NAS benchmarks across three training datasets shows that GradSign generalizes well to real-world neural networks and consistently outperforms state-of-the-art gradient-based methods for MPI evaluated by Spearman’s Rho and Kendall’s Tau. Additionally, we have integrated GradSign into four existing NAS algorithms and show that the GradSign-assisted NAS algorithms outperform their vanilla counterparts by improving the accuracies of best-discovered networks by up to 0.3%, 1.1%, and 1.0% on three real-world tasks. Code is available at https://github.com/cmu-catalyst/GradSign

Committee:
Prof. Zhihao Jia (co-advisor)