Pareto-Frontier-aware Neural Architecture Search

1 Jan 2021  ·  Yong Guo, Yaofo Chen, Yin Zheng, Peilin Zhao, Jian Chen, Junzhou Huang, Mingkui Tan ·

Designing feasible and effective architectures is essential for deploying deep models to real-world scenarios. In practice, one has to consider multiple objectives (e.g., model performance and computational cost) and diverse constraints incurred by different computation resources. To address this, most methods seek to find promising architectures via optimizing a well pre-defined utility function. However, it is often non-trivial to design an ideal function that could well trade-off different objectives. More critically, in many real scenarios, even for the same platform, we may have different applications with various latency budgets. To find promising architectures under different budgets, existing methods may have to perform an independent search for each budget, which is very inefficient and unnecessary. Nevertheless, it would be fantastic if we can produce multiple promising architectures to fulfill each budget in the same search process. In this paper, we propose a Pareto-Frontier-aware Neural Architecture Search (PFNAS) method which seeks to learn the Pareto frontier (i.e., the set of Pareto optimal architectures) w.r.t. multiple objectives. Here, we formulate the Pareto frontier learning problem as a Markov decision process (MDP). Relied on the MDP, we transform and absorb the objectives other than model performance into the constraints. To learn the whole Pareto frontier, we propose to find a set of Pareto optimal architectures which are uniformly distributed on the range of budget to form a frontier. Based on the learned frontier, we are able to easily find multiple promising architectures to fulfill all considered constraints in the same search process. Extensive experiments on three hardware platforms (i.e., mobile, CPU, and GPU) show that the searched architectures by our PFNAS outperform the ones obtained by existing methods under different budgets.

PDF Abstract
No code implementations yet. Submit your code now

Datasets


Results from the Paper


  Submit results from this paper to get state-of-the-art GitHub badges and help the community compare results to other papers.

Methods


No methods listed for this paper. Add relevant methods here