RADARS: Memory Efficient Reinforcement Learning Aided Differentiable Neural Architecture Search

TL;DR

RADARS is a novel framework combining reinforcement learning and differentiable neural architecture search to efficiently explore large search spaces with limited memory, achieving higher accuracy and faster search times.

Contribution

RADARS introduces a scalable RL-aided DNAS framework that prunes architecture candidates to enable large search space exploration within bounded memory constraints.

Findings

Achieves up to 3.41% higher accuracy on CIFAR-10 and ImageNet.

Reduces search time by 2.5 times compared to state-of-the-art RL methods.

Handles large search spaces with limited GPU memory.

Abstract

Differentiable neural architecture search (DNAS) is known for its capacity in the automatic generation of superior neural networks. However, DNAS based methods suffer from memory usage explosion when the search space expands, which may prevent them from running successfully on even advanced GPU platforms. On the other hand, reinforcement learning (RL) based methods, while being memory efficient, are extremely time-consuming. Combining the advantages of both types of methods, this paper presents RADARS, a scalable RL-aided DNAS framework that can explore large search spaces in a fast and memory-efficient manner. RADARS iteratively applies RL to prune undesired architecture candidates and identifies a promising subspace to carry out DNAS. Experiments using a workstation with 12 GB GPU memory show that on CIFAR-10 and ImageNet datasets, RADARS can achieve up to 3.41% higher accuracy with 2.5X search time reduction compared with a state-of-the-art RL-based method, while the two DNAS baselines cannot complete due to excessive memory usage or search time. To the best of the authors' knowledge, this is the first DNAS framework that can handle large search spaces with bounded memory usage.