Improving the sample-efficiency of neural architecture search with reinforcement learning

TL;DR

This paper enhances neural architecture search efficiency by replacing the traditional REINFORCE algorithm with PPO in a reinforcement learning framework, leveraging parameter sharing to reduce training time and improve stability.

Contribution

It introduces the use of PPO for training the controller in ENAS, improving stability and speed over the original REINFORCE method in NAS.

Findings

PPO accelerates NAS training compared to REINFORCE.

Parameter sharing reduces computational resources needed.

Enhanced stability observed during controller training.

Abstract

Designing complex architectures has been an essential cogwheel in the revolution deep learning has brought about in the past decade. When solving difficult problems in a datadriven manner, a well-tried approach is to take an architecture discovered by renowned deep learning scientists as a basis (e.g. Inception) and try to apply it to a specific problem. This might be sufficient, but as of now, achieving very high accuracy on a complex or yet unsolved task requires the knowledge of highly-trained deep learning experts. In this work, we would like to contribute to the area of Automated Machine Learning (AutoML), specifically Neural Architecture Search (NAS), which intends to make deep learning methods available for a wider range of society by designing neural topologies automatically. Although several different approaches exist (e.g. gradient-based or evolutionary algorithms), our focus is on one of the most promising research directions, reinforcement learning. In this scenario, a recurrent neural network (controller) is trained to create problem-specific neural network architectures (child). The validation accuracies of the child networks serve as a reward signal for training the controller with reinforcement learning. The basis of our proposed work is Efficient Neural Architecture Search (ENAS), where parameter sharing is applied among the child networks. ENAS, like many other RL-based algorithms, emphasize the learning of child networks as increasing their convergence result in a denser reward signal for the controller, therefore significantly reducing training times. The controller was originally trained with REINFORCE. In our research, we propose to modify this to a more modern and complex algorithm, PPO, which has demonstrated to be faster and more stable in other environments. Then, we briefly discuss and evaluate our results.