[Paper Review] ADC: Automated Deep Compression and Acceleration with Reinforcement Learning.

Yihui He, Song Han|arXiv (Cornell University)|Feb 10, 2018

Advanced Neural Network Applications36 references59 citations

TL;DR

This paper proposes Automated Deep Compression (ADC), a reinforcement learning-based method that automates neural network model compression for mobile deployment. By efficiently exploring the accuracy-compression-speed trade-off space, ADC achieves state-of-the-art results—2.7% higher accuracy than hand-crafted methods at 4× FLOPs reduction on VGG-16, and 2× FLOPs reduction with 1.65× speedup on MobileNet with minimal accuracy loss.

ABSTRACT

Model compression is an effective technique facilitating the deployment of neural network models on mobile devices that have limited computation resources and a tight power budget. However, conventional model compression techniques use hand-crafted features and require domain experts to explore the large design space trading off model size, speed, and accuracy, which is usually sub-optimal and time-consuming. In this paper, we propose Automated Deep Compression (ADC) that leverages reinforcement learning in order to efficiently sample the design space and greatly improve the model compression quality. We achieved state-of-the-art model compression results in a fully automated way without any human efforts. Under 4x FLOPs reduction, we achieved 2.7% better accuracy than hand-crafted model compression method for VGG-16 on ImageNet. We applied this automated, push-the-button compression pipeline to MobileNet and achieved a 2x reduction in FLOPs, and a speedup of 1.49x on Titan Xp and 1.65x on an Android phone (Samsung Galaxy S7), with negligible loss of accuracy.

Motivation & Objective

• To address the inefficiency and sub-optimality of hand-crafted model compression techniques that require expert tuning.
• To automate the exploration of the large design space balancing model size, inference speed, and accuracy.
• To enable fully automated, push-button compression without human intervention for mobile deployment.
• To improve compression quality beyond what is achievable with conventional, expert-driven methods.

Proposed method

• ADC employs reinforcement learning to search the model compression space, treating compression decisions as sequential actions.
• The agent learns to select optimal compression operations such as channel pruning, knowledge distillation, and quantization.
• The state space encodes current model characteristics (e.g., FLOPs, accuracy), and the action space includes compression operations.
• The reward function is designed to maximize accuracy while minimizing FLOPs and model size, enabling end-to-end optimization.
• The method is trained end-to-end using policy gradient methods to optimize the compression policy.
• The pipeline is fully automated, requiring no manual tuning or domain expertise.

Experimental results

Research questions

• RQ1 Can reinforcement learning effectively automate the model compression process without human intervention?
• RQ2 Does an RL-based approach outperform hand-crafted compression techniques in terms of accuracy and efficiency?
• RQ3 Can the method generalize across different architectures like VGG-16 and MobileNet?
• RQ4 To what extent can the method reduce FLOPs while maintaining high accuracy on mobile platforms?

Key findings

• Under 4× FLOPs reduction, ADC achieved 2.7% higher top-1 accuracy than a hand-crafted compression method on VGG-16 using ImageNet.
• For MobileNet, ADC achieved a 2× reduction in FLOPs with only a negligible drop in accuracy.
• On a Titan Xp GPU, the compressed MobileNet model achieved a 1.49× speedup.
• On an Android phone (Samsung Galaxy S7), the model achieved a 1.65× speedup with minimal accuracy degradation.
• The method demonstrated state-of-the-art compression quality in a fully automated manner.
• The approach eliminated the need for expert tuning, enabling a push-the-button compression pipeline.