MCMC: Multi-Constrained Model Compression via One-Stage Envelope Reinforcement Learning

Area of Science:

• Artificial Intelligence
• Computer Science
• Machine Learning

Background:

• Neural networks are computationally intensive, limiting their deployment on resource-constrained edge devices.
• Existing model compression techniques often focus on single hardware objectives, proving ineffective for multi-constraint real-world applications.

Purpose of the Study:

• To develop an automated pruning method, Multi-Constrained Model Compression (MCMC), for optimizing neural networks against multiple hardware targets simultaneously.
• To minimize the impact on model accuracy while reducing latency, floating point operations (FLOPs), and memory usage.

Main Methods:

• Proposes an improved multi-objective reinforcement learning (MORL) algorithm: the one-stage envelope deep deterministic policy gradient (DDPG).
• The DDPG algorithm is adapted to determine optimal neural network pruning strategies, reducing exploration time and increasing flexibility in target priority adjustment.

Main Results:

• On VGG-16, MCMC achieved an 80% FLOPs reduction, memory savings, and acceleration with a 0.09% accuracy improvement.
• For MobileNet-V1 on ImageNet, MCMC reduced FLOPs by 50%, enhancing speed and memory compression while maintaining accuracy.
• On the JETSON XAVIER NX edge device, MCMC achieved a 71% FLOPs reduction for MobileNet-V1, improving speed, memory compression, and accuracy.

Conclusions:

• MCMC offers an effective automated solution for compressing neural networks to meet multiple hardware constraints.
• The proposed one-stage envelope DDPG algorithm enhances the efficiency and adaptability of model pruning for edge computing applications.