A multi-agent reinforcement learning based approach for automatic filter pruning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep Convolutional Neural Networks (DCNNs) face deployment challenges on resource-constrained devices due to high computational and memory demands.
• Network pruning is a key technique for compressing DCNNs, with reinforcement learning (RL) offering adaptive strategies over rule-based methods.
• Existing RL pruning methods often use a single agent, neglecting inter-layer dependencies and varying sensitivities within DCNNs.

Purpose of the Study:

• To propose an automatic filter pruning method, QMIX_FP, utilizing a multi-agent reinforcement learning algorithm (QMIX).
• To model Deep Convolutional Neural Networks (DCNNs) as a multi-agent system, accounting for layer-specific sensitivities and interactions.
• To enhance model compression and enable efficient deployment of DCNNs on resource-constrained hardware.

Main Methods:

• Developed QMIX_FP, a novel automatic filter pruning approach based on the QMIX multi-agent reinforcement learning algorithm.
• Modeled the multi-layer structure of DCNNs as a multi-agent system to capture layer interactions and sensitivities.
• Incorporated knowledge distillation for fine-tuning pruned networks to accelerate performance recovery.

Main Results:

• Demonstrated the effectiveness of QMIX_FP on benchmark DCNNs (VGG-16, AlexNet) using CIFAR-10 and CIFAR-100 datasets.
• Achieved significant reductions in computational and memory requirements for the pruned networks.
• Maintained network accuracy post-pruning, validating the method's efficacy.

Conclusions:

• QMIX_FP offers an advanced solution for Deep Convolutional Neural Network (DCNN) model compression.
• The multi-agent approach effectively addresses layer interactions, leading to optimized filter pruning strategies.
• This method facilitates the efficient deployment of DCNNs on devices with limited resources without compromising performance.