CNNPruner: Pruning Convolutional Neural Networks with Visual Analytics.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) excel in computer vision but their large size hinders deployment on resource-constrained devices.
• Model pruning aims to reduce CNN size by removing redundant neurons and fine-tuning, but current methods lack flexibility in balancing efficiency and accuracy.
• Existing automated pruning lacks transparency due to the complex interaction between pruning and fine-tuning stages, complicating optimization.

Purpose of the Study:

• To introduce CNNPruner, a visual analytics approach for optimizing Convolutional Neural Network (CNN) pruning.
• To enable users to interactively create pruning plans balancing model size and accuracy goals.
• To enhance understanding of filter importance and refine pruning strategies through visualization.

Main Methods:

• CNNPruner assesses convolutional filter importance using both instability and sensitivity metrics.
• It integrates advanced filter visualization techniques for user comprehension.
• Users can interactively define pruning plans based on desired model size or accuracy targets.

Main Results:

• CNNPruner provides a flexible and transparent method for pruning large-scale CNNs.
• Case studies demonstrate the effectiveness of CNNPruner in optimizing the trade-off between CNN efficiency and accuracy.
• The visual analytics approach facilitates informed decision-making in the model pruning process.

Conclusions:

• CNNPruner effectively addresses the limitations of existing automated CNN pruning methods.
• The visual analytics framework empowers users to achieve desired efficiency-accuracy balances in pruned CNNs.
• This approach facilitates the deployment of large CNN models on devices with limited computational resources.