EPruner

EPruner prunes convolutional neural networks by selecting adaptive exemplar filters using Affinity Propagation on weight matrices to reduce FLOPs and parameter count while preserving accuracy.

Key Features:

• Adaptive Exemplar Filters: Uses Affinity Propagation (message-passing) on weight matrices to determine an adaptive number of exemplars that are preserved as filters.
• Automatic Pruning: Identifies important filters without using training data by preserving exemplar filters and removing others.
• Computational Efficiency: Provides a CPU implementation reported to be an order of magnitude faster than GPU-based state-of-the-art methods.
• Improved Initialization: Uses exemplar weights to initialize pruned networks for fine-tuning, potentially improving post-pruning performance.

Scientific Applications:

• Model compression and acceleration: Reduces computational complexity and parameter count while maintaining or improving model accuracy through exemplar-based filter selection.
• VGGNet-16 on CIFAR-10: Achieves a 76.34% reduction in FLOPs, removes 88.80% of parameters, and improves accuracy by 0.06%.
• ResNet-152 on ILSVRC-2012: Achieves a 65.12% reduction in FLOPs, removes 64.18% of parameters, with a 0.71% top-5 accuracy loss.

Methodology:

Applies Affinity Propagation (message-passing) to weight matrices to select adaptive exemplar filters and prunes non-exemplar filters, enabling pruning without training data and implemented for CPU execution.