Scholarlink

<p>In recent years, neural networks have increasingly been optimized through pruning strategies to reduce model size and inference time. A key advancement in this field is the Goal-Driven Pruning (GDP) approach, which introduced a novel top-down attention-inspired method for class-specific pruning. However, GDP requires manually freezing early layers, uses one-shot layer-wise pruning, and assumes static importance scores—factors that limit its adaptability to new architectures. In this study, we propose <i>Bidirectional-Importance Goal-Driven Pruning (BGDP)</i>, an extension of GDP that improves flexibility and automation. BGDP calculates bidirectional importance scores across layers, enabling iterative, global pruning without layer freezing. It dynamically propagates the influence of target classes throughout the network while preserving the original weights. We evaluate BGDP on three architectures: MLP (MNIST), Modified-VGG16 (CIFAR-10), and pretrained VGG16 (ImageNet). The method achieves up to <i>93% parameter reduction in MLP</i>, up to <i>44% FLOP reduction</i> on Modified-VGG16 (CIFAR-10), and up to <i>18% FLOP reduction</i> on pretrained VGG16 (ImageNet) under a <i>no-fine-tuning</i> setting, with less than <i>4% AUPRC loss</i>. As an additional study, we evaluate BGDP as a <i>post-pruning refinement</i> step on top of three structured pruning baselines (L1, FPGM, and Taylor) under the same no-fine-tuning constraint, providing further compression while meeting the same AUPRC preservation criterion. Compared to GDP, BGDP consistently enables deeper pruning and better efficiency on the evaluated MLP and VGG-style CNN models under the goal-driven, no-fine-tuning setting.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A bidirectional importance metric for goal-driven pruning

  • 期刊论文

  • 发表日期：

    2026年5月2日

查看全文
  • Mehmet Zahid Akpolat,
  • Abdullah Bulbul

摘要

In recent years, neural networks have increasingly been optimized through pruning strategies to reduce model size and inference time. A key advancement in this field is the Goal-Driven Pruning (GDP) approach, which introduced a novel top-down attention-inspired method for class-specific pruning. However, GDP requires manually freezing early layers, uses one-shot layer-wise pruning, and assumes static importance scores—factors that limit its adaptability to new architectures. In this study, we propose Bidirectional-Importance Goal-Driven Pruning (BGDP), an extension of GDP that improves flexibility and automation. BGDP calculates bidirectional importance scores across layers, enabling iterative, global pruning without layer freezing. It dynamically propagates the influence of target classes throughout the network while preserving the original weights. We evaluate BGDP on three architectures: MLP (MNIST), Modified-VGG16 (CIFAR-10), and pretrained VGG16 (ImageNet). The method achieves up to 93% parameter reduction in MLP, up to 44% FLOP reduction on Modified-VGG16 (CIFAR-10), and up to 18% FLOP reduction on pretrained VGG16 (ImageNet) under a no-fine-tuning setting, with less than 4% AUPRC loss. As an additional study, we evaluate BGDP as a post-pruning refinement step on top of three structured pruning baselines (L1, FPGM, and Taylor) under the same no-fine-tuning constraint, providing further compression while meeting the same AUPRC preservation criterion. Compared to GDP, BGDP consistently enables deeper pruning and better efficiency on the evaluated MLP and VGG-style CNN models under the goal-driven, no-fine-tuning setting.