An accuracy-aware extension to LRP-based pruning for CNNs to prevent cascading accuracy degradation in data-scarce transfer learning

Researchers propose an accuracy-aware pruning mechanism for CNNs that improves upon existing Layer-wise Relevance Propagation (LRP) methods to reduce model size without degrading performance in transfer learning scenarios with limited data. The approach dynamically adjusts pruning rates using harmonic mean of class accuracy, achieving 15% improvement in compression efficiency while maintaining task-specific accuracy.

This research addresses a practical challenge in machine learning deployment: optimizing pre-trained neural networks for resource-constrained environments while preserving classification performance. The problem emerges when practitioners use fixed-weight feature extractors from large datasets like ImageNet on domain-specific tasks with scarce training data. Existing filters that contributed to ImageNet performance become redundant for specialized tasks, bloating models unnecessarily.

The technical innovation centers on preventing cascading accuracy degradation—a phenomenon where removing seemingly low-relevance filters causes unexpected downstream accuracy drops through layer interactions. By introducing dynamic pruning rate adjustment guided by class-wise accuracy metrics, the method maintains balanced performance across all output classes rather than optimizing global metrics that can mask minority class degradation.

For practitioners building AI systems on edge devices or bandwidth-constrained environments, this represents meaningful progress toward efficient transfer learning. The 15% improvement in the area-under-curve metric suggests substantially better compression-accuracy tradeoffs, reducing deployment friction for domain-specific applications in medical imaging, autonomous systems, or industrial inspection where training data remains limited.

The broader impact extends to making specialized AI models more accessible. Smaller models require less computational resources, enabling deployment on mobile devices and embedded systems. As organizations increasingly adopt transfer learning for rapid model development, efficient compression techniques become critical infrastructure for practical AI deployment. The research validates that principled approaches to pruning, informed by model relevance analysis, outperform simpler heuristics.

• →New accuracy-aware pruning mechanism improves LRP-based filter compression by 15% in transfer learning scenarios
• →Harmonic mean of class accuracy prevents cascading degradation by maintaining balanced performance across output classes
• →Method enables model compression while preserving task-specific performance in data-scarce environments
• →Results reduce computational overhead for deploying pre-trained networks on resource-constrained devices
• →Technique addresses practical bottleneck in specialized AI deployment across edge computing and mobile applications