Beyond Backbone Backpropagation: A Decoupled Strategy for Efficient Transfer Learning

Short summary

This paper proposes a lightweight transfer learning strategy that adapts only normalization layers and decouples feature extraction from classifier optimization, precomputing features once to avoid end-to-end backpropagation. A margin-based weighted loss in the classifier head reduces ambiguity. Evaluated across seven CNN and Transformer architectures on three medical datasets, the approach significantly cuts training time and CO2 emissions with only marginal accuracy trade-offs, often matching baseline performance.

• •Decouples feature extraction from classifier training to avoid costly end-to-end backpropagation
• •Tested on 7 architectures across 3 medical imaging datasets with competitive accuracy
• •Reduces training time and CO2 by orders of magnitude, suited for resource-constrained settings

Generated with AI, which can make mistakes.