Towards Understanding Transfer Learning with Applications to Medical Imaging
summary
Paper "Understanding Transfer Learning for Medical Imaging" evaluates ImageNet pretraining vs random init on chest X-rays and diabetic retinopathy scans (~200k images), comparing large ImageNet architectures (ResNet/Inception) to tiny CBR convnets (≈1/8–1/20 size). Key findings: transfer often matches random init for final accuracy, ImageNet model quality does not predict medical transfer, representational canonical correlation analysis shows different internal features, and most speedup comes from weight scaling — a mean-var init (random weights scaled to pretrained per-layer stats) reproduces the faster convergence.
AI-written summary · about
chapters (12)
- 0:00Intro
- 3:01Transfer in Deep Learning Applications
- 4:37Github for Transfer Learning
- 11:05Do better ImageNet models transfer better? (Komblith, Shlens, Le), 2019
- 16:54Performance Results on Chest X-rays
- 17:12Evaluating Transfer: Main Takeaways
- 19:51Going Beyond Performance Evaluations
- 21:21CCA for Feature Similarity
- 23:08Similarity of Deep Representations
- 27:44Feature Similarity in Transfer with CCA Compare feature similarity of
- 35:19Feature Similarity and Reuse
- 40:17Open Questions
related talks
30mTransfer learning for image recognition in healthcare...
55mDeep Learning in Medical Image Diagnostics
56mTutorial on Deep Learning II
25mDeep Learning in Biomedicine: Challenges and Opportunities
52mTransfer Learning and Finetuning Deep Convolution Neural Network
42mDeep Learning for Computer Vision
50mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
28mMachine Learning & Healthcare