Putting Deep Hybrid Learning to Work
summary
Deep Hybrid Learning combines CNN feature extraction with classical classifiers, here using a 21-layer CNN with custom incept/squeeze to produce feature vectors that feed XGBoost and RandomForest late-fusion classifiers. The pipeline applies domain preprocessing and SMOTE upsampling on a tiny, imbalanced set of 314-image dataset (262 cancer / 52 non-cancer) and reports non-overfitting training, high AUC, and classification accuracy that exceeds a 92% human benchmark; transfer learning and pure deep models perform worse and slower.
AI-written summary · about
chapters (8)
- 0:00Introduction and speaker background
- 2:07What is deep hybrid learning
- 9:33Deep learning vs machine learning trade-offs
- 14:29Why deep hybrid learning works better
- 18:22Variants of deep hybrid learning
- 22:14Case study: Ovarian cancer detection
- 34:25Results and code walkthrough
- 38:46Summary and Q&A
related talks
55mDeep Learning in Medical Image Diagnostics
24mHybrid ML Methods for the Interpretation and Integration of Heterogeneous Multimodal Data
08mDeep Transfer Learning for Computer Vision
30mTransfer learning for image recognition in healthcare...
47mInterpreting Deep Neural Networks (DNNs)
35mDeconstructing Deep Learning
45mIntroduction To Deep Learning
37mRecent Solutions Using Deep Learning Models/ The How and Why