Hybrid ML Methods for the Interpretation and Integration of Heterogeneous Multimodal Data
summary
Hybrid pipelines produce concrete gains: a monitoring system that issues an alert every 90 seconds uses Viper partitions data into low‑dimensional views and outperforms random forests and SVMs to separate artifacts and explain sensor dislodgement; an end‑to‑end model replaces softmax with backpropagation forests and reduces ImageNet top‑5 error from ~10% to <8%; ShortFuse injects 50 selected covariates into CNN/LSTM kernels and raises osteoarthritis prediction from 67% to 74% (≈2,000 subjects); weak supervision labels ~9,000 UK Biobank MRIs and flags higher major‑event risk for bicuspid aortic valve cases.
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:10Motivation
- 2:52Application: Alert Classification
- 3:58Defining interpretability
- 4:47Feature Selection, with a Twist
- 5:54VIPR-a quick overview
- 6:43Alert Classification with VIPR
- 8:25Hybrid Models
- 9:34Back-propagation Trees
- 11:17Merging Decision Forests to Networks
- 11:51ImageNet Experiment
- 13:54Effect of Structured Information
- 15:41Hybrid CNN
- 19:45Gait Kinematics
- 21:56Bicuspid Aortic Valve (BAV) Disease
- 23:27Survival Analysis
related talks
45mHybrid search: Greater than the sum of its parts?
40mPutting Deep Hybrid Learning to Work
10mML-based Log Analysis for Faster Debugging
28mLarge Multimodal Models: Understanding the State of the Art in Compute
15mMulti-Modal AI Optimization: Combining Text, Vision, and Speech Efficiently
46mML Use Cases at Think Big Analytics
10mUtilizing Word Embeddings and Gradient Boosting
37mA Practical Guide to Robust Multi modal Machine Learning in Education