Split Learning versus Federated Learning for Data Transparent ML, Camera Culture Group, MIT Media Lab
7 likes27,460 views
Ramesh Raskar from MIT Media Lab discusses the importance of utilizing 'small data' in health technology while maintaining patient privacy and facilitating societal benefits. The document explores various techniques for training machine learning models without sharing raw patient data, including federated learning and split learning. It emphasizes the need for innovative strategies to manage data friction and enhance data utilization for population health.
Split Learning versus Federated Learning for Data Transparent ML, Camera Culture Group, MIT Media Lab
- 1. Raskar, Camera Culture, MIT Media Lab Camera Culture Ramesh Raskar Data Transparent ML + Health Privacy vs Societal Benefits Training NN without Raw Data Ramesh Raskar Associate Professor MIT
- 2. Pooling ‘small’ data No Exchange of Raw Patient Data
- 3. Technologies Individual Organization Population Physical Diagnostic Devices EHR DigitalGlobal Pop-health Health Vault Health Data Market Health OS Removing Data Friction New Capture + Analysis Patient Scale World Scale Existing Data + Techniques
- 4. ‘Invisible’ Health Data ‘Small Data’ ‘Small Data’‘Small Data’
- 5. a. Distributed Data b. Patient privacy c. Incentives d. ML Expertise e. Labeling Low Bandwidth Low Compute ‘Small’ Data ML for Health Data
- 6. Gupta, Raskar ‘Distributed training of deep neural network over several agents’, 2017 No Exchange of Raw Patient Data Train Neural Nets
- 7. Training Deep Networks No sharing of Raw data Server Client Invisible Data / Data Friction
- 8. Ease Incentive Trust Regulation Ayushman Blockchain AI/ SplitNN Overcoming Data Friction
- 9. Automating ML : AI building AI Published in International Conference on Learning Representations (2017) Otkrist Gupta, Baker, Naik, Raskar, ICLR 2017 Teacher Student
- 11. Protect Raw Data Share Wisdom Data Utility Train Model EncryptSmashObfuscate Add Noise Private Data Protect Infer Statistics Anonymize Share Wisdom, Not Data
- 12. Federated Learning Nets trained at Clients Merged at Server Differential Privacy Obfuscate with noise Hide unique samples Homomorphic Encryption Basic Math over Encrypted Data (+, x) Split Learning (MIT) Nets split over network Trained at both
- 14. Partial Leakage Differential Privacy Homomorphic Encryption Oblivious Transfer, Garbled Circuits Federated Learning Split Learning Protect data Distributed Training Inference but no training Praneeth Vepakomma, Tristan Swedish, Otkrist Gupta, Abhi Dubey, Raskar 2018
- 16. Technologies Individual Organization Population Physical Diagnostic Devices EHR DigitalGlobal Pop-health Health Vault Health Data Market Health OS Removing Data Friction New Capture + Analysis Patient Scale World Scale Existing Data + Techniques