Ep. 81: Federated Learning with DLT – Mathieu Galtier (CPO Owkin)
Mathieu Galtier, PhD – Neuroscience and computer science (applied mathematics) background. Chief Product Officer of Owkin which was founded in 2016 and has quickly emerged as a leader in bringing AI and ML technologies to the healthcare industry. We’ll be talking about how blockchain technology and federated learning systems are creating amazing new opportunities in medical research and development.
Guest Twitter: https://twitter.com/m_galtier
• Origins of Owkin and the healthcare interest
– life science company – research oriented – really want to cure cancer
o Medical data & Clinical predictive models
o Connect platform for Federated Learning
• Owkin Tech stack overview
o Data prep
o Infra: Kubernetes
o Application layer: Substra (python, Hyperledger fabric)
o Algorithms: any python ML frameworks (pytorch, tensorflow etc)
• open source foundation to host Substra framework for distributed computation
ML for Healthcare
• Why is ML relevant for Healthcare?
o Precision medicine: customization of treatment to individual patients based on their medical data.
o• Clinical Trials Research challenges and opportunities
oFederated Learning for Healthcare
• What is Federated Learning?
o Model training across distributed datasets. Data is never shared. (opposition centralized computation vs traveling models)
o Can you describe how federated learning can solve some of our privacy and security concerns with personal data?
o HealthChain project
o Melloddy project
• What is the HealthChain project’s main objective?
• MELLODDY consortium – Owkin, NVIDIA, and King’s College London – Have you reached your goals? What’s new with this collaboration?
• Academic collaboration → HealthChain
o Industrial coopetition → Melloddy
• BLockchain in Federated Learning
• Why is a blockchain relevant for Federated Learning? What is in the blockchain?
o Trustless permission regime → smart contracts for details governance / data access regimes
o Traceability & reproducibility (no 3rd party to trust) → good for certification
o Decentralized orchestration → increase security; no shadow computation
o Tokenization of data value through contribution score for training models
o Private vs public blockchains: we needed a consortium blockchain
o How can blockchain-orchestrated federated learning enable a new method of building AI models?
o Models have an identity card describing their genealogy. Their performance can be assessed on 3rd party datasets → models are production / certification ready
o Models are the most accurate possible. They serve as the basis (for transfer learning) for numerous data modalities: histology, genomics, chemistry…
•What do you believe in that most people would disagree with?
o There is no free will
• Favorite book that has influenced you
o Foundation by Isaac Asimov. Psychohistory: 2 scales individual and population.
• Keys to ML in Healthcare
Occording to sources, Microsoft has been working on a decentralized identity solution since 2017 and has slowly built out the infrastructure over the past few years.. At the recent Ignite conference, Microsoft announced that it will launch a public preview of its “Azure Active Directory verifiable credentials” spring of 2021.
In the NHS pilot, for example, health care providers can request access to professional certifications from existing NHS health care workers, who can in turn choose to allow that access, streamlining a process for transferring to another facility that previously required a much more involved back and forth. U
Health Unchained Links