Notes

on different stuff

---

Project maintained by dg1an3 Hosted on GitHub Pages — Theme by mattgraham

Federated learning pt I

One of the recurring themes at CMIMI were attempts at implemented federated or distributed learning. The motivation was simple: need lots of data, but HIPAA. Therefore let’s learn without giving up data. Different techniques: Coefficient averaging | Round robin | Gradient contrib Gradient contrib is reminiscent of the computational argument…

But it also raised the possibility of “hyper-federated” learning.

• Estonia coordinates all health data access through a distributed ledger, with revocation by the patient
• FHIRChain is an example of a model that combines block chain with a common interop standard, FHIR, to achieve distributed tumor board
• Proof of interop is another example, in this case with proof of work being replaced by a (somewhat ill-defined) proof of interop.

Which raises the question what other architectural problems can be solved like federated learning?