Are you using DLT as an additional technology for securing data and enhancing trust in Mellody?
We are using a DLT to enhance trust in MELLODDY. The distributed ledger stores non-sensitive metadata that are necessary to do the orchestration of computations on the distributed dataset and to have a ful traceability of operations done within the platform. These non-sensitive metadata could also be stored in a central service instead of using a DLT. The use of the DLT removes the need to trust an actor operating the central service, who could modify the non-sensitive metadata on the central service.
Are there classes of algorithms that are suitable or unsuitable for this type of model transfer?
Sequential learning algorithms are required to do federated learning, neural networks are thus the most natural class of algorithms.
For mellody what kind of representations are used is this descriptor/fingerprint/graph based or?
Fingerprints are used.
For Melloddy, How the tasks of the models are defined? How you ensure that the data of each partner can be used for the same task?
Each partner defines the tasks. Multi-partners multi-task learning makes it possible to avoid alignment on tasks