“Swarm Learning” ???

Recently, Gregorio has been pushing this for medical research. See “Swarm Learning, the Crypto-powered AI revolution that is kicking Cancer’s ass”.

As he defines it, this starts with Federating Learning. “an AI technique in which multiple AI models are trained independently on separate computers” and “Under FL, computers do not share any information with each other, being model weights the only known information shared, allowing for a delocalized treatment of data. However, a central coordinator monopolizes control and commercial exploitation by controlling the learning process of all models.”

He goes on to say “Thus, this is an entirely trust-based model where researchers, all of which are trying to make their particular breakthrough, have to collaborate while trusting all other actors will be fair, creating a skeptical-prone environment that disincentivizes collaboration between research teams . Now, get this federated model, put it on top of a blockchain, and you have the definitive collaboration method, Swarm Learning.”

This putting of the “federated model on top of a blockchain” is meaningless. He is using blockchain as a mantra, a chanting of a magic spell, without a deeper understanding. What he is saying is that research team A collaborates with team B where each team has its own patients and data and they somehow use a blockchain to mediate between them without sharing the underlying data. Team A would put some data on that blockchain. Team B uses that data. Team B puts some of its data on the blockchain and team A uses it. Etc.

But when collaboration happens between research teams, what is often shared is a procedure. Team A tells team B its procedure to do something. This telling can be verbally, or in email. The latter might include images of apparatus or of some clinical data. The trust that happens here is backed up by team B being able to verify team A’s method on team B’s patients.

No blockchain is used. It is not necessary. B must get the procedure in enough detail in decrypted plaintext for it to use. A blockchain is not needed for this. If the procedure can be sent via email, that’s the fastest and most efficient way to do it, and be safe against eavesdropping.

And it might not be possible to use a blockchain in a meaningful way. If clinical data is shared, the data is usually vastly greater than can be written to a blockchain. A node on a blockchain is typically 1 megabyte or so. If the data is summary data…