When federated learning is a scam.
Today I’m going to do something that the light people will really like. You know, people with no substance.
The ones who are on social media giving likes and throwing milk jugs over their heads for the challenge of the week.
The ones who say “in my opinion” before speaking.
All of them are going to love it.
Because I’m going to give “value” content.
Oh yes, valueeeeeeeeeee.
And these people love giving value. It pleases them.
The value thing is in federated learning.
Which consists, in case you don’t know, and you don’t have to, in that you put the machine learning to learn from local data, without taking the data to the cloud.
And you take the learned algorithm, but without taking the data with you.
And people obviously like that.
But I’m going to tell you something.
I’m sure no one else is going to tell you. Really, no one.
Just for being on this list, which is free, you get it.
Although for teaching the things I teach here I could charge thousands of euros at any business school.
The thing is that federated learning is ONLY valid if the data from that place of origin is already clean, well structured, modeled and standardized.
Otherwise, no.
In other words, it works very well for making AI that learns radiology, because there are very standardized formats.
But it does not work, for example, for Natural Language Processing.
Because everyone writes as they please.
And for that reason, whether you like it or not, the text has to be taken out to clean it up.
For this reason, the safest way to work with clinical texts is not to federate, which I have already told you is not possible, but to sign contracts that say what you are going to do with this data.
This is a case that is not solved with technology, but with regulation.
And that’s it.
Come on, don’t get ripped off.
Have a nice day.
PS. To do a study extracting information from medical records, using the cloud, but a cloud managed and controlled by the owner of the information, here.