In other news: Ford Motor Co. has filed a patent application for technology that would allow autonomous vehicles to repossess themselves if drivers miss payments.
https://thehill.com/homenews/3880884-future-fords-could-repo...
A model that scans for high risk patients, then passes that into another model which takes actions, which can also be fed data from the police and mass surveillance with that it can also eliminate coverage for "anti-AI" people.
All it has to do is just recognize Anti-AI behavior as a health or security risk, that's it. There was an anti-crime AI that flagged Congress members too, so its not far fetched.
For "something" you can substitute drugs, or social media, porn, or calling others idiots on the Internet.
Wait a minute...
Grow up. this meatgrinder didnt just condense from aether. Someone wrote this, someone can change it, and anyone in the git log can be held accountable for what they did at any time. Enough of this late-stage capitalist attempt to offshore even the very concept of accountability.
Uh, not really, no. The whole point of neural nets is they independently discover novel and unexpected ways of predicting the probable answer.
Someone chose the features, the model architecture, the training data, and SGD parameters; it might not even be a NN (and probably isn't) — it could be a traditional linear regression model, or gradient-boosted decision trees. Someone analyzed the model performance on a test set and maybe even ran simulations to see how effective the model was at learning a known functional form. Someone made reports on that performance and included estimates like "probability of falsely denying claims on average" broken down by 10-year age bracket. An intern might have helped. A team lead coordinated meetings with whatever VP was running the project and signed off on the model. Senior management budgeted months worth of person-hours on model development and integration/deployment into production systems. And the model could have been rejected or modified at any point. The model training code is probably sitting on someone's laptop right now, and absolutely can be revised at any point in time, either slightly or significantly.
This is no magic in something like this. There are no positronic brains to wonder about. It's just a fucking best-fit line, and someone had to both draw the damn line and sign off on it.
Now if you could only make an algorithm that would learn how to tune algorithms...
That exists! https://www.automl.org/