3 ms·
> "But I already know who @horse_js is, and it's not [...]!" Perhaps. The data here is not 100% conclusive. There are some critical assumptions holding up our
by xg15 8y ago
> "But I already know who @horse_js is, and it's not [...]!"
Perhaps. The data here is not 100% conclusive. There are some critical assumptions holding up our conclusion and [...] has never confirmed (or denied) our findings.
Perhaps the horse lives to tweet another day...
Ironically this highlights one of the main problems with how machine learning is used.
On a very high level, I think you can sum up machine learning algorithms as finding pattern in enormous heaps of noisy data ("training") then trying to apply the discovered pattern to novel data and using the result to guess the answer to a question you posed ("predicting").
The keyword being guess here. Unlike algorithms not based on learning, there is no guarantee that the answer is correct, because you usually don't know if the training data you supplied was sufficient or if the learned patterns were the ones you need. If you knew, you could just hard code the patterns directly and get rid of the whole learning overhead altogether.
Researchers know and communicate this. However, in the press, "AI" seems to be seen as almost the exact opposite: Not only can those fantasy AI systems answer questions about fuzzy human concepts with the precision of a computer, their answers are even better than the human ones - which is why the things we need to worry about are ethics discussions and humanity becoming obsolete...
This could be funny if it were just restricted to science fiction and public discussion, but it becomes problematic when "AI" systems are used to make life-changing descisions like setting insurance premiums or declaring persons suspicious to law enforcement.
- _underfl0w_ 8y ago> it becomes problematic when "AI" systems are used to make life-changing descisions like setting insurance premiums or declaring persons suspicious to law enforcement. Hasn't the latter already happened? I'm without link/source, but I seem to recall reading about there being tests of using a homanoid-looking AI-driven "attendant" at a border somewhere that would judge people based on looks/temperament and try to guess if they were lying about what's in their luggage.
- hobofan 8y agoYes, discriminating machine learning is commonplace, but often hard to uncover, as it's not often obvious to the users of automated systems how those values are constructed. Luckily some critical parts like issurance calculation is regulated (in some parts of the world) to have the requirement of explainable algorithms to prevent this kind of discrimination, so it's not as bleak as it's often made out to be. Of course it's also important that it stays that way.
- CorvusCrypto 8y agoI don't see that as a major problem. In fact most life-changing decisions are already based on probability. Insurance companies already do risk analysis and whether the algorithm uses ML or basic statistics, there is a threshold level of confidence used. I'd also argue it's how our brains work. Many times as we come to a decision we are going off of confidence, not true correctness. I'm the case of declaring suspicious person's, well by definition they are suspects based on confidence, not by truth. Even in court we determine verdicts based on human probabilistic confidence that comes from the evidence.