Hacker News Reader

Comment by sqdbps | Hacker News Reader

RE: GDPR:

"Domingos: The European Union's General Data Protection Regulation (GDPR) is putting too much value on the factor of explainability -- meaning why an algorithm decides this way rather than that way. Let's take the example of cancer research, where machine learning already plays an important role. Would I rather be diagnosed by a system that is 90 percent accurate but doesn't explain anything, or a system that is 80 percent accurate and explains things? I'd rather go for the 90 percent accurate system."

"Domingos: There is this notion predating the GDPR that data can only be used for the purpose it was collected for. This sounds plausible, but if we had been using that principle all along we would not have penicillin. We would have no X-ray. We wouldn't have all of the scientific discoveries that came unexpectedly. Serendipity, discovering new things in old data, is a huge driver in progress."

Hear, hear.

View on HN

If I'm a patient being diagnosed by a black box, I'd rather have 90% accuracy. If I'm a doctor or researcher trying to effectively treat a type of cancer, I'd much rather have the explanation.

Come to think of it, 90% is pretty low for a course of totally unnecessary chemotherapy that cold've been avoided by a human doctor noticing, as humans often do, how dumb the provided explanation was. So maybe even as a patient this isn't an obvious choice.

More generally, I'm super amused by the idea that the Master Algorithm will fall out of the brilliance at few big tech companies, but only if they could have access to my purchase history or porn viewing habits from the past decade.

If the human filtering the results improves the 80% to 85%, then it's still better to use the 90% unfiltered machine. You assume the human will actually improve the accuracy at all, not reduce it, which means he's smarter than the machine in some aspect. If the AI is very good, he might more often degrade the accuracy by rejecting things he doesn't understand or that he has an incorrect bias/opinion about.

Remember that AlphaGo game where the commentator thought it must have made a mistake because it did a move that was obviously wrong but then it surprised everyone by winning because of that?

I’d want both the 90% as well as the explanation. While some parts of these NN are black boxes, many parts are not. I’d want to know what data the model was trained on; if there were any biases in the patients tested so far; were there any issues with batch effects, et cetera...Like you said, if the question of chemo was on the line i wouldn’t want to go with just a black box.

That being said, at some point in the future, when these types of problems with NN can all be ironed out, I think going with the NN will be the obvious choice. The massive numerical complexity of cancer and the human body is too large for any human to understand and treat optimally.

NN + doctor for now; just NN in the (relatively near) future.

I think the larger point he's trying to make is that it's foolish to regulate the potential out of emerging technology (NN) and make it harder to discover new technology (the penicillin and X-ray of their time).

Reply on news.ycombinator.com