1

With the advances in machine learning, data mining and process of big data, soon machines will find patterns that could either interpreted as causation or correlation while we have no idea of the underlying mechanisms.

At the time falsifiability is the most accepted concept to differentiate science and pseudoscience by ranking high risk taking theories above the others.

According to falsifiability we need data to falsify the theory (mostly happens in natural science) which is already based on vast amount of data processed by AI.

The idea doesn't seem secure as past since data gathering and processing costs has changed a lot from Karl Popper days.

Is this going to obsolete falsifiability in the future or where I am wrong in my assumptions?

| improve this question | |
  • 2
    As you can see from the article you linked, falsifiability is specific to Popper and not "most accepted". Although some vaguer versions like "testability" are indeed used to distinguished science, falsifiability as such is mostly of historical interest only. However, I do not see how "we need data to falsify a theory" and "data gathering and processing costs" are relevant here. We need data to do just about anything, and costs are not specific to its use for testing or falsification. – Conifold Aug 8 '18 at 0:56
2

With the advances in machine learning, data mining and process of big data, soon machines will find patterns that could either interpreted as causation or correlation while we have no idea of the underlying mechanisms.

If you don't know what mechanism causes a correlation, then you can't interpret the pattern as causation.

At the time falsifiability is the most accepted concept to differentiate science and pseudoscience by ranking high risk taking theories above the others.

According to falsifiability we need data to falsify the theory (mostly happens in natural science) which is already based on vast amount of data processed by AI.

There is no way to derive a theory from data. Any set of data are compatible with an infinite set of mathematical equations describing the data. Nor is any set of data equivalent to a theory since a theory is about the underlying reality not the data.

If you find some correlation you can guess that there is some underlying causal mechanism and guess about that mechanism, but guessing is not derivation. Once you have guessed some causal mechanism you can test the guessed mechanism by doing more observations: machine learning may or may not be useful during this process.

If you want to understand Popper's position, see the material in this list:

http://fallibleideas.com/books#popper

| improve this answer | |
1

I don't see why falsifiability would go away ... any theory an AI comes up should still be falsifiable to be a good theory. In fact, falsifiability may become more important than ever: once we trust an AI to the point where the AI's claims and judgements are deemed correct exactly because "the AI says so", we're in real trouble. So, we should make sure that there will be independent means to test those theories, claims, and judgments.

| improve this answer | |
  • Consider AI comes up with something like this exaggerated example with a p < 0.05 based on large amount of data gathered from years. That takes a long time to be falsified if ever. Do we need something more relying on explanation of mechanisms besides falsifiability? – Xaqron Aug 8 '18 at 0:47
  • @Xaqron I think we have always had something to refute something like that: common sense! (which largely is about underlyng mechanisms). But I suppose your question then is: will common sense still be applicable as AI's will generate a flood of such theories, some of which have much greater initial plausibility? Well, good question ... maybe our methods of putting these theories to the test will requite other AIs ... but testing we will, and testing we must. – Bram28 Aug 8 '18 at 1:01
  • You right but someone just need right numbers to publish a paper in medical journals. In face recognition we even don't know how ML classifies faces and don't care. But statistics shows that machine even does that better than human. We are loosing insight and this is gonna accumulate in the future. – Xaqron Aug 8 '18 at 1:09
  • @Xaqron Yes, I agree! – Bram28 Aug 8 '18 at 11:27
0

I don't think falsifiabilty (or theory in general) will become obsolete. Machine learning is still subject to the laws of statistics. It is perhaps true that in some restricted domains with lots of data, we will increasingly rely on difficult-to-interpret models (e.g., artificial neural networks) instead of simpler, interpretable models. But in other domains, we will not have enough data, or will want to generalize to different cases where we don't have data. Relatedly, in many contexts we want to do more than predict future data from the same process; we want to have an understanding of the underlying mechanism, for example in order to predict what happens if we change the mechanism, how we might improve the mechanism or use it in other contexts, etc. See a related discussion here.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.