Learning from Data/Observations is Overrated

Original article was published on Artificial Intelligence on Medium

Learning from Data/Observations is Overrated

The title might shock some readers, although I hope not. In fact, I am shocked that I have to convince ‘scientists’ that the statement in the title is a truism.

To be sure there are few things we do learn from observations (from data, or experience, etc.) but not everything — and thus the word ‘overrated’ in the title. And here’s the proof that the statement of the title is true (is a fact).

Claim:
Not everything we come to know is learned from data (experience, or observation). In fact, most of what we know we ‘learn by instruction’ (or ‘by being told’).

Proof (by negation):

  1. Suppose that everything we know we learn from experience (from observations and from data).
  2. Growing up, different people have different experiences and they make different observations — in fact, it is rarely the case that two people have exactly the same experiences.
  3. By (1) and (2), we can say that, in theory, there is nothing that all people should agree on and know exactly the same way. (note, this is not saying there is nothing that different people would agree on, but that there’s nothing that different people should agree on!)
  4. But we know (3) is false, since no 2 human beings can — even of they wanted to — disagree on many facts (a small sample of facts that no two humans can disagree on is shown in Figure 1 below)
  5. But if (3) is false, then either (1) is false or (2) is false (or of course both). But we know (2) is true, so (1) must be false — that is, there are many things that cannot be (and are not) learned by observation/data because if they were different people might learn them differently but that cannot be the case.

End of proof.

Figure 1. Examples of things we come to know not by observations (or experience, or from data) but we come to know because they are true — usually, we learn these by instruction, or by being told, etc. Some of these are physical facts, some are mathematical facts, some are just commonsense facts, etc.

I really feel silly writing this article which states a scientific fact, but the ML runaway train has made many ‘scientists’ make some silly statements (e.g., ‘we can learn anything/everything from data’). And, unfortunately, some of these ‘scientists’ are award-wining scientists that are negatively effecting tens of thousands of young scientists and engineers.

Some logic, and some commonsense, goes a long way. So please, tone down your excitement — it will prove to be very useful in your journey.