>>10959081The fact that ML theories try to use induction is why they suck.
Induction, i.e. gaining knowledge about the world by "extrapolating" or "generalizing" from data, is a myth because there are an infinite variety of ways to extrapolate or generalize from any set of data, and there's no principled way to pick out which one is correct. See Hume's problem of induction.
Whenever it looks like an ML model is using induction, it's really just using one particular way of extrapolating or generalizing from data out of the infinity of possible ways. But since there are so many possible ways, different ML models will extrapolate in different ways from the same data. You have to already have knowledge about which methods of extrapolation work in your circumstances for these ML models to be useful.
The way that knowledge is created in human minds, according to Popper, is by conjecture and refutation. In other words, we come up with ideas (aka conjectures), blindly, not by any process of "induction". Since we come up with ideas blindly, they will most often be wrong. So, once we've conjectured ideas, we attempt to refute them, to figure out if they're wrong. This involves comparing them to empirical evidence, or comparing them with other ideas that we have reason to accept, and looking for contradictions. When there is a contradiction between the idea and the evidence, or between the idea and some other idea that we accept, we throw the idea out and look for a new one. In this way, knowledge creation is a lot like the process of biological evolution, in that it works by random variation and then selective retention. For this reason Popper's epistemology is sometimes called "evolutionary epistemology". This approach involves no induction, and thus doesn't run into Hume's problem. Existing ML models don't work this way though, and so they suck. We'll only get to human-level AI when we implement the type of evolutionary process that happens in the human mind.
