Knowledge Redux
by JS
[My follow-up for the previous post.]
My theory of knowledge acquisition is based on two assumptions:
First, data alone is not sufficient to build belief-state/world associations. We already know that causal relationships cannot be derived from data alone (start with Pearl’s work if you want to go down this particular rabbit hole). We’d need certain causal assumptions (alt. model assumptions) which we can then use in conjunction with data to revise our belief states about the world.
Second, evolution alone is not sufficient to explain our knowledge of the world. I’m fairly certain that I don’t have the specific rules of English encoded in my DNA. I certainly have something, a kind of language schema, but English specifically, as a constraint on whatever general language learning algorithm I start with is almost certainly the product of experience over evolution.
Another argument that evolution alone is not sufficient is that evolution is a poor oracle for the acquisition of belief-state/world associations. Think about it. An agent that survives knows what it needs to know to survive, that is, the necessary belief-states align with the world in a way that enables survival. But the agent can have other belief states that are completely wrong, but that the world doesn’t care about, and so do not affect the survivability of the agent.
Or, if an agent has incorrect belief states and dies off because of them, we can’t know from observation of the death alone what belief states were incorrect. (Well, we could, because we have plenty of reasonable belief-states about this world. But another universe? Probably not so much.)
For evolution to work at all we need to assume that the world is complex enough that agents can’t survive with too many false beliefs, but not so complex that all the agents die all the time because the correct belief states are hidden in an infinite sea of incompatible belief states. We live on this fine line, but I think evolution’s solution (us) walks this line by encoding only the absolute minimum number of belief state assumptions in our genetics. We let data do the rest. Data has a habit of being accurate.
Finding computational models that start with minimal structural, cauasal, and other beliefs about worlds and can build up complex beliefs through data is precisely what I spend my days thinking about.
