I wrote a post asking if humans had ‘inductive biases’ in the same sense that machine learning algorithms do. My conclusion was that they do not. However, there is a different meaning to the term ‘inductive bias’ that should also be explored in relationship to human intelligence.
In a previous post, I gave the example of how children strongly tend to somehow come up with a (false) explanation that wool warms things. I asked if this was an (apparent) form of induction that my theory about induction (that ‘induction’ is always really evolutionary algorithms under the hood) couldn’t explain. Many interesting ideas were offered about how to explain this without referencing induction.
However, I wanted to mention one theory I came across that also explains it. It’s rooted in “Hebbian learning” which is the idea that neurons that fire together (i.e. at the same time) will become wired together. The idea is that this gives humans a sort of ‘bias’ towards assuming correlation is causation:
It turns out that all animals and humans have what researchers call a built-in confirmation bias. Animals and humans are wired to believe that when two things happen closely together in time it’s not an accident; instead the first event caused the second thing to happen.
For instance, if you put a pigeon in a cage with a key that lights up right before a piece of food appears, pretty soon the pigeon will start pecking the lighted key to get food. He does that because his confirmation bias leads him to believe that the first event (the key lighting up) causes the second event (the food appearing). This pigeon happens to peck the lighted key a couple of times, the food appears (because food always appears when the key is lit), and now he concludes that pecking the key when it’s lit causes food to appear.
The pigeon is acting like a person who thinks his team will win the baseball game if he’s got his lucky rabbit’s foot with him, which is why B.F. Skinner called this kind of behavior animal superstition. …
Confirmation bias is built in to animal and human brains, and it helps us learn. We learn because our default assumption is that if Event 1 is followed closely by Event 2, then Event 1 caused Event 2. … That’s why in statistics courses you have to formally teach students that a correlation isn’t automatically a cause. Our brains are wired to see correlations as causes, period. Since in real life a lot of times Event 1 does cause Event 2, confirmation bias helps us make the connection.Animals in Translation, p. 98-99
I’m not sure if I buy this theory yet. But this theory is interesting for two reasons. First, classical conditioning / Hebbian learning is a type of evolutionary algorithm of variation and selection. So there is no underlying incompatibility with Critical Rationalism. Second, it’s an idea more related to how one might make conjectures. So even if ‘correlation is causation’ is not always true, it merely has to be true often enough to be useful for animals. And in humans, it can work as a starting assumption that might later be refuted. If you do see 100 white swans and no black ones, the conjecture “all swans are white” may be false, but it’s not a bad starting conjecture.
This possibly explains the children that thought wool warms things up. Even if no one told them “wool warms things” they may have still noticed that putting on wool socks warmed their feet and then had a bias towards event 1 (putting on wool socks) caused event 2 (their feet got warmer.) On the other hand, one of the many other good proposed theories might be true and this one might be false.
So this view might still be compatible with my theory that all ‘induction’ is really evolutionary algorithms. However, it’s a new twist on that theory in that this is more related to how conjectures are made — particularly in animals, but also in humans for simple conjectures. Therefore, if it is true, it may have some relevance to AGI.