cubefox comments on AISafety.info: What are Inductive Biases?

[cubefox]

For many architectures, we know that neural networks are biased towards low-frequency functions, even at initialization.

If we can determine if a given behavior is low-frequency, then we can have some confidence that neural networks will fail to learn it by default.

These statement seem to contradict each other?

[Algon]

That’s an error. Thank you for pointing it out!

[cubefox]

Okay, that makes more sense now. So the obvious question would be about the interpretation of low/​high frequency. That’s quite a low-level concept. Does it correspond to simplicity/​complexity? Or underfitting/​overfitting?

[Algon]

I think it’s unclear what it corresponds to. I agree the concept is quite low-level. It doesn’t seem obvious to me how to build up high-level concepts from “low-frequency” building blocks and judge if the result is low-frequency or not. That’s one reason I’m not super-persuaded by Nora Belrose’ argument that deception if high-frequency, as the argument seems too vague. However, it’s not like anyone else is doing much better at the moment e.g. the claims that utility maximization has “low description length” are about as hand-wavy to me.