Re:Neural Networks and the Chomsky Hierarchy [DeepMind, 2022]
by wnkrshm on Monday January 23, 2023 @06:50AM
(#34487758)
It's often reduction that furthers understanding (e.g. in physics, to try and reduce phenomena to a small set of assumptions/laws).
Re:Neural Networks and the Chomsky Hierarchy [DeepMind, 2022]
by YeGoblynQueenne on Monday January 23, 2023 @11:11AM
(#34490630)
Absolutely, but we only understand things in the context of what we already know, and we build up our understanding of new things from that.
Besides, a theory can also guide your experiments. For example, in physics, you wouldn't try to calculate the mass of all strawberry icecream in the world because physics doesn't, as a matter of course, consider the properties of strawberry icecream as particularly special or interesting. They're not a concept of physics, that is.
That's how you probe the abilities of neural nets! Empirical evaluation done in the context of a theoretical framework. Not just poking models to see what happens [1].
Huh. No wonder Marcus Hutter's name is in the list of authors.
_____________
[1] Although poking stuff to see what happens is the root of all science, don't get me wrong.