I know this is totally obvious to the researchers but AI - perhaps more accurately “insanely large differentiable functions that one can maximise complex optimisation functions for eye wateringly big data” - is just an excellent fit for biology - molecular and organismal scale
@ewanbirney
So speciation is just biology overfitting lots of models to local problems?
@darachm @ewanbirney The model fitting in this case would equate to selection - but that is not what drives novelty in evolution. The ‘new features’ come from remixing parental genomes and from mutation. The models just work with that material.
@rupertoverall @ewanbirney Right so running that backwards then, since we know that the specific architectures in biology can generate complexity neutrally, what kind of analogous Weirdness do you find in neural net gradient descents?
Or is it all removed because you're putting selection at the scale of each node and not on the level of each network?
But that's also multicell centric - even we still have cell-level selection (immune, cancer)...
just wondering for rhyme of bio in the modeling!
@rupertoverall @ewanbirney lol that's one way to put it 😂
That implies some interesting things of "overfitting"
Can overfitting drive the emergence of complexity, in cryptic function outside of original constraints? What architectures and permit that?