Direct Link to File. 674 words, 3 minute read, 2 paperback pages

M.,

I want to inquire with you over a potential rebuttal to the evolutionary argument against naturalism – or, more specifically, your question of “why can primates do physics?” If humans were ‘mere beasts’ who could only act on their genetic programming, then the ‘problem’ of why we should be able to understand physics would be compelling.

However, this is not the case. The evolutionary survival strategy employed by humans favors the ability to learn and adapt (to the environment and to each other), via the pattern-matching faculties of sophisticated neural networks. This is why it takes humans relatively long to reach maturity in comparison to other animals: we need training. A fruit fly, or a fish, or a snake, all emerge ready to perform their needed tasks as if on auto-pilot. Humans, rather, observe and predict – it is what the neurons in our brains just do1. Thus, it does not seem unreasonable that humans with the ability to learn via observation and prediction should be able to understand phenomena far removed from our ancestors’ hunter-gatherer existence.

I imagine that the reply might be, “Yes, but even that adaptivity only needed (i.e. was conditioned by evolution) to be good enough to learn in the physical domain of basic survival.” But…my research involves general networks with the ability to learn “arbitrary” representations, purely conditioned on their “experience,” whether that be text, or images, or audio. The neurons in our brains are way better than that.

Ok but why, you might ask, would evolution favor the development of such a powerful – and metabolically expensive – apparatus for learning and adaptivity? Surely having a little less such capacity would not adversely affect the basic survival aspects of hunting and gathering? Is it then just an accident of mutation that we ended up with “way more” capacity for learning than is necessitated by typical changes in the natural environment? Have I not simply replaced the question of “Why should we be able to do physics?” with the question of “Why should we be so good at learning (general) things?”

Perhaps, but beyond the natural environment, what if we include competition among human-like entities? Then there does seem (to me) to be good reason those that are a little ‘smarter’ than others would be more likely to survive, to be more resourceful (and perhaps even more vicious! :-/ ) than their cousins, and thus this ability to learn could confer an evolutionary advantage, and the ‘arms race of intelligence’ could produce highly adaptive brain structures able to (in ML lingo) generalize powerfully to increasingly arbitrary new scenarios, including the regularity of the physical world on large and small scales.

Your thoughts?

(PS- A possible rebuttal to my rebuttal: In the neural network systems that human researchers employ now, these networks’ very ability learn and adapt is because they were designed with architectures that favor the kinds of learning they need to do, and there are typically trade-offs made between the “inductive bias” of certain structures – which can learn with relatively little data but aren’t that “general” – and truly generic architectures which may be highly adaptable but are too…“unwieldy” let’s say…to learn efficiently. And even the best neural networks systems designed by humans can’t hold a candle to even simple biological brains. How could Nature have sooo optimized these neural architectures?

Counterpoint to what I just said: there are powerful neuro-evolution systems that learn the architectures too. The “NEAT” algorithm is pretty amazing, in fact.)

References

  1. “Ah, but are there not neurons in fish and snakes? Why is it humans’ neurons ‘observe and predict’ but snakes’ don’t?” …They do, but it’s (at least) a difference of degree. (It may also be a difference of kind.) For snakes or other animals, there’s just a lot less capacity to predict and adapt. “Do you know this for sure or are you just talking out of your ass, Scott?” Mostly the latter right now. ;-) ↩