ELLMM: Daniel Grimmer (Yale)

Location: 451 College St, room B04

Title: Evolutionary Epistemology In Silico: Exploring Nature and Nurture with Neural Networks

Abstract: 

The recent engineering success of neural networks is often seen as favoring Empiricism over Nativism. Indeed, Empiricist accounts of epistemology (i.e., tabula rasa initialization followed by general-purpose learning from massive amounts of sensory data) align closely with the standard way of training neural networks. Crucially, however, this training regime imposes a broadly Empiricist epistemology by engineering fiat, rather than fairly testing it against a Nativist alternative. This talk will demonstrate how this issue can be remedied with a change of training regime; an Evolutionary Meta-Learning framework allows one to use neural networks to efficiently simulate the evolutionary pressures that shaped the human mind.

For instance, one could subject an LLM to fitness pressures that reward rapid language acquisition (from relatively few training examples). I will argue that over many generations such a network would evolve a (broadly Chomskyian) language acquition device, i.e., a highly structured initialization which already contains deep grammatical structure and thereby speeds up the learning process. Such an evolved LLM would constitute a strong response to Chomsky's "False Promise of Chat-GPT" article. Indeed, when appropriately trained LLMs can teach us a great deal about human language use and acquisition.

Time permitting, I will briefly mention how this Evolutionary Meta-Learning framework allows us to investigate a wide variety of topics within Evolutionary Epistemology. Is Konrad Lorenz right that "The a priori of the individual is the a posteriori of the species"? If so, what evolutionary forces shaped our (broadly Kantian) intuitions about space, time, and causation? Does evolution give us any good reason to believe that our cognitive maps correspond to the noumenal territory (e.g., otherwise we would have gone extinct)? These questions are all in-principle answerable by conducting in silico simulation of cognitive evolution.