Evolutionary Meta-Learning in Neural Networks as a Neutral Testing Ground for Nativism and Empiricism

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. We argue that this issue can be remedied with a change of training regime; our Evolutionary Meta-Learning framework allows one to use neural networks to efficiently simulate the evolutionary pressures that shaped the human mind. We present a first-principles derivation showing that the dynamics of a Darwinian Lineage Simulation (DLS) are formally equivalent to a noisy Stochastic Gradient Ascent (SGA) up a log-fitness landscape. Furthermore, we demonstrate that the Baldwin Effect—modeled here as an evolutionary pressure towards rapid learning so as to minimize childhood mortality—is well captured by the MAML++ meta-objective function. This framework offers a neutral testing ground for Nativist and Empiricist adaptation strategies while also aligning both approaches with Sutton's Bitter Lesson. Without hand-coding or brute stipulation, adaptations are chosen by the evolutionary process itself as it navigates its way up an ever-changing fitness landscape.