Design Problems in Life and AI

Abstract

This article aims to draw a connection between organismic evolution and machine learning as recursive optimization processes. Optimization of complex systems presupposes certain forms or designs of the input-output functions. Recent literatures in evolutionary developmental biology have discussed various design features of the genotype-phenotype mapping, including neardecomposability, generative entrenchment, standardization, plasticity, canalization, and scaffolding as means to solve complex adaptive problems through recursive evolution. I point out similar problems and/or techniques exist in the machine learning literature, and sketch some common features in these two distinct fields.