Polyploidism in Deep Neural Networks: m-Parent Evolutionary Synthesis of Deep Neural Networks in Varying Population Sizes

Abstract

Evolutionary deep intelligence was recently proposed to organically
produce highly efficient deep neural network architectures
over successive generations. Thus far, current evolutionary synthesis
processes are based on asexual reproduction, i.e., offspring
neural networks are synthesized stochastically from a single parent
network. In this study, we investigate the effects of m-parent
sexual evolutionary synthesis (m = 1, 2, 3, 5) in combination with
varying population sizes of three, five, and eight synthesized networks
per generation. Experimental results were obtained using
a 10% subset of the MNIST handwritten digits dataset, and show
that increasing the number of parent networks results in improved
architectural efficiency of the synthesized networks (approximately
150x synaptic efficiency and approximately 42–49x cluster efficiency)
while resulting in only a 2–3% drop in testing accuracy.