Growth and evolution of deep neural networks from gene regulatory networks

A simple population based Evolutionary Algorithm (EA) was used to evolve convolutional neural networks for solving an image classification problem (CIFAR10). Each member of the population was defined by a genome. This work proposes the construction of a genome based closely on the naturel world. The genes within such a genome regulate each other's expression and hence build a gene regulatory network (GRN). In the proposed approach, the genome contains no information from the problem space and could be applied to any application in principle. The genome behaves as an evolved program that grows multi-cellular organisms through a developmental process from an initial single cell. The cellular structure is an intermediate phenotype which is then mapped to its final form, a convolutional neural network in this case. The proposed GRN approach was able to evolve successful networks whose level of performance is comparable to a LeNet5 implementation.