EvoCNN: Evolving deep convolutional neural networks using backpropagation-assisted mutations

In this abstract we present our initial results with a novel genetic algorithms based method for evolving convolutional neural networks (CNN). Currently, standard backpropagation is the primary training method for neural networks (NN) in general, including CNNs. In the past several methods proposed using genetic algorithms (GA) for training neural networks. These methods involve representing the weights of the NN as a chromosome, creating a randomly initialized population of such chromosomes (each chromosome represents one NN), and then evolving the population by performing the steps (1) measure the fitness of each chromosome (the lower the average loss over the training set, the better), (2) select the fitter chromosomes for breeding, (3) perform crossover between the parents (randomly choose weights from the parents to create the offspring), and (4) mutate the offspring. While in smaller NNs these methods obtained results comparable with backpropagation, their results deteriorate as the size of NN grows, and are impractical for training deep neural nets. Nowadays these methods have largely been abandoned due to this inefficiency.