Compute-Efficient Neural Network Architecture Optimization by a Genetic Algorithm

A neural network’s topology greatly influences its generalization ability. Many approaches to topology optimization employ heuristics, for example genetic algorithms, oftentimes consuming immense computational resources. In this contribution, we present a genetic algorithm for network topology optimization which can be deployed effectively in low-resource settings. To this end, we utilize the TensorFlow framework for network training and operate with several techniques reducing the computational load. The genetic algorithm is subsequently applied to the MNIST image classification task in two different scenarios.