A Cross-Dataset Evaluation of Genetically Evolved Neural Network Architectures

The design of deep neural networks is often colloquially described as an `art.' Although there are some common, guiding principles such as using convolutions for data with spatial locality or using recurrence for data with temporal characteristics, neural network architectures tend to be manually engineered. Few works currently provide methods to determine optimal architectures and hyper parameters. In this work, we generate empirical evidence for neural network architecture choices. We use a genetic algorithm to evolve 980 neural networks for a variety of common data sets. We analyze the characteristics of the highest performing architectures, compare those traits across data sets, and present a set of generalizable neural network design patterns.

Author

• Gelman, Ben

• Neural networks
• Deep learning
• Genetic algorithms
• Data

Identifier

hdl:1920/11479

OAI identifier oai:identifier

oai:MARS:1920/11479