Visualization System for Evolutionary Neural Networks for Deep Learning

• Our visualization system consists of three major views: a lineage view, a fitness-parameter view, and a network architecture view. The three views visualize different features of the evolutionary process. This allows the results of evolutionary computation to be visualized from different perspectives. Also, the views are connected to each other and provide interactive analysis features.
• The lineage view shows parent-child relationships in the evolution. In this view, we utilize a forced-directed layout to show the lineage, where nodes represent parent/child networks and edges represent their relationships.  The central idea of the lineage view is to provide an immediate intuitive understanding of the dynamics of the whole evolutionary process by combining other features of evolutionary computation. The dynamics of evolution are represented as an ancestry tree while we integrate other features into the tree.
• The fitness-parameter view shows the changes in fitness and hyper-parameters of the networks as evolution progresses. The view allows the users to interactively explore and compare how fitness and selected hyperparameter values of networks change over the course of evolution. We need an efficient method to find an appropriate set of hyper-parameters in searching a high-dimensional hyper-parameter space. When we use the approach based on evolution, an analogy between fitness and hyper-parameters is helpful in finding appropriate hyperparameter sets.
• The network architecture view shows how the layers of each network are stacked and how the network architectures converged as evolution progresses. The major goal of the network architecture view is to explore evolution of the network architectures—how the layers of each network are arranged and how the network architectures are converged over evolution progress. In this view, users can focus on network architectures rather the hyperparameters for layers. The search space of the genetic algorithm includes the arrangement of layers within a network as well. The network architecture is one of the most important aspects that affect the performance of convolutional neural network models.
• The views are tightly connected to each other and provide interactive user interfaces for analysis. The system allows users to link over the three views in order to understand the evolutionary process better through interacting with the presented data through different representations. The interaction techniques of the views, such as exploration, filter, selection, and reconfigure, overcome the limitation of static representations and support to uncover insights.

Citation and DOI

Junghoon Chae, Catherine D. Schuman, Steven R. Young, J. Travis Johnston, Derek C. Rose, Robert M. Patton, Thomas E. Potok. "Visualization System for Evolutionary Neural Networks for Deep Learning" IEEE Big Data 2019.