ANN Crowds: An Investigation of Individual Voices in the Crowd

Abstract This paper presents an in-depth analysis of architectural behavior within the Artificial Neural Network (ANN) Crowd, focusing on its impact on prediction accuracy. The concept of the ANN Crowd is introduced as a means to make predictions from small training set sizes. A single architecture has not proven robust enough in previous studies, so 189 different architectures are employed in the crowd, introducing diversity. Convergence is stopped early to prevent overfitting for each architecture. Therefore, the individual architectures are replicated 100 times with different initial weights. In this manner, the crowd size is increased to 18,900 total voices. Using the Architecture Behavior Interpretation scheme, this study explores how variations in architectural compositions influence the overall performance of the ANN Crowd. Employing One-Way Analysis of Variance (ANOVA) as the primary statistical tool, the research examines differences among architectural groups, maintaining a significance level of 0.05 throughout the analysis. The investigation explores two key questions: (1) whether architectural variations categorized as ‘Good’ and ‘Bad’ based on Normality, Centrality, and Width exhibit statistically significant differences in prediction accuracy, and (2), if the number of nodes within architectural groups significantly impacts the accuracy of prediction models in the ANN Crowd. The study provides valuable insights into the intricate relationship between architectural characteristics and prediction accuracy, offering implications for building ANN Ensembles and enhancing predictive modeling processes.