Fig. 4.

Metrics for different models trained on subsets of the Moral Machine data. (Top) Performance of initial choice models and neural network as a function of dataset size. Five bootstrapped samples were taken for every dataset size. Error bars indicate $\pm$1 SEM. (Bottom) Comparison of a choice model and a neural network before incorporating axes of differences versus after incorporating axes of differences. The addition of these features resolves much of the gap between the choice model and the neural network. Error bars indicate $\pm$1 SEM.