Fig. 3: Statistical violin plots of all hybrid and classical networks trained.

a Annotated violins of the 120-parameter hybrid network and 124-parameter classical network. The shaded blue and orange region shows the region of well-trained networks that scored better than 72% (accuracy of linear support vector machine model). The point on each violin represents the average accuracy of these well-trained networks, with error bars denoting one standard deviation from the mean. The red-shaded region represents poorly trained networks, while networks below the solid red line failed to optimize entirely. The white line is networks A and B seen in Fig. 2 and Fig. 1a, b. b Violin plot of all networks trained in this study. Each violin represents a statistical fit (kernel density estimation) of all networks with a given number of parameters, where each gray horizontal line marks the maximum accuracy achieved by an individual network during training. The line is the mean of the upper of the well-trained networks, and the shaded region is the standard deviation of that mean. c Bar plot showing the percentage of networks which failed to beat the non-probabilistic linear model (<72% accuracy, poorly trained).