FIGURE 4.

Shown, is the log scaled histogram of final model weights of the first layer of a randomly selected model within cross-validation. As seen, the vast majority of weights have converged to values close to {0, 1} in response to the proposed entropy weight regularization. As mentioned, a comparatively small set of weights (an order of magnitude less than successfully separated) show a non-perfect separation towards either extreme. Further analysis of said weights indicate belonging to specific features, contributing to the ultimate variance between each ensemble, as highlighted in Fig. 6.