Given some behavior, $f_i$, and set of inputs into $f_i$ that fall within the performance mode envelope, $E_i$, if we update, modify, or change $f_i$ to a different behavior $f_j$, how distinct or similar are their performances, i.e. $E_i \cap E_j$?
This becomes most clearly relevant when considering samples from a Bayesian Neural Network (BNN). When we create an ensemble of NNs by sampling from a BNN, how distinct are their performance modes? By maximizing distinctness, can we improve performance?
Given some behavior, $f_i$, and set of inputs into $f_i$ that fall within the performance mode envelope, $E_i$, if we update, modify, or change $f_i$ to a different behavior $f_j$, how distinct or similar are their performances, i.e. $E_i \cap E_j$?
This becomes most clearly relevant when considering samples from a Bayesian Neural Network (BNN). When we create an ensemble of NNs by sampling from a BNN, how distinct are their performance modes? By maximizing distinctness, can we improve performance?