In order to extract perceptually relevant properties from the parameter weights of the model, we fit the kernels as various types of features using differential evolution [56], which randomly initializes sets of parameters and generates new sets by taking a set and modifying random values using the values of the other sets. If the child set has a lower error than its parent, it takes its parent’s place. The specific variation of the algorithm is termed “rand/2/bin” [1].
