One of the characteristics of training using backpropagation is that it does not guarantee that the features extracted using a set of filters (in a convolutional layer) are different. Meaning, it is entirely possible to have multiple filters responding to very similar, if not the same, feature. Assume you are training a Neural Network which has an input layer, a single convolutional layer with K filters (f1- fk) followed by a flattening layer, a fully connected layer, softmax activation to give you a classification model. Also, assume that the network is being trained using categorical cross-entropy loss. Your task is to devise constraints and/or mechanisms to encourage the model to bring diversity (through the training process) in the sets of features extracted by the filters in the convolutional layer.