Is there any research into ways to enforce feature selection across classes by network structure?
Given the number of parameters in NN, even convnets are prone to over fitting.
I'm curious if there has been research a way to structure a network which makes the same features be learned across a given class?
L2 regularization and dropout sort of attempts this but doesn't quite achieve the desired results.
I'm more interested in network structure which penalizes different feature combinations for samples within a given class.
In other words, my hypothesis is each class has a distinctive feature set that's approximately in all samples of a given class. I am interested in ways to structure a network to enforce the learning of that feature over learning the features of individual samples.