Information theory for dimensionality reduction in dynamical systems

On compression, I’ve recently been wondering if it makes sense to argue for fitting a MaxEnt model to data and then classifying based on the model rather based on individual features.