The notion of a learned feature is not entirely straightforward with a tool like Polimago. Polimago at least in in terms of how it works - is more similar to a neural network than it is to a tool like e.g. Minos and ShapeFinder. Minos and ShapeFinder are using feature lists that are verified versus image content, so you can - in principle - tell which features have been found and which features haven’t (both tools in fact don’t collect that information but that is for performance reasons) and your classifier operates with a definitive list of features it is looking for.
This is not the case for Polimago: Polimago uses the result of the MRF sequence as the input vector to a classification learning engine based on Tikhonov’s regulratization, and the classification engine then tries to approximate a function mapping an input vector to a result scalar (the output of which is your classification result) using the training data. This function effectively becomes your classifier (in the case of Polimago this would be a high dimension linear function; in deep learning you’d have a trained neural network here). At no point Polimago is using a list of features, and there is no definitive way to tell what precisely it is looking for - the classifier has to a fair extent to be treated as a black box.
There are ways, however, to get a rough idea of what is going on: @MartinK has posted a C# program here that produces a heat map which can give you a rough idea of what a Polimago classifier is sensitive to. In 
13.02.xxx and up the same functionality is also available inside the TeachBench application and as a tutorial for the CVB.Net wrappers.