Page 4 and 9 have the technical definitions.
My off the cuff interpretation is that the rule would only cover convolutional neural nets that are trained to identify and determine the orientation of specific objects in geospatial imagery. If the neural net's input/output aren't wrapped in a GUI it sounds like they still might be OK to export without a license
2.Reduces pixel variation by performing scale, color, and rotational normalization on the positive samples;
3. Trains a Deep Convolutional Neural Network to detect the object of interest from the positive and negative samples; and
4. Identifies objects in geospatial imagery using the trained Deep Convolutional Neural Network by matching the rotational pattern from the positive samples with the rotational pattern of objects in the geospatial imagery.
What counts as "geospatial imagery"? Could this apply to any training UI for self-driving cars, maps, street view, etc.?