For instance, imagine a view from a car onto a dark country road ahead, lined with trees. The road bends so you can't see around the corner, but you see a light getting brighter. Humans would conclude that there must be an oncoming car and they would focus on where exactly the headlights appear relative to the middle of the road.
Nothing else in the picture would matter much. The one thing that does matter cannot be concluded from the picture itself. So humans can compress that image as two points (the headlights) and the a line that represents the middle of the road. Pretty good dimensionality reduction I would say.
Machine learning systems can obviously be trained to do something very similar. But my point is that if you take away the situational context necessary for focus, what's left may not tell us much about how humans process images in most cases.
Also, how many examples of car crashes does a machine learning system need in order to recognise where the headlights should definitely not be? How many head-on car crashes do humans need in order to learn from their mistakes? But that's a wholly different subject.
Please don't take this as a criticism of the paper. I haven't (fully) read it and I know next to nothing about random projection. I'm just generally wondering about how lower level features and high level reasoning are interconnected. It doesn't seem to be a one way street (which doesn't necessarily mean that they are on a collision course though).