This is really interesting, but it appears to hinge on an unstated (and unjustified) assumption: that scientists learn by back propagation, or something sufficiently similar that back propagation is a reasonable model.

It also:

* Bakes in the assumption that there are no internal mechanisms to be discovered ("Each environment is a mixture of multivariate Gaussian distributions")

* Ignores the possibility that their model of falsification is inadequate (they just test more near points with high error).

* Does a lot of "hopeful naming" which makes the results easy to misinterpret as saying more about like-named things in the real world than it actually does.

The existence of "experiments" to choose from in the first place is already theory-given. As soon as you've formulated a space of such experiments to explore, almost all your theory work is done.

According to Popper, scientists learn by putting out theories and then trying to falsify them through experiments.