If neural networks do compression, then they are implicitly providing structure not present in the data. Given a dataset there are always many different ways to

If neural networks do compression, then they are implicitly providing structure not present in the data. Given a dataset there are always many different ways to compress it, each corresponding to some implicit distribution over unseen datapoints. Most of these distributions don’t correspond to real-world generalizations. Yet neural networks do often—not always—choose models for compression that correspond to real generalizations. This is nontrivial.