I spent way too long staring at figure 1, before realizing that they're not using the word "permutation" in the way that people with exposure to CS or discrete math would use it. They're assigning a "neural clique" (possibly not exactly a clique in the graph theory sense) to each subset in the powerset of inputs, and then ignoring the empty set. _Now_ 2^i - 1 makes sense; ordering does not matter to the authors.