One of the major sources of complexity in #NWB is actually a pretty interesting replay of blank nodes in #RDF / #SemanticWeb tech.
So you've got a complex type of thing where your thing not only has properties but those properties are themselves other things with their own properties. Concretely, say I've got some electrophysiological recording- that's a timeseries, yes, but it also has metadata like the electrode group that collected it. That electrode group has multiple electrodes, and each has its own properties like impedance, position, etc.
Neuroscientists would probably model this as a bigass nested untyped anonymous blob, one might call this the "cognitive style of #MATLAB structs." so that's where the format seems to have started...