(also NN wizards please don't pick me apart here; I know that's not how it works exactly, but my point I'm trying to get across is it's not like an LSTM: the network has no feedback to prevent over-reliance on a particular learned concept over time)