Not mentioned in disadvantages is that you can't really explain "how" a neural network produces a given output from given inputs. Some things can be figured out

Not mentioned in disadvantages is that you can't really explain "how" a neural network produces a given output from given inputs. Some things can be figured out from training data, if it were available, but more often than not it isn't, you just get a pre-trained network. So if were comparing it to software, it's like only shipping binary blob which is basically impossible to debug, and in many cases, throwing away most of the source code (the training data). How many people are actually maintaining the full data sets used to train their networks?