Don't just down vote this guys, that doesn't help. No, your naive understanding is not correct. 'Deep learning', and by that I mean a neural network, works at

Don't just down vote this guys, that doesn't help. No, your naive understanding is not correct. 'Deep learning', and by that I mean a neural network, works at a super high level by generalising some input (say X) into some related output (lets say, Y). It's not just random choice; it's like defining a programming function: foo(x) -> y { ... } Where the '...' is implemented by a set of statistical weights and training data and so on. ...but ultimately, you raise a valid point. Training models is incredibly time consuming, and incorporating 'corrections' as new training data is extremely non trivial. It's one of the issues with deep learning. I would even go as far as to say, 'You need to regularly update your training data with new examples and counter examples' as a 'when not to use deep learning'.