Deep learning

Analysis in layers, each one feeding the next. The first layer finds something crude, the second finds patterns in what the first found, and so on up the stack, until relationships appear that no one could have written a rule for. Nothing in there was told what to look for. The depth is the whole trick, and it is all “deep” means.

What it costs you is the explanation. A decision tree can print the path it took: this value was above that threshold, so it went left. A deep network has no such path. The answer emerges from millions of weights nobody set by hand, and asking why it decided what it decided gets you a number, not a reason. Fine when you are sorting images or forecasting demand. A real problem when a person has been refused credit, flagged by a compliance system or triaged in a clinic, and is owed an account of why. That gap is what the whole field of interpretability is trying to close, and it is not closed yet.