20 years ago, as student, I took Expert Systems (rule-based) and "Bioinformatics" (neural networks, backpropagation stuff, now called "deep learning") courses,

20 years ago, as student, I took Expert Systems (rule-based) and "Bioinformatics" (neural networks, backpropagation stuff, now called "deep learning") courses, and it was like Expert Systems based on rules -with or without fuzzy logic- would be the future. While neural networks were in the same "box" as other techniques like genetic algorithms (evaluation function driven partially random guesses, with greedy/non-greedy "evolution", etc.). Back then it was like imminent being able to put the knowledge of an expert e.g. a medicine doctor so the artificial expert system could make decisions as good as the one made by the human. After years, it seems that human experts in some areas were not so easy to replace, not because not being able to mimic the rules related to medical diagnosis, but because the inability to filter "bullshit-input" (creative hypochondriac people, multiple patterns recognized at once, etc.). The "deep learning" (backpropagation stuff) reminds me the euphoria of the expert systems in the nineties, but with palpable results, i.e. despite requiring crazy amounts of computer power, you can see some difficult problems solved. What scares me is that instead of the "clean and understandable" modeling of the rule-based expert systems, the "deep learning" (neural networks backpropagation stuff) is hardly understandable, even by experts (learn how to train a model is one thing, and knowing why it works, is another -e.g. you can correlate success for N training cases, guess that you're covering the model, and then, discovering for the N+1 that there was a correlation/causality issue -e.g. discovering you trained for learning blue things instead of square things, just because the square things were blue, and when a red square appears, it does not get identified-).