I agree with almost everything you say here, except one thing: “Almost invariably, the nicest problems take one of the following forms: 1) I want to explain theoretically something I observe experimentally 2) I want to improve on an existing method by at least an order of magnitude (in accuracy, simplicity, speed).” To me, the nicest problem is a new task that (almost) nobody has tried before. For example, can I automatically extract keyphrases from a document? Can I automatically determine whether a review (e.g., a movie review, a book review, a car review) is positive or negative? Can I automatically answer multiple-choice SAT analogy questions? This is neither (1) nor (2). In a way, you almost can’t fail, because even the tiniest bit of success on a brand new task is an improvement on the state of the art, because there is no state of the art.
I agree with almost everything you say here, except one thing: “Almost invariably, the nicest problems take one of the following forms: 1) I want to explain theoretically something I observe experimentally 2) I want to improve on an existing method by at least an order of magnitude (in accuracy, simplicity, speed).” To me, the nicest problem is a new task that (almost) nobody has tried before. For example, can I automatically extract keyphrases from a document? Can I automatically determine whether a review (e.g., a movie review, a book review, a car review) is positive or negative? Can I automatically answer multiple-choice SAT analogy questions? This is neither (1) nor (2). In a way, you almost can’t fail, because even the tiniest bit of success on a brand new task is an improvement on the state of the art, because there is no state of the art.
Total agree with you. Almost Same here. 🙂