> But then neural networks came, and we had a full generic framework that was able to do upscaling really well, and a bunch of other things, and saying that you wanted grant money to code by hand a new upscaling algorithm using some clever wavelet technique you had in mind got you nowhere.
I wonder though, how "correct" that ANN upscaling really is. Does it reimplement clever ideas, that people came up with in a more computationally expensive way, plus some ANN specific extra stuff? Or does it make pixel values up "to make things look good" but actually moves away further from what reality was?
And what do we really learn from an ANN doing this well? Do we get to extract the algorithm back out of it, arriving at actual new knowledge, or will it remain a not understandable black box, with the only knowledge learned being "this kind of ANN works well for that kind of task"?
> We need to have that moment in cell biology, when we stop having entire research groups looking at a single pathway or just a single protein conformation for years, and instead use a generic framework to solve a problem.
I don't think it is that easy. Probably both is needed, work in painstaking detail and usage of generic methods/frameworks. Some ANN might by its nature not even consider a new method. Or its result might be not understandable and therefore risky to apply to humans.
I wonder though, how "correct" that ANN upscaling really is. Does it reimplement clever ideas, that people came up with in a more computationally expensive way, plus some ANN specific extra stuff? Or does it make pixel values up "to make things look good" but actually moves away further from what reality was?
And what do we really learn from an ANN doing this well? Do we get to extract the algorithm back out of it, arriving at actual new knowledge, or will it remain a not understandable black box, with the only knowledge learned being "this kind of ANN works well for that kind of task"?
> We need to have that moment in cell biology, when we stop having entire research groups looking at a single pathway or just a single protein conformation for years, and instead use a generic framework to solve a problem.
I don't think it is that easy. Probably both is needed, work in painstaking detail and usage of generic methods/frameworks. Some ANN might by its nature not even consider a new method. Or its result might be not understandable and therefore risky to apply to humans.