4 ms·

Deblurring is a classical ill-posed inverse problem where no matter how much math you throw at it (and there's lots of very interesting work in that vein) at th

Deblurring is a classical ill-posed inverse problem where no matter how much math you throw at it (and there's lots of very interesting work in that vein) at the end of the day, data-driven solutions (i.e., ML) will work best for most cases. The only caveat is that ML solutions can be somewhat riskier in that you increase the chances of hallucinating false details, but given the quality of current SOTA it's often an acceptable trade-off (application-dependent, of course). Another caveat: for ML-based solutions, data is key, and a lot of early work on ML-based deblurring suffered from the fact that the "blurry" image data was synthetically created by blurring sharp images. Thankfully, the community has come to realize this and has taken steps to fix it, by collecting "real" blurry datasets and creating more realistic synthetic blurry datasets.

SilverBirch 4y ago
Well this depends on the application right, if I zoom in on an image in twitter and see some detail that wasn't really there... who cares? If I then go on to present that detail in a court of law as evidence of image manipulation, or some characteristic that doesn't exist, or to persuade people the evidence is fallible even though it originally wasn't. Well that's a problem, and it's a difficult problem because it comes down to difficult problems around what you're adveritising you're doing and what people trust you to do. Maybe it doesn't come down to a point of law. Maybe it comes down to trolls on twitter repeating those arguments, but without the burden of proof.
- dimatura 4y ago
  Absolutely! Hence my caveat. Other scenarios where I'd be worried about hallucinations are in algorithms that control self-driving vehicles, or in medical-image analysis. I think accurately quantifying uncertainty for these problems (so that an algorithm, rather than just hallucinate, might say "I don't know") is an important and currently active research topic.

SilverBirch 4y ago

Well this depends on the application right, if I zoom in on an image in twitter and see some detail that wasn't really there... who cares? If I then go on to present that detail in a court of law as evidence of image manipulation, or some characteristic that doesn't exist, or to persuade people the evidence is fallible even though it originally wasn't. Well that's a problem, and it's a difficult problem because it comes down to difficult problems around what you're adveritising you're doing and what people trust you to do. Maybe it doesn't come down to a point of law. Maybe it comes down to trolls on twitter repeating those arguments, but without the burden of proof.

dimatura 4y ago
Absolutely! Hence my caveat. Other scenarios where I'd be worried about hallucinations are in algorithms that control self-driving vehicles, or in medical-image analysis. I think accurately quantifying uncertainty for these problems (so that an algorithm, rather than just hallucinate, might say "I don't know") is an important and currently active research topic.