3 ms·

Deconvolution. Haven't followed that in awhile, but getting to original function was next to impossible. Recording calibration image and/or recent AI developmen

Deconvolution. Haven't followed that in awhile, but getting to original function was next to impossible. Recording calibration image and/or recent AI developments might've taken that into something awesome. I'll go and catch up on material!

toss1 5y ago
We really need to watch it with trusting AI. While my example here isn't photo recognition, the same principle applies. I recently sat for a deposition where the stenographer used an "AI"transcription system. The result was literally pages of errata (vs the standard errata sheet that has space for about a dozen lines). The consistent error I noticed was that the erroneous words were (probably) the word most expected in that position, and NOT the word that I said. So, at a glance, it seemed like a really good transcription. In fact, many errors were barely noticeable to me and I had to go back to the audio recording to confirm. And these were errors that substantially changed the meaning, or even inverted it. This is not merely information loss — the least surprising/lowest information item was inserted instead of the real item — this is actual information CORRUPTION. I'd fully expect parallel phenomena from image - "AI - filling in the item most expected from the training set, and actively corrupting the data by stripping out the highest-value info bits and replacing them with the most expected. Beware
- Keyframe 5y ago
  Yeah, it's called signal reconstruction for a reason. There are classes of it where it's verifiable and decent enough however.
  - toss1 5y ago
    Yes, I'm sure that with properly constrained and well-tested data spaces, it could produce outstanding and very helpful results. But accurate reconstruction in the wild is just sooo far away. And for good reason - it would need to have insane amounts of experience and exposure to every bit of unusual data that existed in the world to get it right...

toss1 5y ago

We really need to watch it with trusting AI. While my example here isn't photo recognition, the same principle applies. I recently sat for a deposition where the stenographer used an "AI"transcription system. The result was literally pages of errata (vs the standard errata sheet that has space for about a dozen lines). The consistent error I noticed was that the erroneous words were (probably) the word most expected in that position, and NOT the word that I said. So, at a glance, it seemed like a really good transcription. In fact, many errors were barely noticeable to me and I had to go back to the audio recording to confirm. And these were errors that substantially changed the meaning, or even inverted it. This is not merely information loss — the least surprising/lowest information item was inserted instead of the real item — this is actual information CORRUPTION. I'd fully expect parallel phenomena from image - "AI - filling in the item most expected from the training set, and actively corrupting the data by stripping out the highest-value info bits and replacing them with the most expected. Beware

Keyframe 5y ago
Yeah, it's called signal reconstruction for a reason. There are classes of it where it's verifiable and decent enough however.
- toss1 5y ago
  Yes, I'm sure that with properly constrained and well-tested data spaces, it could produce outstanding and very helpful results. But accurate reconstruction in the wild is just sooo far away. And for good reason - it would need to have insane amounts of experience and exposure to every bit of unusual data that existed in the world to get it right...