I was reading an article a while ago about image-sharpening via AI. Some commenters then were saying it's impossible, since you can't create information that wasn't in the source image, therefore there can't possibly be clearer detail in the outcome image. What those commenters miss is that the AI can bring in outside knowledge of how objects work, and thus synthesize the missing details. e.g. if you have a fuzzy face photo, then the algorithm can compute known details of the face, combined with what it knows about how faces work in general, and "sharpen" the image beyond the amount of information that's present in the fuzzy photo.[…]
Top-down processing. With enough qualified expectations, missing/incomplete information can be 'assumed. Like with:
Assuming, of course, that the intended original message was not equally possibly "
TAE CHT", or even "
IAE SA+" but with a
severe lack of information
1. (We've all had edgy CAPTCHAs at least partiallt like that, I'm sure...)
Thus the human-directed training regime has a potential to bias the 'expert' system, in the right or wrong direction. We can't he sure that we could make it
better than a siitably trained/experienced human, just more consistent regardless of caffeine (and/or ethanol) levels, etc...
1True story: Round this way, in the early days of "average speed cameras" (and probable elsewhere, and maybe still if they haven't gotten fed up with it enough to add the necessary filtering) they used
ANPR to record the passage of vehicles at various pointa along a multi-mile stretch of road (as, no doubt, everyone here is familiar with, but allow me to continue to explain just in case anyone does not). Recognition of the same number-plate at two points within a time-difference less than the speed limit (+'allowance'?) would seem to require triggered an 'event', eventually to start an infraction penalty process, but at the time of this anecdote it was just a proof of concept being trialled. It would merely record the events and the associated video feeds for off-line study (for example, did any vehicles pass that did not produce a (consistently/accurately) recognisable number-plate, such that the system might need additional 'training'). But it would also flash up the' offending' vehicle plates to the 'live' controllers in the CCTV centre where the system was being nominally controlled from.
But, for 'privacy' reasons (under the concept of it not yet being a system ceryified capable of determining a sufficient presumption of guilt), it flashed up a version of the plate so detected without the first or last characters. Thus if it saw:
...it would flash up "A51 AB". (Although this experiment pre-dated this actual format of plate.) When they turned it on, though, they got a
lot of anomalies flashing up, all of which caused great amuaement to even the most untrained visitor to the control-room.
...being seen practically simultaneously at multiple points along the busy route, with a disturbingly high calculated mean speed prompting a report, was continuously flashing up
"ONG VEHICL"
"ONG VEHICL"
"ONG VEHICL"
"ONG VEHICL"
"ONG VEHICL"
"ONG VEHICL"
"ONG VEHICL"
...
The moral perhaps being more that human experience was pre-trained better than the computer had been, even when the computer had, but removed for the human's 'benefit', extra information.