Hello,
Its raining today and while catching up w a backlog of processing and seeing what's top of the forum updates, I noticed another batch of discussions about "why is this not IOTD" or "backyard vs hosting" etc. It is NOT my intention to open anything like that (nothing would be gained tbh from rehashing), but while watching the rain, a thought suddenly struck me.
There is by now a HUGE data set of images that made TPN, IOTD etc and those that did not. So it ought to be possible to train some algorithms which can look at an image and determine what grade it would make!
Some might argue that removes the human element (aka subjectivity). But given that the data-sets the AI would be being trained on actually incorporate that prior human assessment, thus so would the trained model - and indeed perhaps a model would apply its rubrics more systematically.
I have no idea who submitters, reviewers, judges etc are but tbh there seems no need to have those folks involved any more (especially as any human review process is frankly a bit of a black box anyway if there are no feedback loops built in).
In fact, a sophisticated AI could even give feedback on images (either on an image in its own right, or a comparative score of other images of the same target).
For the avoidance of doubt, this not coming from any place of angst from me about the IOTD process (my images and my skies are nowhere near good enough to stand a chance :laughing-1
.
But I do think that there is enough data now available to try this as an approach. If nothing else it would be interesting to compare what an algorithm comes back with, against where the human assessment goes. Indeed, the astrobin datasets are probably large enough to have deep historical cross-validation sets where the grading of such algorithms could be compared to what the human gradings turned out to be for those images.
Are there any AI engineers in the astrobin community who would/could try building something like this?
Cheers
Its raining today and while catching up w a backlog of processing and seeing what's top of the forum updates, I noticed another batch of discussions about "why is this not IOTD" or "backyard vs hosting" etc. It is NOT my intention to open anything like that (nothing would be gained tbh from rehashing), but while watching the rain, a thought suddenly struck me.
There is by now a HUGE data set of images that made TPN, IOTD etc and those that did not. So it ought to be possible to train some algorithms which can look at an image and determine what grade it would make!
Some might argue that removes the human element (aka subjectivity). But given that the data-sets the AI would be being trained on actually incorporate that prior human assessment, thus so would the trained model - and indeed perhaps a model would apply its rubrics more systematically.
I have no idea who submitters, reviewers, judges etc are but tbh there seems no need to have those folks involved any more (especially as any human review process is frankly a bit of a black box anyway if there are no feedback loops built in).
In fact, a sophisticated AI could even give feedback on images (either on an image in its own right, or a comparative score of other images of the same target).
For the avoidance of doubt, this not coming from any place of angst from me about the IOTD process (my images and my skies are nowhere near good enough to stand a chance :laughing-1
.But I do think that there is enough data now available to try this as an approach. If nothing else it would be interesting to compare what an algorithm comes back with, against where the human assessment goes. Indeed, the astrobin datasets are probably large enough to have deep historical cross-validation sets where the grading of such algorithms could be compared to what the human gradings turned out to be for those images.
Are there any AI engineers in the astrobin community who would/could try building something like this?
Cheers
I'm enjoying seeing the answers - and am relieved to see many fellow luddites out there.
