Re:How We Built the ARKit Sudoku Solver
by empthought on Tuesday October 10, 2017 @06:53AM
(#15440788)
I'll venture to say that one missing piece is the hundreds of million of years of nervous system evolution to distinguish signal from noise.
Re:How We Built the ARKit Sudoku Solver
by mcintyre1994 on Tuesday October 10, 2017 @10:54AM
(#15442282)
FWIW the MNIST dataset they talked about (handwritten digits) only has 60,000 training samples - and that dataset surely has more variance than printed puzzles from magazines.
I think today's machine learning algorithms mimic more evolution than they do how humans actually learn. I think current techniques in machine learning however can one day teach a machine how to learn by building a fixed neural net that has memory that is the learning algorithm.
Re:How We Built the ARKit Sudoku Solver
by Cthulhu_ on Tuesday October 10, 2017 @08:40AM
(#15441253)
You'd think recognising a square (containing squares and more squares if need be) would be relatively simple and not require advanced machine learning / training. Or that recognising a square doesn't take as much. The demo also indicates the sudoku needs to be fairly accurately scanned, similar to a QR code.
Re:How We Built the ARKit Sudoku Solver
by dr_zoidberg on Tuesday October 10, 2017 @10:25AM
(#15442003)
I think he meant for character recognition -- it's explained in the post that they went for an in-house dataset (from sudoku magazines I understand) instead of MNIST. They ran into some issues, found the way to solve them, and improved their training set. This allowed them to reach 98.6% accuracy, and after a few updates to the app over 99%.
Re:How We Built the ARKit Sudoku Solver
by zimpenfish on Tuesday October 10, 2017 @08:47AM
(#15441294)
I don't think the square recognition used any machine learning.
> We use iOS11’s Vision Library to detect rectangles in the image.
Looking at https://github.com/gunapandianraj/iOS11-VisionFrameWork - this definitely doesn't touch CoreML
It's unclear whether Vision uses machine learning behind the scenes though. It's kind of implied in their docs that it uses CoreML behind the scenes (which makes sense with the other things it does like Face recognition and object tracking).
The nice thing is it detects "projected rectangular regions" so even if the puzzle isn't aligned with the camera it still works.
I do wish I had more control though; it runs into trouble sometimes and there's not much I can do other than apply heuristics afterwards to determine whether I should throw out the sample or continue.
Example of a bad read from Vision Rectangle Detection: https://imgur.com/a/RSpTG
Re:How We Built the ARKit Sudoku Solver
by zimpenfish on Friday October 13, 2017 @10:10AM
(#15465798)
> Example of a bad read from Vision Rectangle Detection: https://imgur.com/a/RSpTG
Well, it's technically correct - it did find a rectangle :)
This is super cool, but I can't help but think that something is missing if it takes hundreds of thousands of examples of digits for a machine learning algorithm to be able to differentiate them. It wouldn't take a human child this many. The available machine learning algos are not using near the amount of information available.