5 ms·

> By the time we launched the app it was trained on over a million images of Sudoku squares. This is super cool, but I can't help but think that something is m

> By the time we launched the app it was trained on over a million images of Sudoku squares. 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.

empthought 9y ago
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.
Cthulhu_ 9y ago
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.
- zimpenfish 9y ago
  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 https://github.com/gunapandianraj/iOS11-VisionFrameWork - this definitely doesn't touch CoreML
  - yeldarb 9y ago
    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 https://imgur.com/a/RSpTG
    - zimpenfish 9y ago
      > Example of a bad read from Vision Rectangle Detection: https://imgur.com/a/RSpTG https://imgur.com/a/RSpTG Well, it's technically correct - it did find a rectangle :)
- dr_zoidberg 9y ago
  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%.
rfish 9y ago
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.
mcintyre1994 9y ago
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.

empthought 9y ago

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.

Cthulhu_ 9y ago

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.

zimpenfish 9y ago
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 https://github.com/gunapandianraj/iOS11-VisionFrameWork - this definitely doesn't touch CoreML
- yeldarb 9y ago
  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 https://imgur.com/a/RSpTG
  - zimpenfish 9y ago
    > Example of a bad read from Vision Rectangle Detection: https://imgur.com/a/RSpTG https://imgur.com/a/RSpTG Well, it's technically correct - it did find a rectangle :)
dr_zoidberg 9y ago
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%.

rfish 9y ago

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.

mcintyre1994 9y ago

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.