-
0:
-
1:
-
2:
-
3:
-
4:
-
5:
-
6:
-
7:
-
8:
-
9:
What is this?
123 Recognizer is 3 ai's that try to guess what digit you drawn. 1 being the worst model and 3 being the best model. The right side is the ai's guess. When you draw, the progress bars will change to indicate what the ai think you have drawn. One of the progress bars will turn red to indicate which number the ai thinks you have drawn.
To test out the different model click on either 1, 2 or 3 on the
logo at the top of the page. The explenation of each ai is written
down below.
To change the view on small screen press on the human and it will
turn the view to robot view
Model 1
This is the first machine learning-model I made. The accuracy was 98% and I was super proud. When I tried it out on the web it performed poorly. Turns out the mnist dataset is centered and scaled in a certain way.
This model problems are scale and positioning. If you center your drawing and don't draw to big or to small it works, else it is really bad. You have to help it for it to guess right.
Model 2
Data augmentation was the solution. I shifted the numbers randomly and rotated them randomly. I was able to scale them up but scaling down was not possible the way I did it. I thought it would be enough with movment and rotation. Scaling up was just extra. It worked well but I was wrong. Scaling down would help the model improve a lot.
The main problem with this model is small numbers work bad, especially if they are not in the center.
Model 3
This model is the third and final. It addresses all the previous problems. It shifts and rotates randomly. The numbers are also scaled up and down randomly. The last thing I added was some noise that helped the model generalize more.
