Deep learning recognizes handwritten digits with high recognition rate. We usually use MNIST data sets (60,000 train-images and 10,000 test-images) to demonstrate neural network models.
Original MNIST didn't suitable for this applications and I made my own handwritten data sets.
When you write a number in the blank space, ternary neural networks will recognize the number. You can see the prediction, probability, time cost and number of multiplications.
Ternary Neural Networks have only 9 multiplication patterns.
--------------------------------
0x0
0x1
0x(-1)
1x0
1x1
1x(-1)
(-1)x0
(-1)x1
(-1)x(-1)
--------------------------------
In this applications, details of multiplications in networks are obvious.
If I used Tensor Flow lite, it would be easier to implement neural networks model in Android application. I didn't use libraries to show the number of multiplications. "For loop" recognition isn't efficient but useful to learn details of the networks.