I started this thread because I wanted to know if you had other ideas or if my idea can be measured and improved:
STOP AND THINK!
Before you read on, take a moment to think about how a computer might be able to recognize these two digits. What kinds of features might it be able to look at? How might it be able to identify these features?
I came up with two ideas:

An image has 28x28 rows and columns and now we have colorcoded it. If we add colorcodes of first [4:10] rows and cols for 3 then that sum will be quite different from the sum we get for 7. We can take a mean of the sum for all 6000+ images for that range and then compare the same on validating data.

[4:10] is upper left corner of the image. We can take sums of colorcodes from all four corners and then take a mean of that for all the images and use it in the validation step.
It is not as simple as what authors suggested and also it can be inefficient and bit complex compared to what the authors came up with. Any opinions on this? What about your ideas?