Why is the neural network not trained?

Question

.NET

I want to write my own neural network with 0 that would recognize numbers from 0-9, for this I downloaded the MNIST image databases and started writing code. I implemented the neural network itself quite quickly, but there are problems with training. According to the book "Creating a Neural Network" by Tariq Rashid, I used the formula for getting an error and getting a new weight value.

On a very simple network (1 input 1 hidden layer and 1 output), I applied 1 to the input, and demanded 0 or 1 at the output, and it brought the weights so that the correct value was given. But when I took 786 inputs (pictures with numbers 28x28) 2 hidden layers of 16 neurons and 10 outputs, then for some reason, it does not learn even after 60 thousand. There are almost 1 pictures from the database for any picture everywhere at the output. I gave her these pictures in a random order of 100 pieces.

Here is the source code: GitHub
Why is she not learning and what am I doing wrong?

Answer 1

In this book, in the second part, there will be the creation of the neuron itself in python just for recognizing numbers using MIST. What's the problem with rewriting the author's code?