I developed a library that do matrix operations (dot product, sum, multiplation, etc).

Then, I write a simple neural network, feedfoward, with only 1 hidden layer.

I learned about the maths beahind backpropagation, what is an activation function, why the matrix bias is added, and so on.

 INPUTS * WEIGHTS + BIAS

With this library, I trained a model that can solve XOR.