How difficult is it to write a neural network?

Question

Neural networks

How difficult is it to write a neural network from scratch, for example, in C++? Not in terms of what to take and a library that implements all the functionality, namely from scratch.
I'm interested in understanding the complexity of this from a position. For example, Tensor Flow is how many person-years?
Can one person do it alone now, or should there be several dozen working on it?
I'm not going to write anything myself, it's just that the scale of complexity is interesting and what libraries are needed for this? To work with multidimensional arrays and that's it?

Answer 1

I got about 400 lines. This included the perceptron itself and its "teacher", who slipped him an image and beat / praised depending on the result. Plus, a C# application with a UI was written separately, which already used the generated weights for identification. There is nothing complicated if you know the materiel.
PS. But right now I'm looking at the code and I don't understand why I wrote some sections this way =)

Answer 2

To write as long as it works, not universally, just for your task - is elementary.
Teaching is difficult.
Well and optimization on high-speed performance is for a long time.
------------
If you need to evaluate the complexity of ready-made solutions - look at the source code of those that are available.

Answer 3

There, the problem is not to write in such a way that it works without errors. And in writing in such a way that
1) It worked quickly
2) It could work on the GPU
3) Or work in a cluster
4) It was quite universal in terms of use
5) It was easy to debug
6) It had a convenient API
...
Here on this is, I suppose, the main time spent developing things like Tensorflow.