What neural networks can (and cannot yet) do?

Deep learning (artificial neural networks) is progressing at a rapid pace. In the last few years image recognition performance went from not useful to on a par with human level. Lately, AlphaGo Zero not only beat human masters, but was able to do so entirely learning by playing with itself. And it keeps going; something that was an original discovery 6 months ago may have become an industry baseline.

Moreover, it is relatively easy to start using deep learning - using Python libraries such as Keras or PyTorch. So - should you use it in your project? Is it going to work at all? And if so - is it worth your effort?

I will talk about which problems are likely to be solved by Deep Learning - the kinds of problems, amount of data needed and expected accuracy.

I will show which kind of problems are as easy as using some existing API or "import some_library", which require some work and experience, and which are still out of reach.

I will focus on computer vision, but add a few examples of natural language processing.