How to recognize digits: Yet another tutorial nobody needs

    I don’t know if I’m any better than the people I’m gonna criticize today. Nevertheless, I can’t believe in the scale with which waste is spreading in the community.

    I think ML engineer is a complex job. Beginnings are always hard, but ML is offering us an impressive collection of incredible tutorials that we definitely saw, and really didn’t need.

Let’s take a look:

  • How to recognize digits in Keras

    A nice toy project. We have a small dataset, not very useful though, and now we know how to write literally 15 lines of code to run the state-of-the-art modern neural network.

  • How to recognize digits in Tensorflow

That’s the fifth blog post I’ve seen about it this month. Wow, again we can load all the processed examples in one line of code!

  • How to recognize digits in PyTorch

Haven’t I seen it already? And what was that backprogetion algorithm?

My day looks different. I hear the question that can not be found on the internet:

  • Are we able to run this project? How much data do we need? How to get the data cheap?
  • How good can be a human using this features? How to interpret the features?
  • How to explain the model’s decisions to the client?
  • What type of classifier should I use? (that’s the tutorials part indeed)
  • Does metric correlate with the objective and the final system quality?
  • How to validate the progress? How to assure previous improvements will be valid in the next version? How to control the version?
  • Why invest in the research?
  • Will prediction be fast enough? How many sessions can we run? What infrastructure will be needed? How to load the model?

    This is Machine Learning I know. These are the questions I can answer because of great colleagues I had. I’m sure this is what we need to have reliable ML systems all over the world. I’m sure many biggest fails in ML wouldn’t happen if we had a professional space for learning how to structure and manage ML projects. Not the people that can push the score higher using new Keras layer, but experienced, well-balanced, conscious engineers.

And that’s what Machine Learning Stories are for.

Leave a Comment

Your email address will not be published. Required fields are marked *