Chapter 11 Explainable machine learning

ML algorithms can be quite complex and therefore not easy to analyze how they come to the result presented to the user, the term black box is used often when referring to this matter. Obviously a ML algorithm is not a black box since the algorithm is nothing but basic math operations, but of those there are a lot so that it is difficult to understand for humans to comprehend how the algorithms works. For example a neural net which is quite popular for natural language processing in 2020 is BERT which has in its base version 110 million parameters, and in its large version an astonishing 340 million parameters.

And then comes along Generative Pre-trained Transformer 3 (GPT-3) creaetd by OpenAi with 175 billion machine learning parameters.

Due to different reasons it might be nice or necessary to understand how the algorithm works. In case of a movie recommendation system it is not as important as for algorithms which work in autonomous driving cars were they are safety relevant.

Also biased algorithms can be problematic as as discussed in 2.4 because they discriminate certain groups of people.

Benefits of explainable ML:Smiley face

  • Increase safety of systems using ML
  • Avoid discrimination
    • human resources recruitment system
  • Model improvement
    • find artifacts
  • Training data verification
    • find unsuited training data

In the following chapters a description of methods and tools are given.

Next chapters:

  • Methods
  • Tools