“Use me well and keep me clean — I’ll never tell what I have seen” — AI model. Part 2 on the dream of democratising ML never seemed closer. (with python code)
Note: In this link you can find part 1 where I introduced snorkel approach for data labelling.
The maxim within a title of this article is an extremely broad topic — but in short it comes to on how hard it is to understand an AI model. This is so important due to the fact that in many areas the current regulation require AI models to be explainable.
This is not only interesting from the long term perspective, to control ML within our human boundaries (keep in mind matrix trilogy ;)), but also from an academic point of view, to further develop ML for our advantage.
In this article I’m going to introduce a snorkel slicing functions that helps to better monitor AI/ML models.
Note: traditional machine learning systems optimize for overall quality, which may be too coarse-grained. Models that achieve high overall performance might produce unacceptable failure rates on critical slices of the data —[writer: for example] data subsets that might correspond to vulnerable cyclist detection in an autonomous driving task... — source here
Snorkel slicing functions may be easily implemented to monitor, interpret and fine tune critical parts of the model.
