model cards

In the paper titled ‘Model Cards for Model Reporting‘ (Mitchell et al., 2019), the authors propose a procedure for standardizing the documentation of machine learning and AI models called ‘Model Cards.’ Model cards are 1-2 page documents about a machine learning model. Part of the value of model cards is that they may expose system errors BEFORE a model is productionized, not after, which happens in most cases. As a starting point, think about a nutritional label on the side of a box. The label gives you a sense of what the food is made of and how healthy it is. A model card does the same thing but is more comprehensive. Using the same analogy, imagine that same food label told you where the ingredients were sourced from, what the intention was for creating it, and how it may impact people differently. The model card is comprised of nine sections, each of which conveys an important attribute as to the genesis of the model. In some ways, it’s a sort of who-what-where chronology of the model.

Who created the model?
Why was the model created?
Whom was the model created for?
When was the model created?
…and so on

Model cards are effective AI ethics tools that document and publish known issues. The cards can help to increase transparency, improve fairness, and promote explainability. It is important to understand that the model cards are a mathematical-to-written translation of the model’s origin and performance. This underscores the importance of accurately translating models into a written narrative so that the description of the model’s architecture, performance, and potential dangers is explanative and easily understood.

Example model card below taken from the paper.

References:
Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, D., & Gebru, T. (2019). Model Cards for Model Reporting. https://doi.org/10.1145/3287560.3287596