This workshop focused on a major issue in the development of ethical and trustworthy artificial intelligence: the insidious presence of biases that often lead to discriminatory decision making.

Nathan Noiry and Yannick Guyonvarch, both post-doctoral students at Telecom Paris, developed the following topics:

• Definition of the different types of biases: cognitive (tendency to detect non-existent correlations, confirmation bias, etc.), algorithmic (poor choice of metrics) and statistical (mismatch between source and target data).
• Relationships between bias and fairness, a recent field of machine learning, which consists in setting up algorithms that are judged to be efficient with respect to certain criteria (e.g. gender bias).
• Methods to detect representativeness biases and to correct them (imputation, correlation, weighting…)

You can consult here their notebook presenting a re-weighting method using synthetic data. During the discussion that followed the presentation, some participants proposed to test these functions on real data and to adapt them as an open-source library.

Laure Leter, résidente datacraft