Facial recognition fails on darker faces because of biased training data

Evidence

We can loosely define machine learning as when a computer learns to perform a task by being fed a training set of examples, after which it performs the same task with data it has not been fed previously (Louridas & Ebert, 2016). There are a number of different biases that affect machine learning. First, there might be sampling bias: when the population used to train the model does not accurately represent the population with which the model will interact. This is a major reason for the limited success of facial recognition systems. Facial recognition models perform well at identifying White faces because they are primarily trained on White faces (see Buolamwini & Gebru, 2018). — Joseph Avery, Jim Sidanius, Joel Cooper, Gregory Davis, Amanda N., Bias in the Law, 1152 (#corroborates, via linked claim)

Categorization

• XREF: connects to: Voluntary ‘best practices’ guidelines legitimize unconsented facial recognition by dressing it in toothless language — Shoshana Zuboff, The Age of Surveillance Capitalism, 791; Facial recognition fails because training data overrepresents White faces. — Joseph Avery, Jim Sidanius, Joel Cooper, Gregory Davis, Amanda N., Bias in the Law: A Definitive Look at…, 1152; Machine-learning bias often stems from unrepresentative training data

Broader Connections

• Facial recognition fails because training data overrepresents White faces. #corroborates (0.95) — Joseph Avery, Jim Sidanius, Joel Cooper, Gregory Davis, Amanda N., Bias in the Law, 1152