Bias in Machine Learning

The term ‘bias’ is used in many ways in Machine Learning, and I think this leads to a lot of people talking past each other.

Examples of how the term is used:

• Bias/variance trade-off

• Bias as in weights and biases

• Algorithmic bias:

  • Bias in the sense of a result that doesn’t accurately reflect the real world (e.g. due to a poorly chosen training data set) possibly related to cognitive biases of programmers

  • Bias in the sense that the result does accurately reflect the real world but the real world is biased/unfair

  • Bias in the sense that the data used to train the model does accurately represent the real world, but the system learns incorrect proxies (e.g. a system that captions a woman at a computer as ‘man sitting at a computer’)

Related:

• Notes on AI Bias — Benedict Evans

• Even if it’s unbiased, it doesn’t mean it works