High Accuracy Is Not Enough: Epistemic Bias in Machine Learning Task Formulation

Key Points

• High accuracy in machine learning tasks may conceal underlying epistemic bias, impacting overall effectiveness.
• Evidence shows that even highly accurate algorithms can be misled by incorrect task formulations, affecting their reliability.
• Observational analysis focused on algorithm outputs in various contexts indicates significant variability in performance.
• Awareness of epistemic bias in machine learning could enhance algorithm design and implementation going forward.

Abstract