Algorithm Design and Implementation Biases

Models are not neutral conduits for their data: architecture, objective, regularization, training schedule and evaluation protocol each decide whose patterns are learned and whose errors are seen, so they must be examined group by group like the data themselves.