HUMANS DIFFER: SO SHOULD MODELS: Systematic Differences Call for Per-subject Modeling

摘要

While machine learning is most often learning from humans, training data is still considered to originate from a uniform black box. Under this paradigm systematic differences in training provided by multiple subjects are translated into unavoidable modeling error. When trained on a per-subject basis those differences indeed translate to systematic differences in the resulting model structure. We feel that the goal of creating humanlike capabilities or behavior in artificial systems can only be achieved if the diversity of humans is adequately considered.