Systematic characterization of student–teacher interactions governing RaT bias mitigation

Characterize systematically how the interaction between the student function class and the teacher function class determines the degree to which the residual-as-teacher (RaT) procedure mitigates teacher-induced bias, including conditions under which RaT can fully or partially correct such bias.

Background

The analysis shows that RaT’s performance depends on the interplay between the student and teacher classes; for example, in kernel settings the spectra of the corresponding operators and the teacher’s regularization affect the bias and variance terms and yield a provable separation from soft-matching.

While the paper identifies specific instances where RaT mitigates bias effectively, a general theory that explains and predicts this dependence across broader classes is not provided and is highlighted as an outstanding problem.