Per @aguinot (#776): UberSeg partially removes the one-sided "pull" of asymmetric weights on the model fit; 4-fold weight symmetrisation removes it completely — the least-biased option. Implemented as an off-by-default option (follow-on PR to #770).
Setup: rider on the uberseg A/B runs — same tile set, third config variant (uberseg + symmetrisation).
Metrics: shape/B-mode difference vs plain uberseg, plus the S/N cost (symmetrisation quadruples the masked area).
Decider: group call on the bias-vs-noise trade, informed by image-sim m/c where available.
— Claude, on behalf of Cail
Per @aguinot (#776): UberSeg partially removes the one-sided "pull" of asymmetric weights on the model fit; 4-fold weight symmetrisation removes it completely — the least-biased option. Implemented as an off-by-default option (follow-on PR to #770).
Setup: rider on the uberseg A/B runs — same tile set, third config variant (
uberseg+ symmetrisation).Metrics: shape/B-mode difference vs plain uberseg, plus the S/N cost (symmetrisation quadruples the masked area).
Decider: group call on the bias-vs-noise trade, informed by image-sim m/c where available.
— Claude, on behalf of Cail