Two labellers, three arms

One fits a model and always answers. The other declines.

keypoint-MoSeq

moseq label 0moseq label 1moseq label 2

MoSeq fits a model that assumes behaviour is a sequence of discrete chunks, and slices every frame into one. It always returns an answer, including when it is given noise.

nonparametric v1

nonparam label 11nonparam label 29nonparam label 28

The nonparametric labeller finds stretches that recur across animals and labels only frames that closely match one, leaving the rest unlabelled. It declines to label a smooth continuum entirely.

nonparametric v2

nonparam2 label 34nonparam2 label 18nonparam2 label 55

The same labeller, with each seed's threshold calibrated against a null instead of set to one arbitrary distance. Five times the coverage, at contamination below MoSeq's.

Four published unsupervised segmentation methods, four different architectures — centroid-based, density-based, an HMM, an AR-HMM. Nothing here adjudicates between them; that is why this page validates a labeller instead of adding a fifth.
Four published unsupervised segmentation methods, four different architectures — centroid-based, density-based, an HMM, an AR-HMM. Nothing here adjudicates between them; that is why this page validates a labeller instead of adding a fifth.
0%25%50%75%100%1%10%MoSeq100.0% / 2.82%nonparam v114.0% / 0.00%nonparam276.0% / 2.16%coveragecontamination — assignment rate on the nulldashed: the α sweep
Coverage against the rate at which a labeller fires on a null that is not behaviour. The dashed line is the per-seed α sweep against the microstate null; the three arms sit on it or beside it. MoSeq has no curve because a labeller with no abstain bin has no threshold to sweep. The horizontal axis is logarithmic, so an exact 0.00% cannot be placed on it — that arm is drawn hollow, on the axis itself.
MoSeqnonparam v1nonparam v2
coverage100% by construction14.0%76.0%
abstentionnone — no abstain bin exists86.01%24.02%
artifact mass2.8%0.00%2.16%
surrogate assignmentnot calibrated against one0.00%α = 1e-4 per seed
grammar1,491provisionalINCONCLUSIVE, 0 of 08 of 8 surviving
noise controlwhite noise survives it toono n-gram to testnot yet run
human validationnonenonenone

The 2.8% figure is the conservative half of two: a further 9 labels (18.9% of frames) are enriched for tracking flags but not in every stratum, which means their failure tracks posture — real behaviour that degrades tracking, not the tracker alone. Counting both as one number would overstate the contamination.

The v1 arm's 0.00% surrogate assignment has a 24× margin and does not show its acceptance regions are tight. Phase, AR and OU surrogates all sit about 24× further from the corpus than the corpus sits from itself, so any threshold calibrated against them lands past full coverage.