15.156 14.6.15 Objetivos de entrenamiento
| Objetivo | Uso |
|---|---|
LOSS-ATTR |
masked binary/focal loss for observable attribute heads |
LOSS-HIER |
hierarchy-consistency penalty and parent-aware classification |
LOSS-CANONICAL |
class-balanced multi-label loss for 344 canonical labels |
LOSS-BOUNDARY |
boundary/event alignment loss with tolerance intervals |
LOSS-CONTRAST |
supervised contrastive or metric loss on confusion sets and athlete/route positives |
LOSS-DIFFERENTIAL |
pairwise/set ranking margin for decisive contrasts |
LOSS-UNKNOWN |
open-set/evidential/energy objective using held-out unknown families and outliers |
LOSS-CALIBRATION |
post-hoc or train-time calibration selected on calibration split |
LOSS-TEMPORAL |
smoothness without erasing true boundaries; duration-aware regularization |
LOSS-DEBIAS |
penalty/adversarial test against scene, athlete and camera shortcuts |
LOSS-MISSING |
modality-dropout consistency and explicit missingness handling |
LOSS-SELECTIVE |
coverage-risk or rejection loss for deployable selective prediction |
La pérdida class-balanced y focal son opciones para frecuencia desigual [P46-R17, P46-R18]. Ninguna corrige por sí sola falta de ejemplos, etiquetas débiles, sesgo de dominio o clases semánticamente solapadas.
15.156.1 Supervisión débil
Los timestamps pueden reducir el costo de anotación y aproximar
segmentación frame-wise [P46-R9]. En esta enciclopedia, una etiqueta
derivada por timestamp, regla o modelo conserva
provenance=Weak/PseudoLabel y no se mezcla silenciosamente
con adjudicación manual.