Quiz collegato alla guida · Medio Livello

Quiz sulla perdita della cerniera

Calcola la perdita di cerniera dai margini firmati e collega le violazioni dei margini agli obiettivi SVM e alle distinzioni dalla perdita logistica.

Percorsi guida correlatiPerdita della cerniera
Domanda 1 di 8

A correctly labeled example has signed margin 0.4. What is its hinge loss?

Anteprima di ogni domanda8 domande

Le risposte e le spiegazioni vengono visualizzate mentre rispondi al quiz.

  1. A correctly labeled example has signed margin 0.4. What is its hinge loss?
    • 0,6
    • 0
    • 1.4
    • 0.4
  2. A quale margine segnato la perdita della cerniera diventa inizialmente zero?
    • Exactly zero only
    • One or greater
    • Any negative value
    • At minus one or less
  3. Un caso con etichetta negativa riceve il punteggio 0,5. Qual è il suo margine firmato?
    • 1.5
    • 0,5
    • -0.5
    • -1.5
  4. Perché un punto correttamente classificato può avere una perdita di cerniera positiva?
    • The point must be mislabeled
    • Correct classification always means a negative score
    • Hinge loss ignores all margins
    • It can be on the correct side but inside the unit margin
  5. In cosa differisce la perdita logistica per un punto correttamente classificato ben oltre il margine?
    • It is always exactly zero
    • It continues to have a positive loss
    • It becomes a class probability itself
    • It is not defined for positive scores
  6. Quale compromesso oggettivo è comune in una SVM lineare?
    • A clustering compactness term plus a tree-depth penalty
    • A squared-error term on numeric class labels plus a seasonal term
    • A weight-norm penalty balanced against hinge-loss violations
    • A probability calibration penalty balanced against class frequency
  7. Why should a decision score not be reported directly as a calibrated probability?
    • A decision score is already the observed event frequency
    • Any score above zero is automatically a calibrated probability
    • Every SVM estimates a generative distribution over classes
    • Hinge loss does not itself calibrate decision scores as probabilities
  8. What should be checked when comparing two SVM implementations called linear classifiers?
    • Whether the estimators use hinge or squared hinge and compatible regularization settings
    • Whether identical training accuracy implies that their loss definitions are equivalent
    • Whether raw loss values must match even when their loss definitions differ
    • Whether each implementation reaches 100% training accuracy