Triplet Loss and Metric Learning Quiz

Answers and explanations appear as you take the quiz.

1. What three inputs make up a triplet in triplet loss?
   • Anchor, positive, and negative
   • Input, hidden layer, and output
   • Query, key, and value
   • Train, validation, and test samples
2. Why does triplet loss include a margin term?
   • To speed up matrix multiplication
   • To increase the number of classes
   • To prevent the model from collapsing all embeddings to one point
   • To convert distances into probabilities
3. Which famous 2015 system popularized triplet loss for face recognition?
   • AlexNet
   • FaceNet
   • ResNet
   • BERT
4. What does a trained metric-learning model output for each input?
   • A class label only
   • A bounding box
   • A single probability score
   • An embedding vector whose distances reflect similarity
5. Why is metric learning well suited to open-set problems like adding new faces?
   • It retrains the whole network for each new identity
   • It only works when all classes are known in advance
   • Comparison relies on distance, so new identities need no retraining
   • It memorizes a fixed list of outputs
6. Embeddings in metric learning are often L2-normalized. What effect does this have?
   • It places embeddings on a unit hypersphere so distances stay bounded and comparable
   • It removes the need for any loss function
   • It converts the network into a classifier
   • It doubles the embedding dimension
7. Under max(0, d(a,p) minus d(a,n) plus margin), d(a,p)=0.3, d(a,n)=0.8 and margin=0.2 yield what loss?
   • 0.3
   • 0.7
   • 1.3
   • 0
8. A positive margin is 0.2 and every embedding collapses to the same point. What loss does that triplet formula give?
   • 0, because both distances are equal
   • 0.2, so collapse has not satisfied the separation objective
   • 1, regardless of the margin
   • A negative loss of minus 0.2