Siamese Networks and Triplet Loss Quiz

Answers and explanations appear as you take the quiz.

1. What is the defining architectural feature of a Siamese network?
   • Two or more identical branches that share the same weights
   • A single branch with no hidden layers
   • Separate networks with completely independent weights
   • A recurrent loop over time steps
2. In triplet loss, what are the three inputs?
   • Input, hidden, output
   • Anchor, positive, and negative
   • Train, validation, test
   • Encoder, decoder, classifier
3. What does the triplet loss objective enforce?
   • Anchor and negative should be identical
   • All embeddings should collapse to a single point
   • Anchor-positive distance should be smaller than anchor-negative distance by a margin
   • Positive and negative should have equal distance to the anchor
4. Why is 'hard-negative mining' important for triplet loss training?
   • It removes all positives from training
   • It makes the network shallower
   • It replaces the distance metric with a label
   • Easy negatives already satisfy the margin and produce zero useful gradient
5. What capability makes Siamese networks valuable for 'one-shot learning'?
   • They can compare a new input to a single reference and judge similarity, even for unseen classes
   • They need thousands of labeled examples per class
   • They only classify into fixed, predefined categories
   • They cannot generalize beyond training identities
6. How does a Siamese network typically decide if two inputs match?
   • By summing their raw pixel values
   • By comparing the distance between their embedding vectors against a threshold
   • By counting the number of layers used
   • By assigning each a random class label
7. For triplet loss max(0, d(a,p) − d(a,n) + m), let d(a,p)=0.2, d(a,n)=0.8 and m=0.3. What is the loss?
   • 0.3
   • 0.5
   • 0
   • 1.1
8. What happens when a negative example is already farther from the anchor than the required margin?
   • The anchor and positive embeddings are discarded
   • The loss forces the negative closer to the anchor
   • The network switches to classification mode
   • The triplet loss becomes zero and that example gives no useful gradient