Online and Hard Negative Mining Quiz

Answers and explanations appear as you take the quiz.

1. Why do most randomly sampled negatives provide little training signal in triplet loss?
   • They are usually mislabeled
   • They are too small to fit in memory
   • They are already far from the anchor, so the loss and gradient are near zero
   • They always equal the positive distance
2. What distinguishes online mining from offline mining?
   • Online mining uses the internet; offline does not
   • Online mining selects hard examples within each mini-batch on the fly
   • Online mining only works on text
   • Online mining requires no forward pass
3. What is a 'semi-hard' negative as defined in FaceNet?
   • A negative closer than the positive to the anchor
   • A negative that equals the anchor exactly
   • Any randomly chosen negative
   • A negative farther than the positive but still within the margin
4. In batch-hard mining, which negative is chosen for each anchor?
   • The nearest negative in the batch
   • The farthest negative in the batch
   • A random negative
   • The negative with the median distance
5. Why does larger batch size tend to improve mined metric-learning results?
   • It reduces the embedding dimension
   • It eliminates the need for a margin
   • It removes the loss function
   • It provides a richer pool of candidate hard negatives per step
6. What is a known risk of selecting the absolute hardest negatives?
   • They are always too easy
   • They make the batch matrix smaller
   • They may be mislabeled or near-duplicate positives that corrupt training
   • They remove the gradient entirely
7. With positive distance 0.4 and margin 0.2, which negative distance is semi-hard under the described rule?
   • 0.5
   • 0.2
   • 0.8
   • 1.1
8. Why does ordinary online mining within one minibatch not necessarily find the hardest negative in the entire dataset?
   • Every batch contains the complete dataset
   • Its available candidate pool is the examples in that batch
   • Distance calculations prove no other examples exist
   • The model discards all negative labels