Contrastive Learning Quiz

Answers and explanations appear as you take the quiz.

1. What is the core objective of contrastive learning?
   • Pull similar (positive) pairs together and push dissimilar (negative) pairs apart in embedding space
   • Predict the exact pixel values of an image
   • Classify every input into one of a fixed set of labels
   • Minimize the number of model parameters
2. In a self-supervised image method like SimCLR, how is a positive pair typically created?
   • Two different images of the same labeled class
   • An image and a randomly chosen image from the dataset
   • Two random augmentations (e.g., crops, color jitter) of the same image
   • An image paired with its segmentation mask
3. Which loss function is most associated with contrastive learning?
   • Mean squared error
   • InfoNCE
   • Cross-entropy over class labels
   • Hinge loss for SVMs
4. Why do methods like MoCo use a memory bank or queue?
   • To store labels for supervised fine-tuning
   • To cache the final predictions for faster inference
   • To compress the model for mobile devices
   • To supply many negative examples without needing an enormous batch size
5. What problem do BYOL and SimSiam specifically guard against without using explicit negatives?
   • Overfitting to the training labels
   • Exploding gradients in the optimizer
   • Representation collapse, where all embeddings become identical
   • Vanishing learning rate
6. How does CLIP extend the contrastive idea?
   • It aligns images with their text captions in a shared embedding space
   • It contrasts two augmentations of the same audio clip
   • It removes the need for any neural network
   • It uses reinforcement learning instead of similarity
7. A contrastive image encoder maps every photo to the same vector. Which failure does this illustrate?
   • A useful separation of image categories
   • An increase in the number of negative samples
   • A perfect nearest-neighbor representation
   • Representation collapse
8. Why is MoCo’s training queue different from caching completed inference responses?
   • It serves final chatbot responses without model execution
   • It supplies encoded comparison examples for a contrastive training objective
   • It replaces every positive image with a label
   • It removes the need to update any encoder