CLIP and Vision-Language Models Quiz

Answers and explanations appear as you take the quiz.

1. What is the core idea behind CLIP?
   • Compressing images into smaller files
   • Placing images and text into the same shared embedding space
   • Generating 3D models from photos
   • Translating speech into text
2. What training approach does CLIP use?
   • Reinforcement learning from human feedback
   • Hand-labeling every object in every image
   • Contrastive learning that pulls matching image-text pairs together
   • Copying weights from a speech model
3. What does zero-shot classification mean in the original CLIP transfer setup?
   • Using text-defined candidate categories without new task-specific supervised classifier fitting
   • Using a model that never received any pretraining data
   • Running without any computation or electricity
   • A guarantee that every category is classified correctly
4. How does CLIP measure whether an image and a caption match?
   • By file size comparison
   • By counting pixels of each color
   • By checking the image's timestamp
   • By cosine similarity between their vectors
5. Roughly how much data was CLIP trained on?
   • A few thousand hand-labeled images
   • Only text, no images
   • Around 400 million image-text pairs from the web
   • A single curated museum collection
6. How does the cited unCLIP image-generation paper use CLIP representations?
   • The original standalone CLIP encoder directly outputs every image pixel
   • A prior supplies CLIP image embeddings to a generative decoder
   • It removes all image representations and uses only filenames
   • It uses CLIP only as a network-encryption protocol
7. Which image-encoder families were evaluated in the original CLIP paper?
   • Only recurrent text encoders
   • Only a mandatory linear pixel average
   • Only a raw audio codec
   • Modified ResNets and Vision Transformers
8. A CLIP batch contains 32 paired images and texts. How many entries are in its full cross-modal score matrix?
   • 1024
   • 32
   • 64
   • 992