Visual Question Answering Quiz

Answers and explanations appear as you take the quiz.

1. What two inputs does a Visual Question Answering system take?
   • An image and a natural-language question
   • Two images to compare
   • A question and an audio clip
   • Only an image with no text
2. Why was the VQA v2.0 dataset created with 'balanced' answers?
   • To make images higher resolution
   • To reduce language bias so models can't guess from text alone
   • To remove all yes/no questions
   • To shrink the dataset for faster training
3. What is 'language bias' in VQA?
   • Translating questions into the wrong language
   • Using too many languages at once
   • Answering based on question wording while ignoring the image
   • Preferring short images over long ones
4. How do many VQA models combine image and question information?
   • By concatenating raw pixels and letters
   • By generating a new image from the question
   • By ignoring the question entirely
   • By fusing their encoded representations, often with cross-attention
5. Classic VQA systems often produced answers by doing what?
   • Classifying over a fixed vocabulary of common answers
   • Drawing a bounding box only
   • Editing the original image
   • Returning the input question unchanged
6. Which is an example of a more advanced VQA challenge?
   • Resizing an image to a thumbnail
   • Answering questions about text and numbers inside a document image
   • Sorting images alphabetically by filename
   • Compressing video files
7. How does the original VQA benchmark score an answer when people phrase the same answer differently?
   • It requires an exact match to one fixed phrase only
   • It scores image resolution instead of answers
   • It compares the prediction with multiple human answers using a consensus-based score
   • It uses the model’s confidence as the ground truth
8. Why can high VQA accuracy still fail to prove visual grounding?
   • Images cannot be represented by neural networks
   • Questions are always randomly generated
   • The benchmark contains no answer labels
   • A model may exploit dataset or language regularities instead of the image