VQA Visual Question Answering

what is “AI-complete”?

• require multi-modal knowledge, beyond a single domain
• well-defined quantitative evaluation metric

provide a dataset

• images from MSCOCO
• 3 questions per image
  • open end, free form
  • low-level and commonsense level
  • are images necessary / too strong commonsense?
• network
  • image -> hidden layer
  • question -> lstm / bow
  • then concat these layers, fc