Classic visual question answering is played by a human who asks the question and an artificial intelligence who answers them. Implementing a full blown VQA Model as neural network is a complex task which goes beyond of a single blog post. What is shown instead is inverse VQA in which the game engine asks the questions and the human has to answer them.
The world in the python prototype consists of 10x10 grid with geometric shapes. If the human answers a question correct he gets a +1 reward. So its a test if the human is able to map the shown situation to the questions.
No comments:
Post a Comment