Visual Question Answering
The task of answering natural-language questions about images - combining vision and language understanding.
Think of it like
Like a student pointing to a diagram in a textbook and asking "what color is this?" The model sees the image and understands the question.
Example
"What color is the car?" while shown a photo. "How many people are in this room?" over a screenshot. "Is the dog sitting?" - yes or no.
How it actually works
VQA requires grounding language in visual content: parsing the question, locating relevant image regions, and fusing visual features with textual semantics. Biases in datasets (e.g., yes/no imbalance) and object co-occurrence (dogs often near grass) trip up models. Modern approaches use vision-language models like CLIP or specialized VQA architectures.
For product teams
VQA powers accessibility features, search-by-query on images, and interactive image understanding apps.
For engineers
Typically a vision encoder + language model bridged by a fusion mechanism; training requires paired image-question-answer data.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome