Achieving Human Parity on Visual Question Answering

The Visual Question Answering (VQA) task utilizes both visual image and language analysis to answer a textual question with respect an image. It has been popular research topic increasing number of real-world applications in the last decade. This paper introduces novel hierarchical integration vision AliceMind-MMU (ALIbaba’s Collection Encoder-decoders from Machine IntelligeNce lab Damo academy - MultiMedia Understanding), which leads similar or even slightly better results than human being does on VQA. A framework is designed tackle practical problems VQA cascade manner including: (1) diverse semantics learning for comprehensive content understanding; (2) enhanced multi-modal pre-training modality adaptive attention; (3) knowledge-guided model three specialized expert modules complex task. Treating different types questions corresponding expertise needed plays important role boosting performance our architecture up level. An extensive set experiments are conducted demonstrate effectiveness new work.