Abstract

MAC Net is a compositional attention network designed for Visual QuestionAnswering. We propose a modified MAC net architecture for Natural LanguageQuestion Answering. Question Answering typically requires LanguageUnderstanding and multi-step Reasoning. MAC net's unique architecture - theseparation between memory and control, facilitates data-driven iterativereasoning. This makes it an ideal candidate for solving tasks that involvelogical reasoning. Our experiments with 20 bAbI tasks demonstrate the value ofMAC net as a data-efficient and interpretable architecture for Natural LanguageQuestion Answering. The transparent nature of MAC net provides a highlygranular view of the reasoning steps taken by the network in answering a query.

 

Quick Read (beta)