Adapted GooLeNet for Visual Question Answering

Visual Question Answering (VQA) aims at answering a question about an image. In this work, we introduce an effective architecture --Adapted GooLeNet (AG)-- into a typical VQA method MUTAN instead of LSTM for question features capturing. This improvement can capture more levels of language granularities in parallel, because of the various sizes of filters in AG. The empirical study on the benchmark dataset of VQA demonstrates that capturing sentence features on different levels of granularities benefit sentence modelling by utilizing AG.