I introduced neural module network which can be learned end-to-end in Visual Question Answering task.
This architecture, N2NMNs, learned layout policy that can build layout of layers and learned modules included in layout.
1. Learning to Reason: End-to-End Module
Networks for Visual Question Answering
Ronghang Hu, Jacob Andreas, Marcus Rohrbach et al.
ICCV 2017
Presented by Choi Seong Jae
2017. 11. 11
12. Training: Behavioral cloning from expert polices
• Optimizing loss function in Eqn. 4 from scratch is a
challenging reinforcement learning problem
• Optimizing the layout policy
• Optimizing attention weights for each module
• Learning the parameters in the neural modules
13. Training: Behavioral cloning from expert polices
Is there a red shape above a circle?
Leaves
Internal
Root
attend
re-attend or combine
measure and classify
J. Andreas, M. Rohrbach et al. Neural module networks, CVPR 2016