Compositional Attention Networks for Machine Reasoning

[howardyclo]

Metadata

• Authors: Drew A. Hudson and Christopher D. Manning
• Organization: Stanford University
• Conference: ICLR 2018
• Paper: https://arxiv.org/pdf/1803.03067.pdf
• Code: https://github.com/stanfordnlp/mac-network
• Video: https://www.youtube.com/watch?v=jpNLp9SnTF8

[howardyclo]

Summary

This paper purpose a novel recurrent, attention and memory based neural architecture "Memory, Attention and Composition (MAC) cell" for VQA, featuring multi-step reasoning.

Three operations in MAC cell:

• Control unit: Updates the control state to attend at each iteration to some aspect of a given question.
• Read unit: Extracts information out of a knowledge base, guided by the control and memory states.
• Write unit: Integrates the retrieved information into the memory state, iteratively computing the answer.