arxivst stuff from arxiv that you should probably bookmark

End to End Module Networks

Post · Apr 25, 2017 16:40 ·

n2nmns vqa end-to-end state-of-the-art clevr cs-cv

Visual Q&A models are incredibly useful, but many of them are still fragile. This paper proposes an end-to-end solution that does away with parsers while reducing errors by almost 50%. Holy cow, batman.

Highlights From the Paper

  • Reduces overall error by 48.3% relative over CNN+LSTM+SA.
  • Outperforms previous work by over 30% absolute accuracy on questions requiring attribute comparison of two objects.
  • Dynamically predicts a network structure for each instance, without the aid of external linguistic resources at test time.
  • The model can be first trained with behavior cloning from an expert layout policy, and further optimized end-to-end using reinforcement learning.


Arxiv Abstract

  • Ronghang Hu
  • Jacob Andreas
  • Marcus Rohrbach
  • Trevor Darrell
  • Kate Saenko

Natural language questions are inherently compositional, and many are most easily answered by reasoning about their decomposition into modular sub-problems. For example, to answer "is there an equal number of balls and boxes?" we can look for balls, look for boxes, count them, and compare the results. The recently proposed Neural Module Network (NMN) architecture implements this approach to question answering by parsing questions into linguistic substructures and assembling question-specific deep networks from smaller modules that each solve one subtask. However, existing NMN implementations rely on brittle off-the-shelf parsers, and are restricted to the module configurations proposed by these parsers rather than learning them from data. In this paper, we propose End-to-End Module Networks (N2NMNs), which learn to reason by directly predicting instance-specific network layouts without the aid of a parser. Our model learns to generate network structures (by imitating expert demonstrations) while simultaneously learning network parameters (using the downstream task loss). Experimental results on the new CLEVR dataset targeted at compositional question answering show that N2NMNs achieve an error reduction of nearly 50% relative to state-of-the-art attentional approaches, while discovering interpretable network architectures specialized for each question.

Read the paper (pdf) »