Neuro-Symbolic Visual Reasoning: Disentangling “Visual” from “Reasoning”

  • Saeed Amizadeh ,
  • Hamid Palangi ,
  • Oleksandr Polozov ,
  • Yichen Huang ,
  • Kazuhito Koishida

ICML 2020 |

PDF

Visual reasoning tasks such as visual question answering (VQA) require an interplay of visual perception with reasoning about the question semantics grounded in perception. Various benchmarks for reasoning across language and vision like VQA, VCR and more recently GQA for compositional question answering facilitate scientific progress from perception models to visual reasoning. However, recent advances are still primarily driven by perception improvements (e.g. scene graph generation) rather than reasoning. Neuro-symbolic models such as Neural Module Networks bring the benefits of compositional reasoning to VQA, but they are still entangled with visual representation learning, and thus neural reasoning is hard to improve and assess on its own.

To address this, we propose (1) a framework to isolate and evaluate the reasoning aspect of VQA separately from its perception, and (2) a novel top-down calibration technique that allows the model to answer reasoning questions even with imperfect perception. To this end, we introduce a differentiable first-order logic formalism for VQA that explicitly decouples question answering from visual perception. On the challenging GQA dataset, this framework is used to perform in-depth, disentangled comparisons between well-known VQA models leading to informative insights regarding the participating models as well as the task.