Towards Explainable Neural-Symbolic Visual Reasoning

Hybrid AI systems combining neural networks and symbolic reasoning produce better explanations: The paper argues that integrating connectionist (neural) and symbolic (rule-based) approaches is more effective than either alone for creating explanations that non-technical users can understand, addressing a key gap in current high-performance AI models.

There is currently no agreed-upon standard for what makes a good AI explanation: The authors highlight that despite growing interest in explainable AI (XAI), there is "no general consensus on how to produce and judge explanations," meaning organizations and developers may be using inconsistent methods to explain their AI systems to users.

Explanations can be used to identify and correct AI bias: The paper demonstrates that by using their proposed reasoning model to explain a neural network's decisions, it becomes possible to detect and correct bias in how the network makes its choices—suggesting explanations are not just for transparency but also for improving fairness.

Visual reasoning tasks like image captioning benefit from explainability: Using image captioning as a practical example, the authors show how their approach could be applied to real-world AI systems that interpret visual information, making these systems more trustworthy for applications where understanding the reasoning matters.