Scholarlink

In the field of Visual Question Answering (VQA), there's a burgeoning frontier where symbolic reasoning converges with neural networks, exemplified by Neuro-Symbolic Artificial Intelligence (NSAI). This research paper embarks on an investigative journey to uncover the degree to which NSAI, enhanced by the manual integration of logical rules, surpasses traditional neural network models in the realm of VQA tasks. In simple words, this paper aims to discover the potential of NSAI over traditional methodologies. Our study involves a comparative examination of NSAI alongside conventional neural network approaches when applied to DAQUAR (Dataset made for Question Answering on Real-world images) datasets. This paper focuses on the intricacies of manually infusing logical rules into the NSAI framework, thereby creating a hybrid system that harnesses the strengths of both symbolic and neural computing paradigms, which are discussed. To gauge the performance disparities, a diverse set of benchmarks from the DAQUAR dataset, encompassing real-world scenarios, open-ended questions, and domain-specific contexts, was employed. To assess the performance, two evaluation metrics were used: Accuracy and WUPS (weighted average precision score), which measures the semantic similarity between the predicted answer and the ground truth answer based on WordNet.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative Analysis of NSAI and Traditional Neural Networks for Visual Question Answering on DAQUAR Datasets

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Srivaths Gondi,
  • Divyansh Aggarwal,
  • Sulabh Bansal,
  • Gaurav Kumawat

摘要

In the field of Visual Question Answering (VQA), there's a burgeoning frontier where symbolic reasoning converges with neural networks, exemplified by Neuro-Symbolic Artificial Intelligence (NSAI). This research paper embarks on an investigative journey to uncover the degree to which NSAI, enhanced by the manual integration of logical rules, surpasses traditional neural network models in the realm of VQA tasks. In simple words, this paper aims to discover the potential of NSAI over traditional methodologies. Our study involves a comparative examination of NSAI alongside conventional neural network approaches when applied to DAQUAR (Dataset made for Question Answering on Real-world images) datasets. This paper focuses on the intricacies of manually infusing logical rules into the NSAI framework, thereby creating a hybrid system that harnesses the strengths of both symbolic and neural computing paradigms, which are discussed. To gauge the performance disparities, a diverse set of benchmarks from the DAQUAR dataset, encompassing real-world scenarios, open-ended questions, and domain-specific contexts, was employed. To assess the performance, two evaluation metrics were used: Accuracy and WUPS (weighted average precision score), which measures the semantic similarity between the predicted answer and the ground truth answer based on WordNet.