Scholarlink

<p>Diagrammatic reasoning challenges machines to interpret visual patterns, infer abstract relationships, and apply logical rules-tasks at which humans naturally excel. Yet, capturing multi-scale dependencies and generalizing across diverse visual tasks remains difficult for artificial systems. To address these challenges, we present PAtNet, a novel neural architecture that introduces two key methodological innovations: a Path Aggregation Network (PAN) for robust multi-scale feature representation, and a Deep Predictive Reasoning Block (DPRB) that models contextual relationships and predicts missing or occluded feature components. These modules work in concert to enhance both hierarchical feature extraction and relational reasoning. Through comprehensive experiments on the RAVEN, I-RAVEN, and RAVEN-FAIR benchmarks, PAtNet consistently outperforms previous state-of-the-art methods, with accuracy improvements of 1.3%, 0.3 %, and 1.2%, respectively. Ablation studies further demonstrate that the synergy between multi-scale representations from PAN and the contextual inference of DPRB leads to more robust and interpretable diagrammatic reasoning, advancing the field toward deep learning models grounded in cognitive principles.</p>

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

Advancing abstract reasoning for RPMs with a path aggregation network and deep predictive reasoning

  • 期刊论文

  • 发表日期：

    2025年11月3日

查看全文
  • Amresh Kumar Singh,
  • Sandeep Khanna,
  • Chiranjoy Chattopadhyay

摘要

Diagrammatic reasoning challenges machines to interpret visual patterns, infer abstract relationships, and apply logical rules-tasks at which humans naturally excel. Yet, capturing multi-scale dependencies and generalizing across diverse visual tasks remains difficult for artificial systems. To address these challenges, we present PAtNet, a novel neural architecture that introduces two key methodological innovations: a Path Aggregation Network (PAN) for robust multi-scale feature representation, and a Deep Predictive Reasoning Block (DPRB) that models contextual relationships and predicts missing or occluded feature components. These modules work in concert to enhance both hierarchical feature extraction and relational reasoning. Through comprehensive experiments on the RAVEN, I-RAVEN, and RAVEN-FAIR benchmarks, PAtNet consistently outperforms previous state-of-the-art methods, with accuracy improvements of 1.3%, 0.3 %, and 1.2%, respectively. Ablation studies further demonstrate that the synergy between multi-scale representations from PAN and the contextual inference of DPRB leads to more robust and interpretable diagrammatic reasoning, advancing the field toward deep learning models grounded in cognitive principles.