Seven Mental · 心智七篇
中 EN
← Knowledge Atlas · Entity

Artur d'Avila Garcez

Artur d'Avila Garcez：City University of London 教授，神经符号 AI 第三波综合作者，Logic Tensor Networks，XAI 忠实性原则批判
实
ENTITY · ARTUR GARCEZ · City, University of London · Neurosymbolic Third Wave

Artur d’Avila Garcez

Professor at City, University of London — core researcher in neurosymbolic computing

Garcez co-authored “Neurosymbolic AI: The 3rd Wave” (arXiv:2012.05876) with Luís C. Lamb, synthesizing 20 years of neurosymbolic computing research, published in Artificial Intelligence Review (Springer Nature, 2023). Core contributions: the systematic argument for propositional fixation, the detailed exposition of the six-class Kautz taxonomy, and the critical reassertion of the XAI fidelity principle.

Core contributions
Propositional fixation — systematic argumentThe theoretical boundary showing neural networks cannot express full propositional logical reasoning — connectionism’s most fundamental expressive limit
Kautz six-class taxonomyNeurosymbolic integration spectrum from loosely coupled (Type 1 Symbolic+Neural) to tightly coupled (Type 6 Neuro[Symbolic])
XAI fidelity critiqueCriticizes methods like LIME for abandoning fidelity: “an explanation that does not accurately describe how an ML system works cannot be called an explanation”
Research themes
Logic Tensor Networks
Logic Tensor Networks as one concrete path for neurosymbolic integration
Non-monotonic reasoning
Combining non-monotonic reasoning with neural networks — handling uncertainty and default reasoning
Knowledge extraction
Extracting interpretable symbolic rules from trained neural networks
→ Neurosymbolic AI · Neurosymbolic AI Taxonomy · Knowledge Extraction FidelityGarcez & Lamb (2023)

Artur d’Avila Garcez

身份： 计算机科学家，City, University of London 教授，神经符号计算领域核心研究者

主要贡献： 20 年神经符号 AI 研究的综合与推进，Logic Tensor Networks（LTN）相关工作，XAI 忠实性原则的系统阐述

核心工作

神经符号 AI 综合（与 Luís C. Lamb 合著）

Garcez 与巴西联邦大学的 Luís C. Lamb 合著了论文”Neurosymbolic AI: The 3rd Wave”（arXiv:2012.05876），发表于 Artificial Intelligence Review（Springer Nature，2023）。

论文综合了 20 年神经符号计算研究，提出了：

  1. 神经网络逻辑表达边界（命题固着）的系统论证
  2. Kautz 六类分类法的详细阐述
  3. 可解释 AI 中忠实性原则的批判性重申
  4. 神经符号循环的实用框架

XAI 忠实性批判

Garcez 明确批评 LIME 等流行 XAI 方法放弃忠实性的做法，认为”一个不能准确描述 ML 系统工作方式的解释”不能被称为真正的解释。见 knowledge-extraction-fidelity。

研究主题

  • 神经符号集成理论基础
  • 逻辑张量网络（Logic Tensor Networks）
  • 神经网络的知识提取
  • 可解释 AI 与 AI 可信度
  • 非单调推理与神经网络

References