Scholarlink

<p>Recent philosophical work on explainable AI has explored mechanistic methods targeting the causal structure of deep neural networks. However, the ever-growing complexity of deep learning models challenges purely decompositional strategies. This paper advances an alternative perspective based on topological explanations—an approach that both complements and, in some respects, surpasses the mechanistic perspective. By examining how topological properties of deep neural networks relate to model behavior in a non-causal, abstract, and counterfactual manner, I argue that this framework is particularly well suited to addressing why-questions in AI contexts. Drawing on cases of model-driven and data-driven explanations, I show that the topological explanatory strategy offers distinctive epistemic advantages for explainability research.</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

Topological explanations of deep neural networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月10日

查看全文
  • Marcin Rabiza

摘要

Recent philosophical work on explainable AI has explored mechanistic methods targeting the causal structure of deep neural networks. However, the ever-growing complexity of deep learning models challenges purely decompositional strategies. This paper advances an alternative perspective based on topological explanations—an approach that both complements and, in some respects, surpasses the mechanistic perspective. By examining how topological properties of deep neural networks relate to model behavior in a non-causal, abstract, and counterfactual manner, I argue that this framework is particularly well suited to addressing why-questions in AI contexts. Drawing on cases of model-driven and data-driven explanations, I show that the topological explanatory strategy offers distinctive epistemic advantages for explainability research.