Revisiting Explainable AI through Model-Independent Concept Dictionaries
To address these limitations, we propose DictXAI, a method that defines concepts directly in the input domain via a dictionary---a large, potentially overcomplete set of predefined elements, each carrying an interpretable meaning.
ProofPaper ↗
Key points
- Modern applications of AI rely on increasingly complex models.
- Explainable AI (XAI) has emerged as a set of techniques aimed at improving model transparency.
- We demonstrate the actionable nature of DictXAI explanations, showing that they can attribute AI malfunctions (e.g., Clever Hans effects) directly to identifiable artifact patterns in the data, while fostering human-AI alignment on intricate biomedical signals.
- Overall, our results show that DictXAI provides more interpretable, actionable, and architecture-agnostic insights than classical XAI or existing concept-based approaches.
Sources (1)
- [1]Revisiting Explainable AI through Model-Independent Concept DictionariesarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 7, 03:56 PM
To address these limitations, we propose DictXAI, a method that defines concepts directly in the input domain via a dictionary---a large, potentially overcomplete set of predefined elements, each carrying an interpretable meaning.
Modern applications of AI rely on increasingly complex models.
Extractive summary: sentences quoted from the sources.