Reproduce “Persistent Convolution: A Topological Framework for AI Alignment Testing and Semantic Space Characterization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tyler Ashoff, Jordan Rodu
The paper introduces "Persistent Convolution," a topological framework that uses persistence diagrams, landscapes, and silhouettes to characterize the internal semantic structure of AI model embedding spaces. This framework enables statistically rigorous alignment testing against human-curated knowledge and allows for multi-modal, dimensionally agnostic comparisons, even when outcome-based metrics are saturated.
Key insights
- Topological data analysis can reveal differences in internal semantic organization among AI models that achieve identical or near-identical task performance.
- The framework provides a statistically rigorous method to formally test and quantify the alignment of AI model representations with predefined human conceptual structures.
- Topological representations offer a dimensionally and modality-agnostic way to compare and characterize the semantic structure of diverse AI models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.