On the Potential of CLIP for Compositional Logical Reasoning
Voice is AI-generated
Connected to paperThis paper is a preprint and has not been certified by peer review
On the Potential of CLIP for Compositional Logical Reasoning
Justin Brody Franklin and Marshall College
AbstractIn this paper we explore the possibility of using OpenAI's CLIP to perform logically coherent grounded visual reasoning. To that end, we formalize our terms and give a geometric analysis of how embeddings in CLIP's latent space would need to be configured in order for the system to be logically coherent. Our main conclusion is that, as usually configured, CLIP cannot perform such reasoning.