On the Potential of CLIP for Compositional Logical Reasoning

In 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.