Search papers, labs, and topics across Lattice.
This paper introduces PromptSpLiCE, a novel method for interpreting the changes in concept profiles of vision-language models like CLIP after prompt learning by expressing class-conditioned text embeddings as sparse combinations of a fixed natural-language dictionary. The study reveals that, on average, only 1.6 of the top-10 initial concepts remain in the top-10 after prompt learning, indicating substantial shifts in concept representation that correlate positively with accuracy improvements across multiple datasets. Additionally, the authors derive a local gradient expression that elucidates the sensitivity of image-aligned concept directions, enhancing our understanding of prompt learning dynamics.
Prompt learning transforms concept profiles so dramatically that only 16% of initial top concepts persist post-optimization, revealing a strong link between these changes and accuracy gains.
Prompt learning adapts vision-language models such as CLIP by optimizing continuous prompt vectors, but the learned prompts are difficult to interpret in natural language. We present PromptSpLiCE, a post-hoc method that expresses each class-conditioned text embedding as a sparse combination of concepts from a fixed natural-language dictionary. Using the same dictionary before and after prompt learning allows us to compare changes in their concept profiles. We evaluate PromptSpLiCE on CoOp, a representative prompt-learning method, across 11 image-classification datasets. The concept profiles change substantially: on average, only 1.6 of the initial top-10 concepts remain in the top 10 after learning. Across datasets, profile change is positively associated with accuracy gain. We also derive a local gradient expression that provides geometric intuition for why image-aligned concept directions distinct from the current prompt can have greater loss sensitivity.
