Interpreting CLIP with Sparse Linear Concept Embeddings (SpLiCE)
This addresses the need for transparency in multimodal AI applications by providing a task-agnostic method to interpret and edit CLIP representations, though it is incremental as it builds on existing CLIP technology.
The paper tackles the problem of CLIP embeddings being high-dimensional and not easily interpretable, proposing Sparse Linear Concept Embeddings (SpLiCE) to transform them into sparse linear combinations of human-interpretable concepts, which maintains high downstream performance while improving interpretability.
CLIP embeddings have demonstrated remarkable performance across a wide range of multimodal applications. However, these high-dimensional, dense vector representations are not easily interpretable, limiting our understanding of the rich structure of CLIP and its use in downstream applications that require transparency. In this work, we show that the semantic structure of CLIP's latent space can be leveraged to provide interpretability, allowing for the decomposition of representations into semantic concepts. We formulate this problem as one of sparse recovery and propose a novel method, Sparse Linear Concept Embeddings, for transforming CLIP representations into sparse linear combinations of human-interpretable concepts. Distinct from previous work, SpLiCE is task-agnostic and can be used, without training, to explain and even replace traditional dense CLIP representations, maintaining high downstream performance while significantly improving their interpretability. We also demonstrate significant use cases of SpLiCE representations including detecting spurious correlations and model editing.