SpLiCE is a method for interpreting CLIP embeddings by decomposing them into sparse, nonnegative linear combinations of human-interpretable concepts. CLIP embeddings, while effective for various vision tasks, are dense and hard to interpret. SpLiCE leverages the structured nature of CLIP's latent space to transform these embeddings into sparse semantic components, enabling better interpretability without requiring concept labels or training. Through extensive experiments on real-world datasets, SpLiCE representations are shown to maintain equivalent performance to traditional CLIP embeddings while being more interpretable. The method is applied to various tasks, including detecting spurious correlations, model editing, and quantifying semantic shifts in datasets. SpLiCE's representations are also used for image tagging and concept-based explanations. The method is post-hoc, requiring no prior concept labels, and is applicable to a wide range of downstream tasks. SpLiCE's approach is based on the assumption that CLIP encoders are linear in concept space and that data is sparse in concept space. The method uses dictionary learning to approximate CLIP representations with sparse concept decompositions. SpLiCE has been validated on multiple datasets and has shown promising results in terms of interpretability and performance. The method is also useful for understanding and summarizing datasets, as it can reveal underlying semantic distributions. SpLiCE's approach provides a way to interpret CLIP embeddings in terms of the semantics of the underlying data they encode, making them more useful for applications requiring transparency. The method is supported by empirical evidence and has been shown to be effective in various tasks, including zero-shot classification, image retrieval, and concept-based explanations. SpLiCE's approach is a significant advancement in the field of interpretable AI, offering a way to make CLIP embeddings more interpretable and useful for downstream applications.SpLiCE is a method for interpreting CLIP embeddings by decomposing them into sparse, nonnegative linear combinations of human-interpretable concepts. CLIP embeddings, while effective for various vision tasks, are dense and hard to interpret. SpLiCE leverages the structured nature of CLIP's latent space to transform these embeddings into sparse semantic components, enabling better interpretability without requiring concept labels or training. Through extensive experiments on real-world datasets, SpLiCE representations are shown to maintain equivalent performance to traditional CLIP embeddings while being more interpretable. The method is applied to various tasks, including detecting spurious correlations, model editing, and quantifying semantic shifts in datasets. SpLiCE's representations are also used for image tagging and concept-based explanations. The method is post-hoc, requiring no prior concept labels, and is applicable to a wide range of downstream tasks. SpLiCE's approach is based on the assumption that CLIP encoders are linear in concept space and that data is sparse in concept space. The method uses dictionary learning to approximate CLIP representations with sparse concept decompositions. SpLiCE has been validated on multiple datasets and has shown promising results in terms of interpretability and performance. The method is also useful for understanding and summarizing datasets, as it can reveal underlying semantic distributions. SpLiCE's approach provides a way to interpret CLIP embeddings in terms of the semantics of the underlying data they encode, making them more useful for applications requiring transparency. The method is supported by empirical evidence and has been shown to be effective in various tasks, including zero-shot classification, image retrieval, and concept-based explanations. SpLiCE's approach is a significant advancement in the field of interpretable AI, offering a way to make CLIP embeddings more interpretable and useful for downstream applications.