IMP: Benchmarking Image Polysemy in Vision-Language Models
Primary Area: datasets and benchmarks
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Image Polysemy, Vision-Language Representation Learning, Benchmark
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: A benchmark for testing the ability of Vision-Language Models to handle Image Polysemy.
Abstract: Current vision-language models predominantly use contrastive losses to learn from the co-occurrence of image and text. While effective for certain tasks, this approach assumes semantic equivalence between these two modalities. This assumption runs counter to the diverse meanings that a single image can convey, which in turn may compromise visual understanding. To investigate the impact of this assumption, we introduce a novel dataset: $\textbf{IMP}$, designed to challenge and evaluate vision-language models on image polysemy. Our empirical results reveal that current models fall short in recognizing the multiple semantic dimensions of images, underscoring the need for more robust approaches for learning vision-language representations. Code and data will be made available on publication.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1888
Loading