The paper proposes a new methodology for comparing the similarity structures of colors between humans and Large Language Models (LLMs). The key points are as follows:
-
Beyond traditional correlation analysis, it utilizes the unsupervised alignment technique called Gromov-Wasserstein Optimal Transport (GWOT).
- GWOT - An interesting article on using Gromov-Wasserstein Optimal Transport for determining seating arrangements at a wedding reception.
-
It compares the color similarity structures of color-normal and color-deficient human participants with GPT-3.5, GPT-4, and color space models (RGB and LAB) using a 93-color dataset.
-
Results show that:
- The color similarity structure of GPT-4 is closest to that of color-normal individuals, with a matching rate of 91.4%.
- Although GPT-3.5 also showed significant alignment (11.8%), it was not as pronounced as GPT-4.
- Color space models, despite relatively high correlation coefficients, did not perform well with unsupervised alignment.
-
This methodology revealed detailed structural similarities and differences that cannot be detected by simple correlation analysis.
-
These results provide insights into how accurately LLMs capture human color perception and demonstrate the usefulness of unsupervised alignment methods for comparing human and LLM representations.
-
It’s amusing to think about just giving hex codes and asking for similarity, isn’t it?