[ExI] Quantitative Qualia and the Science of Redness

[Stuart LaForge]

Here is an interesting study that reports using machine learning 
techniques to analyze color qualia quantitatively. Basically they had 
426 people with typical color vision and 257 individuals who were color 
blind take a computer survey where they judged the similarity of 
color-pairs chosen randomly from a pool of 93 colors using a point 
scale. The results were then used to train an artificial neural network 
(ANN) by unsupervised pairwise alignment of individual's similarity data 
for the color pairs without reference to the color name or label. In 
other word, it was told to look for similarities based on the numerical 
distance reported by each pair of individuals for each color pair, 
without being told what the colors were and then used the alignments to 
form clusters corresponding to a "color map". When the relative 
differences between colors reported by the research subjects were 
clustered without reference to the color, it nonetheless turned out that 
clusters corresponded to the various color labels and the color maps of 
the normally-sighted group were similar to one another. The color maps 
of the color-blind people were, also, similar to one another. However, 
the color maps of the color-sighted people were different from the color 
maps of the color-blind people.

https://www.cell.com/iscience/fulltext/S2589-0042(25)00289-5

https://www.lesswrong.com/posts/LYgJrBf6awsqFRCt3/is-red-for-gpt-4-the-same-as-red-for-you

Using machine learning to analyze qualia like this is fascinating. Just 
like an LLM can learn the contextual meaning of words without being 
explicitly programmed with the definition of the words simply by 
statistically analyzing the average numerical distances between words in 
a corpus of text, this technique should allow AI to recognize and use 
colors without being explicitly programmed with any particular 
definition of say red. This would render the question of whether an AI 
can truly see a color to be equivalent to whether an LLM actually 
understands what it is saying.

Brent, you have have a thing for both color qualia and surveys so this 
paper should be right up your alley.

Stuart LaForge