"A lemon" to yellow, "mocha" to brown, etc. Can AI give an abstract name to a color? Experimental result

ByBjorn Hovdal

Abstract names may be given to colors such as "tomato" in red and "lemon · chiffon" in yellow. To this,Make it easy to distinguish intermediate colors that are difficult to distinguishBy purpose and by namePurpose of making it easier for consumers to chooseIs there a possibility that Artificial Intelligence (AI) can give an abstract name to a color like a human? The point of researcher'sJanelle ShaneHe is experimenting.

Letting neural networks be weird • New paint colors invented by neural network
http://lewisandquark.tumblr.com/post/160776374467/new-paint-colors-invented-by-neural-network

Letting neural networks be weird • Paint colors designed by neural network, Part 2
http://lewisandquark.tumblr.com/post/160985569682/paint-colors-designed-by-neural-network-part-2

An AI invented a bunch of new paint colors that are hilariously wrong | Ars Technica
https://arstechnica.com/information-technology/2017/05/an-ai-invented-a-bunch-of-new-paint-colors-that-are-hilariously-wrong/

Reprogramming the AI ​​that wanted to name paint colors and failed miserably | Ars Technica
https://arstechnica.com/information-technology/2017/07/new-experiments-reveal-that-ai-are-still-terrible-at-naming-paint-colors/

First, Mr. Shane learned the list of paint 7700 kinds of paint manufacturer "Sherwin Williams" on the neural network with RGB color code. As an algorithm, "char - rnn" which predicts the next character from one string sequence is adopted. By learning a combination of Shah Win Williams' color list and RGB, the neural network generates two of "color name" and "RGB color code" and can give an attractive name for color I checked it.

As a result of the experiment, the longer the learning time using the dataset was, the better the naming was able to be made, but it was named as "bluish-purple color" "Sane Green" It was a very surreal, Shane said. Although it was possible to recognize the basic red and gray by repeating learning several times, the neural network still got the sky blue "Gray Pubic" (gray pubic bone), the dark green "Stoomy Brown Brown) "and named it.


In a series of results, Ms. Shane says "Neural networks really like" brown "" beige "" gray "" neural networks had the worst idea regarding how to name colors. "

Shane said that "Neural networks really like" brown "," beige "and" gray ", because many colors of beige and gray were created as below.


It seems to be concluded that it is still difficult to make AI name a color ... ... When Mr. Shane publishes a blog article, the AI ​​coders who read the article suggest that "What should I do?" To be sent much. Mr. Shane who received the proposal adjusted the creative level of AI and given some new data sets, another result was born.

First, in the first experiment, we assigned a variable regarding "color temperature" to the neural network, but this variable interferes with the neural network combining and characterizing character by character, and creation of AI He said that he had lowered Tyvis. When eliminating the color temperature variable there, the sense of the name is subtle, but the match rate of the type of color and the name has increased.

The list of colors actually named in the above algorithm is as follows. Certainly, rough recognition such as "white" "gray" "green" "brown" is right.


Also, experts suggested adjustment of the data set. Originally, Shane used RGB color to reproduce a wide range of colors by mixing three primary colors of red (Red), Green (Green), Blue (Blue) in the data set, but AI randomly ran the series It produced the form, giving it a name. However, the color tone expressed by the color code produced at this time was not so clear as to be distinguishable by human eyes, and in addition, it was said that there were many muddy hues close to gray and brown. there,HSVYaLabI also challenge using color schemes such as.

However, even with HSV or Lab, Shane can conclude that it did not have a better naming than RGB, and ultimately again RGB is the best option, Shane concludes.

However, after that, one coder sent me a new data set. This dataset contained the names of the paints of the two companies Behr and Benjamin Moore and a list of original color names. The result of using this dataset is as follows, with the result that the name matches the color image, such as "lemon" in yellowish color, "dry custard" in dark beige.


As for the result of using this data set, Mr. Shane said "It was surprisingly good", and Ars Technica of the news media also expresses "It shows the possibility of machine learning."