Do you like the thick brush strokes and soft color palette of Impressionist paintings like Claude Monet? Or do you like Rosco’s bold colors and abstract shapes? Although individual artistic tastes have certain mysteries, a new California Institute of Technology study shows that a simple computer program can accurately predict which painting a person will like.
New research published in the journal Natural human behaviorLeveraged Amazon’s crowdsourcing platform Mechanical Turk with more than 1,500 volunteers to evaluate paintings in the Impressionist, Cubist, Abstract, and Colorfield genres. Volunteer answers were entered into a computer program, and after this training period, the computer was able to predict volunteer artistic preferences much better than it happened by accident.
“I was surprised at this result because I thought the evaluation of art was personal and subjective,” said Kiyohito Iigaya, a postdoc who works in the laboratory of a professor of psychology at the California Institute of Technology. I will. John Odhaty, Related members Tianqiao and Chrissy Chen Institute for Neuroscience At the California Institute of Technology.
The findings not only showed that computers can make these predictions, but also led to a new understanding of how people judge art.
“The important thing is that we have insight into the mechanisms people use to make aesthetic decisions,” says Odherty. “That is, people seem to combine them using basic image features. This is the first step in understanding how the process works.”
In this study, the team will break down the visual attributes of a painting into low-level features such as contrast, saturation, and hue, and high-level features that require human judgment and include such features. I programmed the computer. As if the picture is dynamic or stationary.
“Computer programs estimate how much a particular feature is considered when deciding how much you like a particular work of art,” explains Iigaya. “When making these decisions, both low-level and high-level features are combined. When the computer estimates it, it successfully predicts a person’s preference for another work of art that has never been seen before. I can do it.”
Researchers have also found that volunteers tend to fall into three general categories. People who like paintings using real objects, such as Impressionist paintings. Someone who likes colorful abstract paintings like Rosco. People who like complex paintings, such as Picasso’s Cubist portraits. The vast majority of people fell into the first “real objects” category. “Many people liked Impressionist paintings,” says Iigaya.
In addition, researchers have found that deep convolutional neural networks (DCNNs) can be trained to learn to predict volunteer artistic preferences with a similar level of accuracy. DCNN is a type of machine learning program that provides a computer with a series of training images to help you learn the classification of objects such as cats and dogs. These neural networks have interconnected units, like neurons in the brain. Networks can “learn” by changing the strength of the connection from one unit to another.
In this case, the deep learning approach did not include the selected low-level or high-level visual features used in the first part of the study, so the computer had to “determine” the features to analyze on its own. had.
“Deep neural network models learn exactly like the real brain, so we can’t really know exactly how the network solves a particular task,” explains Iigaya. To do. “It may be very strange, but when I looked inside the neural network, I found that it was building the same functional category as the one I chose.” These results determine aesthetic preferences. It suggests that the functions used to do this may appear naturally in a brain-like architecture.
“We are now actively investigating whether this is the case by examining people’s brains while people are making the same type of decisions,” says Odherty.
In another part of the study, researchers also showed that their simple computer program, already trained on artistic tastes, could accurately predict which photo volunteers would like. They showed volunteers pictures of scenes such as pools and food, and saw results similar to paintings. In addition, researchers have shown that reversing the order is also effective. After first training volunteers in photography, they were able to use the program to accurately predict the subject’s artistic preferences.
Computer programs have succeeded in predicting volunteer artistic tastes, but researchers say there’s still more to learn about the nuances that go into personal tastes.
“There are aspects of preference that are unique to a particular individual that have not been successfully explained using this method,” says O’Doherty. “This more peculiar element may be related to semantic features, or other individual personal features that may affect the meaning, past experience, and evaluation of the painting. Computer model. It may be possible to identify and learn these characteristics in, but doing so is a personal preference in a way that may not be generalized among individuals, as we have found here. Includes a more detailed study of. “
A study titled “”Aesthetic preferences for art can be predicted from a combination of low-level and high-level visual featuresThe National Institute of Mental Health (through the Center for Social Decision Neurobiology, California Institute of Technology), the National Institute of Substance Abuse, the Japan Society for the Promotion of Science, the Swarts Foundation, the Suntory Foundation, and William H. And Helen Lang Summer Undergraduate Research Fellowship. Other California Institute of Technology authors include Sanghyun Yi, Iman A. Wahle (BS ’20), and Koranis Tanwisuth, now a graduate student at UC Berkeley.