Picasso or Monet? Abstraction or reality? The California Institute of Technology introduced a study that reveals the way we look at art. A simple computer program can predict who will find which painting appealing. 

The output of studies working with 1 500 volunteers was brought by the Nature Human Behavior magazine. The contacted people were rating paintings from several art periods, from impressionism, through cubism to abstraction. Their answers were then transferred to a computer software. Afterwards, the computer was able to predict their preference very well. 

One of the project authors, Kiyohito Iigaya, working in a psychology lab, was astonished by the result: “I thought that rating paintings is a very personal and subjective thing.” According to professor John O´Doherty, it seems that people use elementary factors of paintings and combine them. “That’s the first step to understanding how the process works,” says.

The program first “breaks” the visual attributes of the painting to the level that authors call low level elements, like contrast, satiety and shade – and also high level elements, that require human opinion on whether the painting is dynamic or static. Then, it estimates the degree of the specific element consideration when making a decision about how much the person likes the painting.

Furthermore, scientists found that they could also train a deep convolutional neural network (DCNN) to be able to predict art preferences of volunteers with a similar accuracy level. DCNN is a type of machine learning software in which there is a range of training paintings presented to the computer, so it can learn to classify objects like cats vs dogs. These neural networks have units that are interconnected like brain neurons. By changing the strength of connecting one unit to another, the network is able to “learn”.

“In the case of models with deep neural networks, we don’t know exactly how the networks handle a specific task, because models are learning by themselves, like actual human brains,” explains Iigaya. “It can be very mysterious, but when we took a look at the neural network, we were able to say that it constructs the same function categories that we chose.” These results indicate an opportunity for the elements used for aesthetic preference determination to naturally show in the brain architecture.

In the next part of the study, scientists also demonstrated that their simple computer program, that had been already trained to recognize art preferences, is able to precisely predict the photos the volunteers would like to see. They were presented photos of pools, food and more scenes and saw similar results as those that were related to paintings. It was also shown that reversing the order worked as well: after the first training of photo volunteers, they could, using a program, precisely predict subjects’ art preferences.