Machines are getting highbrow. One artificial intelligence has learned how to create new styles of art 鈥� now another is teaching itself art history.
By analysing thousands of paintings produced over hundreds of years, the AI was able to spot connections between generations of painters that matched accepted theories in the art world. It might even teach us something new. 鈥淭he machine could be seeing some complex links that we have no idea about,鈥� says , an art historian at the College of Charleston in South Carolina.
Paintings are often grouped according to their style: think of the Renaissance marvels of Botticelli and Michelangelo or the abstract expressionist masterpieces of Rothko and Pollock. Describing exactly what constitutes a style is tricky, even for art historians. But one way to do it involves something computers excel at: analysing large amounts of data.
Advertisement
and his colleagues at Rutgers University, New Jersey wanted to see if machine learning software would spot the same sort of patterns across art movements that humans do. They fed almost 77,000 paintings from the early Renaissance to pop art into a suite of convolutional neural networks, a form of deep learning that is good at image recognition.
The system was trained to identify 20 different styles and then went on to teach itself finer-grained differences between painters and paintings.
Without knowing anything about the progression of art over the last few centuries, the machine was able to arrange the works chronologically, with one style transitioning into the next. 鈥淭his showed that style changes smoothly, which confirms art historians鈥� hypothesis,鈥� says Elgammal.
The AI found connections between distinctive styles and highlighted historically important artists. For example, it identified the work of Paul C茅zanne as a bridge between post-impressionism and cubism. 鈥淎rt historians had already understood that C茅zanne was important,鈥� says Mazzone, who worked with Elgammal鈥檚 team. 鈥淏ut, wow, how striking that the machine also saw that.鈥�
Hidden links
It also found cycles, linking art from the Renaissance period 鈥� such as some of El Greco鈥檚 paintings, for example 鈥� with certain abstract paintings from the early 20th century. Some modern artists have discussed drawing inspiration from El Greco, says Mazzone. Even so, she was surprised how strongly the machine picked up this connection.
鈥淚t suggests it might be even deeper than art historians really give it credit for,鈥� she says. She is excited about the possibility of finding stylistic links that humans have not picked up on.
鈥淭his paper shows once more that various methods from art history can be reconsidered with machine learning and computer vision,鈥� says digital humanities researcher Peter Bell at the Friedrich-Alexander University Erlangan-Nuremberg in Germany.
There are limitations with the team鈥檚 approach, however. 鈥淭heir understanding of style remains restricted to what can be seen visually,鈥� says Sabine Lang at the University of Heidelberg in Germany. 鈥淐ontext, material and paint application are disregarded.鈥�
Mazzone accepts some may be uncomfortable about using computers to understand art. 鈥淭here鈥檚 a fear this could suggest there鈥檚 only one interpretation that鈥檚 correct,鈥� she says. 鈥淏ut that鈥檚 not what鈥檚 going on here鈥�.
She thinks art historians should embrace these new avenues of analysis. 鈥淐omputer scientists are already working with art,鈥� she says. 鈥淒o we want to be part of the conversation or not?鈥�
