Modern (and slightly less modern) Art

The goal of the lesson is to practice distinguishing between the artistic styles of the Renaissance, Impressionism, Surrealism, and Cubism. Students will classify paintings by style using an online data entry form. Once they finish, we will be able to observe their decisions and discuss any mistakes. At the end, we can compare their results to those of a computer that will attempt to classify the same images.

Preparation

Before the lesson, go to https://data.pumice.si/classification and prepare a data entry form: select “Images from the Renaissance, Impressionism, Surrealism, and Cubism”, click the “Prepare activity page” button, and you’ll get a link such as https://data.pumice.si/musician-castle.

Introduction

Discuss the different art movements with the students and show some example paintings of Impressionism, Surrealism, Cubism, and the Renaissance. If desired, you can open the prepared Orange workflow. Double-click the Image Viewer widget and it will display 53 example paintings from the four styles. This is a good moment to talk about the characteristics of each style.

Image Classification

1. Share the link to the data entry page with the students. Each student will first enter their first and last name. Alternatively, you can divide them into pairs or groups, and each group or pair can choose a name.

2. 

The following pages will show images to be classified. Let them do it.

Observing Students’ Results

1. 

   In Orange, double-click the Classification by Students and set the URL: if the students used the link https://data.pumice.si/musician-castle for data entry, then the URL for the results will be https://data.pumice.si/musician-castle/data.

   In the fictional data shown here, two student groups participated — “yellow” and “green”. In class, there will of course be more.

2. 

   Double-click Student Selection and Confusion Matrix to the right of it. In the former, you select a group; in the latter, you see a table where the rows represent the actual artistic styles and the columns the students’ predictions. The numbers in the cells show how many images of a given style a particular group misclassified as another.

   In the screenshot (which shows fictional data—we’ll look at real ones in class), we see that the group Green only misclassified one Renaissance painting, placing it among the Surrealist ones.

   The group Yellow didn’t do as well: they put one Renaissance and one Impressionist painting among the Surrealists, one Surrealist and one Cubist among the Impressionists, and labeled one Impressionist as Renaissance. Most notably, they classified six Surrealist paintings as Cubist. In fact, they only correctly recognized one Surrealist painting!

3. 

   Now open the Image Viewer that is connected to the student confusion matrix, and arrange the windows so both are visible at the same time. If you click a cell in the confusion matrix, the Image Viewer will display the paintings that were misclassified by the students. This lets us see which features they overlooked or misunderstood.

   We can examine the Surrealist paintings that confused the Yellow, but it will be especially interesting to see which Renaissance painting managed to mislead the otherwise infallible Greens.

   At first glance, it really does look more Surrealist than Renaissance. But if we zoom in a bit, we can see that it’s far too… well, Renaissance to be Surrealist.

Could a Computer Do It Too?

In the second part of the lesson, we can show how a computer tackles the same task the students just completed. This is what the lower part of the workflow is for.

The Test and Score widget receives three things:

• The training images, which were also shown to the students. The table includes numbers that describe the images and the label of the artistic style.
• A learning algorithm that will learn to distinguish between the styles.
• The test images — these are the same ones the students classified. Here too, each image has a label, but the computer will not see this label during classification; it’s only used afterward to check which images were classified correctly and which were not.

Just like before, we open the Confusion Matrix — this time the one in the lower part of the workflow — and connect the Image Viewer to it to see the images the computer misclassified.

Maybe we can understand how a Renaissance and a Surrealist painting ended up in the Impressionism category — but The Metamorphosis of Narcissus has absolutely no business being in the Renaissance. What on earth made it put it there?

Wrapping Up With a Bit of Explanation

If the teacher feels up to it, they can end with a short explanation about artificial intelligence.

We’ve gotten used to the idea that AI is all-powerful and simply “knows.” But it is not, and it doesn’t.

When large language models became widely used, someone discovered that even half-century-old chess programs could beat them at chess, which of course sparked media ridicule. But large language models are computer programs that generate text resembling the texts they “read” during training. Nothing more. So naturally, they can’t play chess — and the real joke is on the one who expected they could.

In this lesson, we used an artificial neural network designed for recognizing image content (later fine-tuned slightly for artworks). With this predisposition, it would succeed at the task we gave it only if Renaissance images showed very different motifs than Impressionist ones, and those from Impressionism differed clearly from those in Surrealism and Cubism. That’s what the network is specialized for: it pays attention to what is depicted, not the style. So the fact that it achieved the result it did — with only 4 errors out of 27 images, an accuracy of 89% — is actually quite impressive.

Of course, we could build an artificial neural network specifically trained to distinguish between styles. To do that, we’d show it hundreds or thousands of paintings from each style. The fact that it learned to distinguish styles from a mere 53 training images is, in itself, a remarkable achievement.