Woah. In an ML class I took last fall we got to design our own classification challenge problems that the entire class then competed on.

The one our group made used this exact dataset (the museum website)! We called it VANGOGHORNO haha. We basically pulled the images from the collection and mixed them with non-Van Gogh paintings. It was surprisingly a very learnable dataset, the top two teams got around 97% accuracy and many teams got to 90%. We resized the images and included paintings from vaguely similar styles like impressionism so we were surprised people did that well. I think the top team used some kind of transfer learning to identify entities in the paintings and learned from that.