Soongsil University AI research team publishes study on Western art analysis in PNAS

A research team from Soongsil University, in collaboration with KAIST and the University of Virginia, has developed an artificial intelligence model capable of analyzing stylistic and thematic changes in Western painting over the last five centuries. The groundbreaking findings were published in the prestigious journal Proceedings of the National Academy of Sciences (PNAS).

Traditional art history research often focused on formal elements like color and composition. However, this approach failed to capture contextual information such as figures, objects, and narratives within the artworks. To address this limitation, the team utilized a dataset of 72,447 Western paintings and employed a pre-trained multimodal AI model called CLIP (Contrastive Language-Image Pre-training).

By comparing CLIP's contextual embeddings with traditional autoencoder-based embeddings that only consider formal elements, the researchers discovered a significant difference. AI models trained solely on formal aspects struggled to accurately predict the creation period of paintings. In contrast, AI models that incorporated contextual information, including figures and themes, demonstrated a much higher accuracy in predicting the artwork's year of creation and artistic style. This confirmed that context plays a more vital role than formal elements in understanding art across different eras and styles.

Furthermore, the AI analysis revealed broader socio-cultural trends in Western art. The study observed a consistent decline in religious subject matter over time and a gradual shift from portrait-centric styles towards more abstract forms. The research team also conducted an experiment where they provided the AI with a limited prompt about elements that might appear in the next century's art and asked it to modify previous centuries' paintings. The results showed the generation of images similar to the actual styles of those periods, validating the analysis's accuracy. This study is presented as a practical example of Explainable AI (XAI), demonstrating how AI interprets human culture and history through art.

“

We have presented a new approach to quantitatively grasp how creativity in art has changed over time through new approaches using multimodal AI and data science. As AI has been rapidly introduced into human knowledge production activities in recent years, it has now shown that AI can be a new perspective for understanding and interpreting human culture and history, beyond simply being a tool to aid creation.