Online-Vortrag // Jan von Bonsdorff: Finding high-level image content: Teaching an AI to recognize context, agency, visual narration and metaphors
AI-based search motors should be more knowledgeable about image content. How does one prepare the metadata of an image so that the AI understands, not only what is shown, but what the image is about? What formations in the picture would be optimal for the learning process of the neural networks, like visual narratives, agency, or visual metaphorics? How comprehensive should the contextual information be?
Jan von Bonsdorff is professor of Art history at Uppsala University, Sweden. He has written on art transfer in Medieval Northern Europe, on Scandinavian painting in the 19th century, and has a keen interest in Digital Humanities.
[Caption: Planned work flow of annotating and feeding images and meta-data into neural networks, guiding them to high-level metaphors and context. The image plus the natural-language description is processed through WordNet, BERT and other ontologies and programs suited for visual-linguistic pre-training (Recursive Neural Networks). In a second stage, the intermediate result clusters are sent to a Convolutional Neural Network, teaching it eventually to find similar visual-semantic clusters, i. e. images similar in pictorial conventions and context. Graphic: Jan von Bonsdorff, based on an advertisement from 1956.]
____________
PARTICIPATION: The lecture will take place via Zoom. You can join the Zoom meeting at the following link: https://us02web.zoom.us/j/85659345839?pwd=UmFZYU0xN1NxMGJ1MjlQM054NXgvZz09. Meeting-ID: 856 5934 5839 | Password: 148258.