Using Multidimensional Vector Embeddings to Study Temporal Dimensions of Historical Newsreel Data

Mila Oiva, Tallinn University (Estonia)

How to study the visual discourses, emerging from large-scale visual heritage data, in a way that allows the findings to rise from the data, instead of assigning the types of visual representations in advance? And how to grasp temporal changes in these visual discourses? This presentation showcases uses of the Collection Space Navigator (CSN) (Ohm et al.) to analyze central shot frames of historical newsreels to reveal continuities and changes in visual discourses of Soviet newsreels 1944-1992. We used a pre-trained ResNet50 model to encode the newsreel frames into 2048-dimensional feature vectors. After that we used the CSN tool to visualize the encoded images with two-dimensional UMAP projection. This approach allows visual exploration of image collections, which in our case study can be used to reveal also temporal visual patterns.

Ohm, Tillmann, Mar Canet Solá, Andres Karjus, and Maximilian Schich. “Collection Space Navigator: Interactive Visualization Interface for Multidimensional Datasets,” 2023. https://collection-space-navigator.github.io/.

Mila Oiva holds a PhD from Cultural History with the specialization on history of knowledge, media history, and computational research methods. Her current research interests focus on circulation of information and formation of knowledge in long temporal continuum. She runs a project dedicated to studying Soviet and Estonian newsreels in 1922-1997 computationally. Earlier she has been studying the 19th century global news flows, circulation of fake historical narratives in the Russian language internet forum discussions, and the Cold War era transnational information circulation.