This workshop is a deep dive into contemporary model architectures and pipelines following the semiotic question of meaning making linked to text, images as well as the multimodality between them. We will focus on how contemporary transformer-based AI systems based on Large Language, Image, or Multimodal Models are trained on selected data inputs (tokens, images, or both) that then translate to high-dimensional numerical representations. What becomes possible through such a representation in latent geometries (for example in multimodal systems)? And how is the meaning-making and knowledge creation to be thought?

Through several short talks and shared discussions, the workshop will explore the inner workings of AI systems including the question of training, ‘representation’ and ‘prediction’ in contemporary AI systems, as well as questions of translation and exchange. By discussing the meaning making of such architectures and spaces through different aspects from data scraping and cleaning, to architecture adapting, benchmarking, and safety alignment, we will ask: How is it to be thought that these numerical representations within high-dimensional space are ‘representing’? In what ways does meaning and knowledge travel and emerge in these data systems differently?

The workshop is organised by Professor Mercedes Bunz with Dr Daniel Chavez Heras, Department of Digital Humanities. If you are interested in attending, please email mercedes.bunz@kcl.ac.uk.

With short contributions from:

• James E. Dobson (Dartmouth College, US)
• Leonardo Impett (University of Cambridge)
• Tobias Blanke (University of Amsterdam)
• Daniel Chavez Heras (King's College London)
• Mercedes Bunz (King's College London)

Programme

• 14.00-14.20 Welcome, introduction and intro to the topic
• 14.20-14.50 Pattern Matching in Vector Space: The Ontology of Multimodal Models. James E. Dobson  (Dartmouth College)
• 14.50-15.10 Neural Exchange Value: Towards a Materialist Epistemology of AI. Leonardo Impett (University of Cambridge/Hertziana)
• 15.10-30 Discussion
• 16.00-16.20 The Epistemic Politics of Compression in Deep Learning Models: How Latent Spaces Erase Facial Difference. Tobias Blanke (University of Amsterdam)
• 16.20-16.40 Topologies of Time: Representation of Temporal Parts in High-Dimensional Space. Daniel Chávez Heras (King's College London)
• 16.40-17.00 Lost in representation: On Red Teaming language vector spaces between features and concepts. Mercedes Bunz  (King's College London)
• 17.00-17.30 Discussion