Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

Evaluating Human-Machine Representational Alignment in Hyperbolic Space

Otto Béla Márton¹, Christina Sartzetaki¹, Pascal Mettes¹, Iris Groen¹; ¹University of Amsterdam

Presenter: Otto Béla Márton

While humans naturally organize concepts hierarchically, this characteristic remains poorly represented in many deep neural networks (DNNs). This may lower generalisability and could cause DNNs to fail in unexpected ways. Virtually all DNNs make use of Euclidean geometry, but hyperbolic geometry is more naturally suitable for hierarchical structures. Using the THINGS dataset of human similarity judgments of object triplets, we examine the alignment between humans and Euclidean versus hyperbolic models, including both a hyperbolic version of a task-optimized DNN (CLIP) and a hyperbolic adaptation of sparse positive embeddings trained directly on the human behavioural data. Confirming the suitability of hyperbolic geometry, we find that the hyperbolic models predict human behavioural similarity judgments significantly better than their Euclidean counterparts.

Topic Area: Visual Processing & Computational Vision

Extended Abstract: Full Text PDF