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

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Orientation Bias and Abstraction in Working Memory: Evidence from Vision Models and Behaviour

Fabio Bauer¹, Or Yizhar, Bernhard Spitzer²; ¹Max Planck Schools, ²Max-Planck Institute

Presenter: Fabio Bauer

Remembering visual orientations involves systematic biases in human working memory. We tested whether vision models exhibit analogous orientation biases when exposed to rotated real-life objects and if they represent orientations independent of object identity. Using representational similarity analysis (RSA), we compare the representational geometries from eight vision models and human behavioral reports to theoretical patterns of orientation encoding and bias. Our analysis differentiated between two representational domains: the 180° space (2-fold rotationally symmetric), and the 360° space (distinguishing 'up' vs. 'down'). To examine the extent to which orientation representations generalized, we compared artificial neural network (ANN) activations within- and between-objects. We found that vision models showed orientation encoding in 180° space and exhibited a pronounced attraction bias, unlike the characteristic repulsion effects observed in human participants. Further, the vision models display limited 360° orientation encoding, with inadequate cross-object generalization. In contrast, human working memory reports readily reflected orientations in 360° in a generalized fashion. Thus, while contemporary vision models can represent stimulus-specific orientation information, they fail to replicate abstract object-independent orientation encoding and bias that humans effortlessly achieve. Our findings underscore critical limitations of current vision models for studying visual working memory processing.

Topic Area: Object Recognition & Visual Attention

Extended Abstract: Full Text PDF