Poster Presentation

Poster E in Poster Session E: Thursday, August 6, 10:30 am – 12:15 pm, Kimmel Center, Shorin & Rosenthal Rooms

Geometric Dynamics Across Recurrent Vision Models

Kexin Cindy Luo¹, George A. Alvarez¹, Talia Konkle¹; ¹Harvard University

Presenter: Kexin Cindy Luo

The human visual system contains rich feedforward, feedback, and bypass pathways that shape how object categories are represented over time. Inspired by this neuroanatomy, a growing set of deep vision models incorporates recurrent dynamics, yet how category representations differ across recurrent mechanisms remains unclear. Here we characterize the geometric signatures of recurrence in four distinct model families: two-pass long-range feedback networks (LRM/LRA; Konkle & Alvarez, 2023), two within-layer recurrent models (BL, CORnet-RT; Spoerer et al., 2020, Kubilius et al., 2018), and a recurrent model with skip connections and gating (ConvRNN; Nayebi et al., 2018). We find that these recurrent models are geometrically heterogeneous in two key aspects. First, at inference time, these models show distinct representational shifts across processing steps: LRM/LRA locally compacts category representations toward prototypes while preserving global between-category structure, whereas BL, CORnet-RT, and ConvRNN each exhibit distinct global drifts alongside local changes. Second, training with different recurrent architectures yields distinct learned decision-stage geometry: LRM/LRA develops near-orthogonal category prototypes, a pattern absent in their feedforward counterpart (AlexNet) and other recurrent vision models, but also present in some feedforward architectures (e.g., ResNets). These divergent signatures reveal that "recurrence" is not a single computational strategy in current models: different recurrent architectures exhibit distinct geometric trajectories and decision-stage structures. Such heterogeneity motivates more careful differentiation among recurrent vision models in both computational and neuroscientific contexts.

Topic Area: Computational Models of Vision & Visual Cortex