Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster B in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Power-Law Enhancement for Rate-Based Binding in Object-Centric Perception

Ishanvir S. Choongh¹, Manu Madhav¹; ¹University of British Columbia

Presenter: Ishanvir S. Choongh

The binding problem—how distributed sensory features are integrated into coherent object representations—remains a central challenge in perception. Binding by enhanced firing rates (“BBRE“), proposed as an alternative to synchrony-based theories, states that attention amplifies neuronal activity for same-object features, creating perceptual separation without oscillatory coordination. Despite its theoretical appeal, “BBRE“ has not been computationally operationalized or tested. Here we introduce the Minimal Viable Binding Architecture (“MVBA“), a neural network that operationalizes “BBRE“ as a smooth power-law function (f(x) = sign(x) · |x|^α, α∈ [1,3]), applied within separated spatial ("WHERE") and feature ("WHAT") binding streams inspired by dual-stream theories of the visual system. We trained seven ablated model variants unsupervised on synthetic multi-object scenes and evaluated reconstruction quality as a proxy for binding fidelity. The full model achieved a 71.4% MSE improvement over baseline (p < 0.001, d = 2.37), with 18.4% attributable to power-law enhancement alone (d = 0.49). Spatial binding benefited most from enhancement (49% MSE reduction). Critically, the combined improvement (71.4%) was less than the sum of individual stream contributions, consistent with the non-additive integration characteristic of attention in Feature Integration Theory. These results demonstrate that a “BBRE“-motivated gain mechanism improves reconstruction quality in an object-centric model, offering an initial computational framework for testing rate-based binding theories.

Topic Area: Computational Models of Vision & Visual Cortex