A Computational Model for Binding by Enhanced Firing Rate: Implementing Smooth Power-law enhancement in Object-Centric Representations

Ishanvir S. Choongh, Manu Madhav

Keywords: binding problem, visual perception, object-centric representation, binding by enhanced firing rate (BBRE), Slot Attention, power-law enhancement, Feature Integration Theory, computational neuroscience, unsupervised learning, ablation study

TL;DR: MVBA model tests if power-law enhancement of firing rates (mimicking neural attention) improves visual object binding, full model achieving 18.4% better reconstruction on multi-object scenes compared to fixed alpha variant..

Abstract: The binding problem remains a core challenge in perception. A proposed neural mechanism is binding by firing-rate enhancements (BBRE), where attention amplifies neuronal activity for same-object features without the need for synchronized oscillations. We interpret BBRE as a contrast mechanism and test its functional utility computationally via the Minimal Viable Binding Architecture (MVBA), a proof-of-concept model extending Slot Attention with a smooth power-law enhancement ($f(x) = \operatorname{sign}(x) \cdot |x|^{\alpha}$) to amplify strong signals and suppress weak ones in isolated spatial (`WHERE`) and feature (`WHAT`) binding modules, inspired by dual-stream theories of the visual system. Seven ablated variants were trained unsupervised on a synthetic dataset using MSE reconstruction loss. Results showed 71.4\% MSE improvement overall $\textit{(baseline: 1.174 }$ to $\textit{full: 0.335)}$, with 18.4% from power-law enhancement alone. Visualizations validated reduced feature bleeding. MVBA empirically supports BBRE as a contrast mechanism, bridging neuroscience, psychology, and machine learning for biologically plausible object-centric learning.

Submission Number: 66