Mixture-of-Experts View of the Double Flash

Read full text
Authors: Aman Chawla
Cited by: 0

Abstract

The sound-induced double-flash illusion demonstrates that perception is dynamically constructed from interactions between multiple sensory modalities. We propose a framework viewing the visual system as a deep mixture-of-experts (MoE) analogue, in which distinct input channels and cortical modules act as specialized experts, and ephaptic coupling and frequency-dependent modulation implement soft, inputdependent weighting. This perspective provides insight into the modularity, inputdependence,hierarchical depth, and probabilistic nature of perception, explaining the emergence of cross-modal illusions and their variability across observers. We formalize the computational mapping between biological mechanisms and MoE architectures and discuss the philosophical implications for perception and cognition. We further extend this framework by introducing Differential Ephaptic Networks (DENs), in which the classical boundary electroneutrality constraint is relaxed, allowing ephaptic coupling matrices to become signed. This yields a biologically groundedanalogue of signed mixture-of-experts architectures, where parallel axonal channels act as proto-experts whose contributions may be selectively amplified or actively suppressed through extracellular field interactions. The resulting DEN–MoE synthesis suggests that competitive expert gating can occur at the tract level prior to synaptic integration, providing a novel substrate for timing variability, multisensory perceptual effects, and illusion-like phenomena. Together, these results motivate a reinterpretation of white-matter pathways as active computational mixtures shaped by boundary physics rather than passive transmission structures. There are also implications for neuromorphic computing - hardware implmentations of DEN–MoE.

Fields of Study & Subjects

Multisensory perception and integration Neural dynamics and brain function Visual perception and processing mechanisms

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
31 January 2026
Document type
Preprint