Generalizing SCAMPI: Robust Detection of Neural Code Juggling

Abstract

Neurons responding to multiple simultaneous stimuli may fluctuate between stimulus-linked activity patterns rather than encode only a pooled response. Detecting such “code juggling” from trial-wise spike counts typically relies on model comparison frameworks that use single-stimulus responses as benchmarks. However, existing approaches commonly assume Poisson benchmark models, an assumption that is often too restrictive for neural data where spike counts exhibit overdispersion and latent heterogeneity.This work develops a mixed-Poisson generalization of the SCAMPI framework for detecting neural code juggling. Single-stimulus responses are modeled using finite Poisson mixtures, allowing benchmark conditions to involve multiple latent firing regimes. Dual-stimulus responses are then compared across fixed, slow-juggling, fast-juggling, and overreaching mechanisms within a unified predictive recursion marginal likelihood framework. To address identification issues and provide uncertainty quantification, we introduce a shrinkage-modified predictive recursion algorithm and incorporate benchmark uncertainty using a Bayesian bootstrap for Poisson mixtures. Simulation studies show that accounting for benchmark uncertainty improves model discrimination and stabilizes inference when benchmark sample sizes are limited. Applying the method to extracellular recordings from primary auditory cortex reveals that while many neurons exhibit benchmark-like responses, a substantial minority show non-fixed structure consistent with multiplexing. Evidence for multiplexing is broadly distributed across neurons and shows different patterns during sustained responses compared with the initial onset window.