Neural Circuit with Top-Down Inhibitory Feedback Outperforms Optimal Bayesian Integration in Multisensory Integration

Bayesian integration is posited as a fundamental computational mechanism underlying multisensory integration, and feedforward neural networks have been proposed to instantiate optimal Bayesian integration (OBI). However, empirical and theoretical research highlights the prevalence of neural feedback projections, raising questions about how recurrent neural networks might contribute to multisensory OBI. We simulated a two-layer neural circuit computational model with reciprocal projections performing a perceptual discrimination task, in which sensory inputs comprise single or dual modalities. The model with reciprocal projections between sensory and decision-making modules can match, underperform, or outperform OBI, depending on feedforward-feedback interplay. This model performance variability accords with prior experimental data. In addition, our theoretical analysis reveals the importance of non-linear interactions within neuronal assemblies in mediating such multisensory integration behaviors. Our work suggests that sensory modalities can be entangled through top-down feedback, challenging the traditional view of their independence, while explaining deviations from OBI.