Multisensory suppressive neurons can implement Bayesian-like nonlinearly weighted sensory combination

We address two related questions in multisensory integration: What are suppressive multisensory and multimodal neurons good for, and how could neurons implement the earliest stages of weighted cue averaging often found in multisensory integration psychophysics? Mildly suppressive multisensory and binocular neurons are ideally positioned to implement weighted averaging. Of the many possible weighted averages, the two most interesting theoretical possibilities are Maximum Likelihood Estimation (MLE) - a Bayesian methodology - and nonlinear magnitude weighting, posited by Erwin Schrödinger for binocular averaging. The MLE approach weights reliability (expressed as inverse relative variance), which is trickier to implement in neurons than nonlinear magnitude weighting.

We tested these two models on three classes of cortical multisensory neurons. We find that suppressive audio-visual, visual-tactile and audio-tactile neurons are well positioned to implement a weighted average of their two inputs, a result consistent with our prior findings in binocular suppressive neurons. The actual neural firing rates more closely resemble Schrödinger's nonlinear weighted average, but the output of this Schrödinger model is well correlated with the output of a rigorously implemented MLE reliability-weighted average.

One possible interpretation is that evolution pushed sensory integration towards a Bayesian-like outcome and settled - at least at an early neural level - for a nonlinear approximation that was easier than inverse-relative-variance-weighting to implement in neural systems.