Contributed Session III: Identifying Specific Neural Substrates for Bayesian-like Computations in Binocular Vision and Multisensory Processing.

Bayesian thinking is influential in vision but the grounding of Bayesian computation in wetware is poorly understood. Bayesian reliability (inverse variance) weighting of inputs is predicted by Maximum Likelihood Estimation Theory and has some psychophysical support, but evidence for neural reliability weighting is sparse and neural modeling of reliability weighting is tricky. However, reliability averaging is just one possible perceptual weighted average. An alternative - nonlinear magnitude-weighted averaging - was suggested by Schrodinger in 1926 to account for suprathreshold binocular perception and is available to repurpose for other sensory cue combinations. We identified macaque suppressive binocular neurons that implement nonlinear magnitude-weighted averaging and approximate Bayesian averaging, without suffering the computational difficulties that Bayesian averaging implies. We then applied the binocular modeling to suppressive multisensory (visual-tactile, audio-tactile, and audio-visual) neurons. Although magnitude-weighting is a better fit than reliability-weighted averaging for cortical firing rates (in all four cases and in three different species), nonlinear magnitude-weighted averaging is well correlated with reliability averaging. Magnitude-weighted averaging could serve as a surrogate for Bayesian calculations; mildly suppressive binocular and multisensory bimodal neurons could be neural correlates of Bayesian-like computation in the brain.