Near-optimal multisensory integration revisited

How do neurons integrate multisensory input optimally? We investigated this question by testing monkeys in a multisensory task in which they integrate visual and vestibular inputs to determine their direction of self-motion. Given the statistics of cortical spike trains in response to the visual stimuli used in these experiments, near optimal multisensory integration requires that the visual and vestibular inputs be dynamically reweighted as a function of the reliability of the visual input. Single cell recordings suggest that this is what MSTd neurons do (Fetsch et al. 2012). This reweighting can be implemented with a divisive normalization (DN) in which inputs are combined with fixed linear weights followed by a squaring nonlinearity and normalization (Ohshiro et al. 2011). This solution can also account for many other aspects of multisensory integration such as the principle of inverse effectiveness.
However, theoretical work indicates that another form of divisive normalization (called sequential divisive normalization or SDN), which essentially skips the linear combination step, is required to achieve optimal cue integration under these conditions (Onken et al. SfN 2012). Unfortunately, our previous analysis of single cell recordings in MSTd cannot distinguish between DN and SDN because, in order to tell the difference, one needs to measure the inputs onto the cell while single cell recordings only measure output spike trains. Even at the behavioral level, these nonlinearities are difficult to distinguish, because both DN and SDN can produce near optimal behavior for the levels of cue reliability that have been used in previous experiments.
Therefore we develop two new tests to explore this issue. We first tested monkeys in a condition in which the visual input is completely uninformative (incoherent motion). We found that monkeys continue to behave near optimally, a result that can only be explained by the SDN model. Next, we used the approach of Churchland et al. (2011) to determine how the inputs onto MSTd cells are combined by measuring the amount of variance in the response of the cells due to the inputs (the so called variance of the conditional expectation). This analysis reveals that the SDN model provides a better fit than the DN model to the variance of the conditional expectation. We conclude that monkeys integrate visual and vestibular inputs near optimally even when the visual input only adds noise and that MSTd neurons might use a close approximation to SDN for combining multisensory inputs.