| Many studies have shown that trial-to-trial variability is correlated across neurons. Yet, neuronal responses are often assumed independent---mostly to simplify the decoding process. When correlations are intentionally ignored while decoding, a cost is usually incurred. Many methods have been proposed to quantify the cost (if any) of assuming the neural code independent when in fact it is dependent or correlated. However, when these methods are applied, certain simplifications are often made, such as investigating pairs of neurons only or assuming the correlated response multivariate normal. However, by making these strong and poorly justified assumptions on the nature of correlations, the obtained cost of ignoring correlations would be misrepresented.;Here we evaluate the cost of ignoring correlations using a new method based on information theory and graph theory, in which the probability distributions describing the correlated behaviour of neurons are learnt and modelled using Bayesian networks. That is, instead of assuming a predetermined correlation structure, we determine both the probability distributions and the conditional independence relations from actual data, resulting in a more accurate description of the correlated behaviour of neurons. We find that by modelling correlations as such, the effect of correlations on the information in a population is greater than what was previously found. Our results suggest that trial-to-trial variations in the brain are stimulus dependent and could be a natural part of the encoding process as opposed to being mere noise. |
