Mutual information analysis of neural codes through joint density estimation by the variational Bayes method

摘要

Mutual Information is often used to quantify the response property of a neuron to sensory or neural inputs. To calculate mutual information from experimental data, it is necessary to estimate the joint probability density of the input and the output. A common method is to make a histogram of data samples by discretizing them into appropriate bins. However, the result is highly dependent on the choice of bin size and is subject to approximation error, especially when the number of data is limited. We propose an alternative method in which the input-output joint density is estimated without discretization. Specifically, we use the variational Bayes method for estimating the parameters as well as the complexity of mixture Gaussian models. A better performance compared to conventional methods is verified through a numerical experiment with a simple Poisson neuron model. Its applicability to realistic problems is demonstrated in the experiment with electrically-coupled inferior olive neuron models.