Information geometric analysis for multi-neuronal spike trains - Estimation of directional interaction and direct/indirect synaptic connections

A problem presented at the US Bio PSW Ohio MBI 2012.

Categories

Presented by:
Dr Masami Tatsuno (Neuroscience, University of Lethbridge)
Participants:
E Cytrynbaum, L Matamba Messi, G Miakonkana, B Solomon, M Tatsuno, D Vats

Problem Description

Understanding how the brain works is one of the most challenging questions of modern science. To study the group dynamics of neurons and their role in cognitive functions, multi-neuronal spike patterns have been analyzed by various statistical methods. However, the estimation of possible changes occurring in the underlying neural networks has remained a difficult problem. Recently, information geometry (IG) has been shown to provide a direct estimation of neural interactions. However, the present IG framework fails to estimate a directional interaction between neurons.

The proposed project aimed at developing a statistical method for estimating directional neural interactions. To this end, we have firstly tried extending the IG approach. By taking into account lagged spiking activities, it was expected that IG is capable of estimating a directional interactions. However, we realized that the problem was not that simple, unfortunately. We then started developing two other statistical approaches; one was estimating a probability of all possible spiking events and the other was using a generalized logistic model (GLM) with an L1 regularization. We have confirmed that both methods worked with simulated spike trains generated by a small neural network (N=2). When the size of the network increases (eg., N=10), we found that the GLM method outperformed the probability counting method. Furthermore, with assumption of sparse connectivity and more observation of time points, we showed that the GLM was able to estimate not only symmetric connections but also asymmetric directional connections.

Inspired by these promising preliminary results, we are planning to develop the method further by implementing model selection algorithms, improving approximation and verifying the performance with more numerical simulations. We would then like to apply the method to multi-neuronal electrophysiological data recorded from freely behaving rodents that form a new memory. Our goal is to elucidate underlying neural mechanisms and computational principles of memory formation and consolidation. The study would significantly promote our understanding of how the brain works.

Download the full problem description