CMU-CS-10-121
Computer Science Department
School of Computer Science, Carnegie Mellon University



CMU-CS-10-121

Statistical Modeling of Spiking Activity
in Large Scale Neuronal Networks

Ryan Christopher Kelly

May 2010

Ph.D. Thesis

CMU-CS-10-121.pdf


Keywords: Neurons, primary visual cortex, V1, GLM, local field potentials, LFP, spike train, network states

Traditional visual neuroscience research has focused on determining the relationship between the activity of single neurons and the stimuli from the outside world, and more recently the interactions within pairs of neurons. These studies have typically recorded from neurons or pairs of cells in isolation. Recent advances in neural recording devices have made it possible to record simultaneously from hundreds of cells. Such data provide new insights into the interactions among the neurons, the connectivity of neurons in a local network, as well as the neural algorithms of information processing. These methods also present new challenges: the scaling of existing system identification and decoding techniques to address the dramatic increase in dimensionality and computational complexity, and the development of new statistical methods to infer the dynamic interaction and connectivity in neuronal ensembles during information processing.

We recorded neuronal activity from the primate primary visual cortex using 96-channel multi-electrode arrays during the presentation of a variety of visual stimuli. We observed that the large fluctuations in firing rate were shared across many cells in the array, regardless of stimulus. These network state changes are related to many other widely known neural phenomena: large spiking stochasticity, slow timescale correlation between cells, and neural oscillations. We sought to understand the extent to which these fluctuations could be captured with the data available.

A statistical technique, the generalized linear model (GLM), has recently begun to be used to model neural activity, due to both its flexibility and computational tractability. In this context, the models we built had explicit terms for the stimulus effects, coupling effects from other cells recorded simultaneously, and more global network effects. We found that the network terms could indeed explain many of the spikes, indicated that neuronal variability cannot be merely considered to be internal noise, but is widely shared across a population of cells. This approach shows how to incorporate the extra-stimulus data in identifying single cell firing properties, as well as taking a step toward reconciling our understanding of single cells with the computations being performed by the larger network.

146 pages


Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu