Robert Meyer, GRK 1589 / BCCN Berlin

Correlations and Coding in Visual Cortex

Understanding the neural code, that is deciphering how joint neural responses represent external stimuli, is one of the cardinal problems in Neuroscience. Since neurons are inherently noisy, the neural code is probabilistic. One key question regarding the probability distributions of neural responses is whether neural activity is correlated. Researchers observed so called noise correlations, shared variability among ensembles of neurons for repeated presentation of the same stimulus, in the visual cortex and other sensory areas. To this day the cause of these correlations remains unclear. Many hypotheses have been formulated about the origin of shared variability. In this thalk we investigate a particular hypothesis in depth. We study analytically and numerically the role of recurrent connectivity as a cause of noise correlations. First, we introduce a novel Python library designed to support and manage numerical simulations such as spiking neuron networks. This library, called pypet, facilitates reproducible research by allowing the scientist to disentangle her core simulation from administrative tasks like scheduling or serialization of data. Besides being well tested and documented, the library provides a rich set of features including native multiprocessing and easy parameter exploration. Next, we investigate analytically and numerically the relation between recurrent connectivity and correlations. Using a recent mean-field approach, we show that Mexican hat connectivity with wider inhibitory than excitatory recurrent connection spread can amplify certain spatial frequencies. Moreover, for homogeneous input and Mexican hat connectivity, we observe the emergence of multiple moving bumps which, in turn, yield noise correlations. With an increasing distance between pairs of cells, the noise correlations are modulated sinusoidally and the amplitude decays exponentially. This holds for wide ranges of parameter settings as well as one- and two-dimensional network models. Moreover, noise correlations persist for heterogeneous stimuli, but the spatial modulation changes. Lastly, we test how the shared variability affects stimulus encoding. In general, the measured correlations decrease the stimulus encoding quality in terms of reduced Fisher information. However, if only a subset of neurons is taken into account to decode the stimulus from the spiking responses, Mexican hat networks achieve better performance than other topologies that do not produce noise correlations.

Additional Information

PhD thesis defence in the GRK 1589 "Sensory Computation in Neural Systems"

Organized by

Klaus Obermayer / Robert Martin