Tom Tetzlaff (Institute of Mathematical Sciences and Technology (IMT), Norwegian University of Life Sciences (UMB), Ås, Norway)

The Bernstein Center Freiburg

Bernstein Seminar

Tom Tetzlaff

Institute of Mathematical Sciences and Technology (IMT)
Norwegian University of Life Sciences (UMB)
Ås, Norway

 

Decorrelation of neural-network activity by inhibitory feedback

Abstract:
Spatial correlations in spike-train ensembles can seriously impair the en- and decoding of information in the population rate or in the fine spatio-temporal structure of these spike trains. Recent theoretical and experimental studies showed that spike correlations in neural networks can be considerably smaller than expected based on the amount of shared presynaptic input in such systems. I will demonstrate by means of a simple linear model and simulations of networks of integrate-and-fire neurons that pairwise correlations and hence population-rate fluctuations in recurrent networks are actively suppressed by inhibitory feedback. To investigate the role of feedback, we calculated the power- and cross-spectra of the network response for the intact recurrent system and for the case where the 2nd-order statistics of the feedback signals is perturbed while the shared-input structure and the 1st-order statistics are preserved. In general, any modification of the feedback statistics causes a shift in the power and coherence of the population response. In particular, the neglect of correlations within the ensemble of feedback channels or between the external stimulus and the feedback can amplify population-rate fluctuations by orders of magnitude. This effect can be observed both in networks with purely inhibitory and in those with mixed excitatory-inhibitory coupling. In purely inhibitory networks, shared-input correlations are canceled by negative correlations between the feedback signals. In excitatory-inhibitory networks, the responses are typically positively correlated. Here, the suppression of input correlations is not a result of the mere existence of correlations between the responses of excitatory (E) and inhibitory (I) neurons, but is instead a consequence of the heterogeneity of response correlations across different types of neuron pairs (EE, EI, II). If correlations between EE, II and EI pairs were identical, input correlations could not fall below the level imposed by the amount of shared input. I will further show that the suppression of correlations and population-rate fluctuations in recurrent networks is in general frequency (time-scale) dependent. The affected frequency range is determined by the population-rate transfer properties. Spike responses of neural populations in both experiment and theory typically exhibit low-pass characteristics. High-frequency fluctuations are therefore not affected by feedback.

Host: Janina Kirsch