Mechanisms of Information Filtering in Neural Systems

Abstract

Claude Shannon’s information theory has been applied to neural information transmission and information rates of up to hundreds of bits per second have been estimated for single spiking neurons. For Gaussian white-noise signals, one can meaningfully resolve the single number of information rate with respect to frequencies and pose the corresponding question: does a neuron encode information preferentially about slow, intermediate, or fast components of the stimulus? This question is answered in an approximate way by the spectral coherence function, a frequency-resolved measure of information transfer that is related to the lower bound on the mutual information rate. Integrator dynamics in conjunction with uncorrelated intrinsic noise yields an overall low-pass filter. In this paper, the features of the neural dynamics that lead to experimentally observed band-pass or high-pass filter shapes of the coherence are reviewed. The mechanisms at the cellular level such as resonating subthreshold dynamics and slow channel noise are discussed and furthermore the filter effects encountered at the population level such as heterogeneous synaptic short-term plasticity in conjunction with multiple stimuli are explained and a scenario in which the synchronous output of a population is regarded as the signal carrier. Finally, the frequency-resolved mutual information rate, which goes beyond the coherence, is discussed.