The firing rate of a neuron in the mammalian cortex fluctuates in coordination with the activity of its neighbors. The nature of this relationship varies across behavioral states, and affects the reliability of the neuron's sensory representation. We found that the rich range of statistical structures in multi-neuron recordings could be reproduced by different operating regimes of a single deterministic network model of spiking neurons.
We fit the parameters of the spiking network model to the statistics of spontaneous and driven activity from 46 different electrophysiology datasets of 20-100 neurons recorded in the sensory and motor cortices of rats and gerbils, using novel computational techniques. First, we used graphics processing units to simulate networks of 512 spiking neurons at 10000x real-time speed. Second, we used Bayesian optimization to find parameters which best reproduced a collection of summary statistics for each dataset: autocorrelation function, mean and variance of spike counts, and stimulus response.
The model successfully fit both the diversity of autocorrelation timescales and the magnitude of the correlations present in the neuronal activity. To investigate the consequences for coding, we drove the networks with external inputs. We consistently observed that noise correlations within each network were smaller for stimuli evoking high firing rate responses than for stimuli evoking less firing. This prediction was verified in recordings from both awake and anesthetized auditory cortex.
Further, evoked responses were least correlated in networks with the largest inhibitory-to-excitatory firing rate ratios. The high inhibitory activity abolished population fluctuations and enhanced coding properties. We confirmed this prediction in two ways. First, looking at multiple sound-evoked recordings from auditory cortex, we found that high levels of fast-spiking inhibition did indeed correlate with reduced noise correlation. Next, we continuously activated auditory cortex PV-positive neurons virally transfected with stable step-function opsin (SSFO), and found that stimulus-independent coordinated population variability decreased while signal-driven activity increased.
Our modelling work suggests that networks with a common architecture can generate widely different multi-neuron patterns depending on their precise parameters. These results provide a computational tool for relating the statistical structure of multi-neuron recordings to neural network connectivity and mechanisms.