Robust Inference of Neuronal Correlations from Blurred and Noisy Spiking Observations

摘要

Emerging large-scale neuronal recording technolo-gies, such as two-photon calcium imaging, typically provide blurred and noisy surrogates of spiking activity. Extracting the underlying neuronal correlations, which are key to understanding neural function and circuitry, from such data is thus a challenging task. Though deconvolution techniques are often applied to such data to recover spiking activity, they require high temporal resolution and signal-to-noise ratio conditions to be effective. In addition, their solutions are biased towards obtaining accurate first-order statistics (i.e., spike detection) via spatiotemporal priors, which may be detrimental to recovering second-order statistics (i.e., correlations). Existing methods for inferring neuronal correlations from two-photon data thus suffer from significant bias and variability. In this work, we propose an algorithm to directly estimate neuronal correlations from ensemble two-photon imaging data, by integrating techniques from point process modeling and variational Bayesian inference, with no recourse to intermediate spike deconvolution. We demonstrate through simulation studies that the proposed method outperforms existing approaches in accurately capturing the underlying neuronal correlations.