Neurons and Cognition (q-bio.NC)

Abstract: Neural population responses in sensory systems are driven by external physical stimuli. This stimulus-response relationship is typically characterized by receptive fields with the assumption of identical and independent Gaussian or Poisson distributions through the loss function. However, responses to repeated presentations of the same stimulus vary, complicating the understanding of neural coding in a stochastic manner. Therefore, to appreciate neural information processing, it is critical to identify stimulus-response function in the presence of trial-to-trial variability. Here, we present a Bayesian system identification approach to predict neural responses to visual stimuli with uncertainties, and explore whether incorporating response fluctuations by using synaptic variability can be beneficial for identifying neural response properties. To this end, we build a neural network model using variational inference to estimate the distribution of each model weight. Tests with different neural datasets demonstrate that this method can achieve higher or comparable performance on neural prediction compared to Monte Carlo dropout methods and traditional models using point estimates of the model parameters. At the same time, our variational method allows to estimate the uncertainty of neural transfer function, which we have found to be negatively correlated with the predictive performance. Finally, our model enables a highly challenging task, i.e., the prediction of noise correlations for unseen stimuli, albeit to a moderate degree. Together, we provide a probabilistic approach as a starting point for simultaneously estimating neuronal receptive fields and analyzing trial-to-trial co-variability for a population of neurons, which may help to uncover the underpinning of stochastic biological computation.