Linking Bayesian models of perception and neural responses with spontaneous activity

Abstract:

In order to correctly interpret sensory stimuli and represent them efficiently despite the ambiguities and noise that are pervasive in natural conditions, the brain builds from experience internal models of its environment. A number of recent behavioral studies suggest that the way the brain combines evidence from sensation with the information captured by internal models is well described by Bayesian statistics, implying that neurons must be able to represent and manipulate probability distributions. Uncovering the neural basis of such computations, which remains largely unknown, is crucial to allow a fruitful exchange between computational and experimental studies.

We recently proposed that spontaneous neural activity in the visual cortex reveals an important functional aspect of internal models, namely their prior expectations about the environment. Since in an efficient model of vision the distribution of prior expectations should match the distribution of the features inferred from natural images, we predicted that the distribution of spontaneous neural responses should be identical to the one evoked by natural images, but not by other stimuli. We analyzed the population activity of neurons in primary visual cortex of ferrets over development, from eye-opening to maturity, and found an increasing similarity between the distribution of spontaneous activity and that of activity evoked by natural scenes but not by artificial stimuli, revealing a gradual adaptation of the internal model to the statistics of the visual environment. We confirmed the same hallmarks of internal models in the primary auditory cortex of adult ferrets, suggesting that our findings might uncover a general feature of representation and computation in sensory cortex.

Our hypothesis provides a direct mapping of Bayesian models of perception on neural activity. This link opens the opportunity for a data-driven development and evaluation of models of visual processing, and the development of methods for the decoding of neural responses that would take into account the information contained in ongoing cortical activity.

back