Typical approaches for studying neural coding focus on information transmission in neural circuits by quantifying how easily a stimulus can be recovered from evoked neural responses. These studies provide statistical descriptions of the mapping between stimuli and evoked responses.

An alternative approach assesses the “statistical optimality of an internal model for probabilistic inference.” As best I can tell this approach considers a sensory system to hold a statistically optimal internal model of the environment if the neuronal activity evoked by a stimulus closely matches spontaneous activity in the system. In a new paper that takes this approach (“Spontaneous Cortical Activity Reveals Hallmarks of an Optimal Internal Model of the Environment” published January 7, 2011 in Science) the authors argue that neural activity is the result of the interaction between an internal model of the environment, embedded in underlying neural circuits, and the sensory input.

Note: The paper under review has an associated 23 page PDF file containing supplementary material. Download the supplementary material from here.

In this paper, they describe experiments in the primary visual cortex that test if evoked neuronal responses to natural visual stimuli closely match spontaneous neuronal responses. The analysis is formally based on the concepts of the likelihood of features and the prior distribution of features. The likelihood of features describes the probability with which any given input image can be expected to arise from a particular combination of features. The prior distribution of features is the probability with which any particular combination of features can be expected to occur.

Spontaneous activity in the sensory cortex reflects the prior distribution of features. In contrast, stimulus evoked activity in the sensory cortex reflects the posterior distribution, which describes the probability that any given combination of features may have given rise to a particular input. The posterior distribution can be computed by Bayes’ rule from the likelihood of features and the prior distribution of features.

The authors carried out the tests to see if evoked neuronal responses to natural visual stimuli closely matched spontaneous neuronal responses in animals at four different ages:

• After the eyes first open.
• After the maturation of orientation tuning and long range horizontal connections in the primary visual cortex.
• In the young adult with a fully matured primary visual cortex.
• In the older adult with a fully matured primary visual cortex.

They found that the similarity between the spontaneous and evoked neuronal activities in the primary visual cortex increased with age and was specific to responses evoked by natural scenes. Their interpretation was that this showed a progressive “adaptation of internal models to the statistics of natural stimuli at the neural level.”