The specificity of sensory network organization in living systems, achieved through evolution, development and short-term plasticity, can be viewed as an ultimate form of memory collection resulting from the adaptation to different scales of the environmental statistics. According to the metaphor of a global “fit between mind and world” (James, 1890), our sensory neural systems can be thought of as encoding a highly compressed version of the informational content of our natural environment. Here we show that the reliability of the neural code in adult primary visual cortex (V1) reflects in a mirror way the complexity and natural relevancy of the sensory input statistics.
Using intracellular recordings in the anaesthetized mammal, we find that the activity evoked in V1 during exposure to natural scenes continuously updated by eye-movements, displays highly reproducible dynamical states at the subthreshold membrane potential (Vm) level and a temporal impulsional code at the spiking output level. In contrast, responses to simple artificial stimuli (“optimal” gratings) are highly unreliable, which supports the prevalence of rate coding for low-dimension or unknown stimuli. In natural-like conditions, the contrast between the temporally dense informative synaptic input and the sparse spiking output shows that cortical computation removes input redundancies by detecting transient precisely coactive assemblies.
Introducing a statistical definition of complexity and ordered redundancies, we show that both noise and redundancy reduction observed in natural-like conditions are a direct consequence of the principle of mutual-information maximisation, suggesting a general framework for environmental adaptation. This modulation of the code by the relevancy of both the transient and global inputs statistics, expressed as a balance between externally imposed states and internal ongoing states, may correspond to the well known self-generative property of recurrent networks. From the computational point of view, the irreversible dissipation of the input constraints operated by the cortex is interpreted as the entropic cost to pay for observing and engramming (or forgetting) the information present in the environment.