Neuronal noise is often assumed to be a major factor limiting information processing in the brain. However, the degree to which trial-to-trial variability in single neuron activity represents "noise" from the point of view of the organism remains controversial, in part because the partition of variability into "signal" and "noise" is model-dependent. We have therefore compared how two models of cortical encoding partition variability. We find that, whereas in conventional (dense) models of population encoding, trial-to-trial variability in neuronal firing implies that the population response is noisy, in sparse overcomplete models trial-to-trial single neuron variability can, surprisingly, coexist with a nearly perfect---noiseless---population representation.
These two models can be distinguished experimentally by examining the pairwise noise correlations of stimulus-evoked responses. Specifically, (i) noise correlation should be stimulus-dependent under a sparseness prior, while it should not be stimulus-dependent under a dense prior; and (ii) the distribution of the noise correlations should be more kurtotic---less Gaussian---as the neuronal representation becomes sparser. With modern techniques for recording simultaneously from large populations of neurons, such characteristics could be assessed experimentally.
Although our model framework should be viewed as a first approximation of more complex neuronal sensory processing, this study provides a novel perspective on the implications of the neuronal variability in cortical computation.
Poster (345KB, PDF):
Computational and Systems Neuroscience (COSYNE) 2008, Salt Lake City, Utah.