In many sensory systems, neurons in later stages are less active than those in earlier stages. At each level a higher degree of specificity is necessary to trigger responding. Sensory data may lie along a continuous curved surface of the n-dimensional state space of possible stimulus forms. If true, that would mean that the state space would be overrepresented by individual neurons, which would free up higher levels of cognition to identify additional patterns in the data. However, neurons are metabolically expensive, so this benefit would have to be weighed by natural selection against the cost.
Kurtosis is one way to measure sparseness in terms of deviations from the Guassian distribution, but it has problems sometimes due to overweighing outliers. Empirically, sparse coding has been found in V1 of primates, the auditory and barrel cortices of rats, layer 6 of the motor cortex of rabbits, and the nuclues higher vocal center of zebra finch, activates in response to specific song sequences. Further research could potentially find more, and quantify additional measures of kurtosis.
Olshausen BA, Field DJ. 2004 Spare coding of sensory inputs. Current Opinion in Neurobiology 14:481-487.