Characterizing the sparseness of neural codes

Area of Science:

• Computational neuroscience
• Information theory
• Visual processing

Background:

• Neural codes for natural visual information are often considered 'sparse'.
• Sparseness is defined as 'population sparseness' (few active neurons) or 'lifetime sparseness' (high kurtosis).
• These two definitions are related but not identical.

Purpose of the Study:

• To investigate the relationship between population sparseness and lifetime sparseness in neural coding schemes.
• To evaluate the effectiveness of lifetime kurtosis as a measure of population sparseness.
• To identify coding schemes that achieve high population sparseness.

Main Methods:

• Measured population sparseness using three metrics: population kurtosis, Treves-Rolls sparseness, and activity sparseness.
• Measured lifetime kurtosis for several biologically inspired coding schemes.
• Compared the correlation between population sparseness measures and lifetime kurtosis.

Main Results:

• The three measures of population sparseness were in close agreement.
• Lifetime kurtosis was uncorrelated with population sparseness across the tested codes.
• Gabor-like codes, often assumed to be highly population sparse, showed low population sparseness.
• Principal components filters yielded the highest population sparseness.

Conclusions:

• Lifetime kurtosis is not a reliable indicator of population sparseness.
• Population sparseness should be measured directly using metrics like population kurtosis, Treves-Rolls, or activity sparseness.
• Principal components filters represent a promising coding strategy for achieving high population sparseness in visual processing.