Random versus maximum entropy models of neural population activity

Area of Science:

• Computational neuroscience
• Statistical mechanics
• Information theory

Background:

• The principle of maximum entropy (MaxEnt) is a powerful tool for inferring statistical models from observed data.
• Its application in correlated systems, particularly in neuroscience, is widespread but its performance on complex empirical data remains understudied.
• Understanding the accuracy of MaxEnt is crucial for interpreting collective neural activity.

Purpose of the Study:

• To evaluate the adequacy of the maximum entropy distribution for modeling collective spiking activity in retinal neurons.
• To compare the performance of MaxEnt against random ensembles of distributions using the same constraints.
• To identify conditions under which MaxEnt excels or falters in describing neural data.

Main Methods:

• Reanalyzing existing data on collective spiking activity of retinal neurons.
• Constraining statistical models using mean firing rates and pairwise correlations.
• Comparing the accuracy of the maximum entropy distribution to a random ensemble of distributions.

Main Results:

• Maximum entropy models generally approximated the true distribution of neural activity better than random models.
• This advantage of maximum entropy increased with the size of the neuronal population.
• Maximum entropy's superiority diminished in networks exhibiting strong correlations.

Conclusions:

• The maximum entropy principle is a robust method for modeling neural population activity, particularly for larger groups.
• The presence of strong correlations can limit the effectiveness of maximum entropy models.
• Further research is needed to refine MaxEnt or explore alternatives for highly correlated neural systems.