Bistability, non-ergodicity, and inhibition in pairwise maximum-entropy models

Area of Science:

• Computational Neuroscience
• Statistical Physics
• Machine Learning

Background:

• Pairwise maximum-entropy models are widely used to predict neuronal population activity based on time-averaged correlations.
• These models assume a uniform reference measure, which may not accurately capture complex neuronal interactions like inhibition.

Purpose of the Study:

• To investigate the limitations of standard pairwise maximum-entropy models in predicting neuronal population activity.
• To identify the cause of unrealistic bimodal distributions observed in model predictions.
• To develop a modified maximum-entropy model that accurately incorporates neuronal inhibition.

Main Methods:

• Analysis of experimental neuronal recordings from macaque monkey motor cortex.
• Application of standard pairwise maximum-entropy models and evaluation of their output distributions.
• Development and testing of a modified maximum-entropy model with a non-uniform reference measure to account for inhibition.

Main Results:

• Standard pairwise models produce unrealistic bimodal population-averaged activity distributions, with a mode suggesting high population activity.
• This bimodality problem affects models of neuronal populations of 200+ neurons and hinders learning and data generation.
• The modified model, incorporating a non-uniform reference measure, successfully eliminates unrealistic bimodalities and improves model behavior.

Conclusions:

• Standard pairwise maximum-entropy models fail to accurately represent neuronal population activity due to their inability to model inhibition.
• A modified maximum-entropy model with a non-uniform reference measure provides a more biologically plausible and computationally tractable approach.
• This revised model offers a promising solution for predicting neuronal population dynamics, especially in the presence of significant inhibition.