A novel modeling approach to identify groups of coordinated neurons which accounts for high-order neural correlations

Perception and behavior are mediated by interconnected neurons that form neural circuits. Traditionally, groups of neurons exhibiting coordinated activity within these circuits have been detected through the use of dimensionality reduction techniques. Here, we proposed a novel approach that is based on modeling groups of coordinated neurons with minimally complex spin models. These models are maximum entropy models with interactions of arbitrary order that are grouped into "communities", with no interactions between the communities and all possible interactions within each community. With these models, we can perform exact Bayesian model selection. Besides, the novelty of our approach is that it accounts for high-order neural activity patterns (i.e. multi-neuronal activity motifs) in the detection of groups of correlated neurons. Additionally, we show that this technique provides robust predictions, without the need for statistical thresholding. We discuss the strengths and weaknesses of both the traditional approach and our approach, and illustrate our findings on artificial data and on real neural data obtained from the auditory cortex. Our new approach is a promising preliminary direction in the search for functional connectivity structures going beyond pairwise interactions.