Model Based Inference of Large Scale Brain Networks with Approximate Bayesian Computation

Brain networks and the neural dynamics that unfold upon them are of great interest across the many scales of systems neuroscience. The tools of inverse modelling provide a way of both constraining and selecting models of large scale brain networks from empirical data. Such models have the potential to yield broad theoretical insights in the understanding of the physiological processes behind the integration and segregation of activity in the brain. In order to make inverse modelling computationally tractable, simplifying model assumptions have often been adopted that appeal to steady-state approximations to neural dynamics and thus prevent the investigation of stochastic or intermittent dynamics such as gamma or beta burst activity. In this work we describe a framework that uses the Approximate Bayesian Computation (ABC) algorithm for the inversion of neural models that can flexibly represent any statistical feature of empirically recorded data and eschew the need to assume a locally linearized system. Further, we demonstrate how Bayesian model comparison can be applied to fitted models to enable the selection of competing hypotheses regarding the causes of neural data. This work establishes a validation of the procedures by testing for both the face validity (i.e. the ability to identify the original model that has generated the observed data) and predictive validity (i.e. the consistency of the parameter estimation across multiple realizations of the same data). From the validation and example applications presented here we conclude that the proposed framework provides a novel opportunity to researchers aiming to explain how complex brain dynamics emerge from neural circuits.