Towards a Dynamic Causal Modeling approach by using Bayesian Convolutional Networks and Stochastic Variational Inference

Thumbnail Image
Issue Date
Journal Title
Journal ISSN
Volume Title
The identification of the neural dependencies and hierarchical representations in the brain constitutes one of the principal goals of neuroscience. We propose a Dynamic Causal Modeling (DCM) approach that utilizes convolutional neural networks and stochastic variational inference. We employ the Neural Information Flow (NIF) that describes the visual cortex areas’ hierarchy and neuronal activity through convolutional layers and an observation model, respectively. We convert the model into Bayesian NIF by introducing probability distributions over all the learnable parameters and estimate the latent parameters’ posterior densities by maximizing the model’s evidence lower bound (ELBO) via stochastic variational inference. All the Bayesian NIF parameters have biophysical interpretation. Stochastic variational inference supports the application of complex Bayesian models to large datasets, whereas the estimated ELBO can be used for Bayesian model comparison like in DCM. Therefore, our proposed approach points to a better explained DCM that can manage large-scale fMRI datasets. We demonstrate our method’s capacity to predict accurate neural responses by applying a large-scale functional magnetic resonance dataset over a simple model architecture representing the early visual cortex. Our results imply that the Bayesian NIF model can predict brain-like responses while successfully capturing the observational model’s receptive field representations. Further, the ELBO converges and thus can be used for Bayesian model comparison with different model architectures.
Faculteit der Sociale Wetenschappen