This note derives the variational free energy under the Laplace approximation, with a focus on accounting for additional model complexity induced by increasing the number of model parameters. This is relevant when using the free energy as an approximation to the log-evidence in Bayesian model averaging and selection. By setting restricted maximum likelihood (ReML) in the larger context of variational learning and expectation maximisation (EM), we show how the ReML objective function can be adjusted to provide an approximation to the log-evidence for a particular model. This means ReML can be used for model selection, specifically to select or compare models with different covariance components. This is useful in the context of hierarchical models because it enables a principled selection of priors that, under simple hyperpriors, can be used for automatic model selection and relevance determination (ARD). Deriving the ReML objective function, from basic variational principles, discloses the simple relationships among Variational Bayes, EM and ReML. Furthermore, we show that EM is formally identical to a full variational treatment when the precisions are linear in the hyperparameters. Finally, we also consider, briefly, dynamic models and how these inform the regularisation of free energy ascent schemes, like EM and ReML.

No abstract

This paper presents a variational treatment of dynamic models that furnishes time-dependent conditional densities on the path or trajectory of a system's states and the time-independent densities of its parameters. These are obtained by maximising a variational action with respect to conditional densities, under a fixed-form assumption about their form. The action or path-integral of free-energy represents a lower bound on the model's log-evidence or marginal likelihood required for model selection and averaging. This approach rests on formulating the optimisation dynamically, in generalised coordinates of motion. The resulting scheme can be used for online Bayesian inversion of nonlinear dynamic causal models and is shown to outperform existing approaches, such as Kalman and particle filtering. Furthermore, it provides for dual and triple inferences on a system's states, parameters and hyperparameters using exactly the same principles. We refer to this approach as dynamic expectation maximisation (DEM).

We describe a Bayesian estimation and inference procedure for fMRI time series based on the use of General Linear Models (GLMs). Importantly, we use a spatial prior on regression coefficients which embodies our prior knowledge that evoked responses are spatially contiguous and locally homogeneous. Further, using a computationally efficient Variational Bayes framework, we are able to let the data determine the optimal amount of smoothing. We assume an arbitrary order Auto-Regressive (AR) model for the errors. Our model generalizes earlier work on voxel-wise estimation of GLM-AR models and inference in GLMs using Posterior Probability Maps (PPMs). Results are shown on simulated data and on data from an event-related fMRI experiment. D 2004 Elsevier Inc. All rights reserved.

No abstract

In this paper, the Bayesian Theory is used to formulate the Inverse Problem (IP) of the EEG/MEG. This formulation offers a comparison framework for the wide range of inverse methods available and allows us to address the problem of model uncertainty that arises when dealing with different solutions for a single data. In this case, each model is defined by the set of assumptions of the inverse method used, as well as by the functional dependence between the data and the Primary Current Density (PCD) inside the brain. The key point is that the Bayesian Theory not only provides for posterior estimates of the parameters of interest (the PCD) for a given model, but also gives the possibility of finding posterior expected utilities unconditional on the models assumed. In the present work, this is achieved by considering a third level of inference that has been systematically omitted by previous Bayesian formulations of the IP. This level is known as Bayesian model averaging (BMA). The new approach is illustrated in the case of considering different anatomical constraints for solving the IP of the EEG in the frequency domain. This methodology allows us to address two of the main problems that affect linear inverse solutions (LIS): (a) the existence of ghost sources and (b) the tendency to underestimate deep activity. Both simulated and real experimental data are used to demonstrate the capabilities of the BMA approach, and some of the results are compared with the solutions obtained using the popular lowresolution electromagnetic tomography (LORETA) and its anatomically constraint version (cLORETA). IntroductionOur interest lies in the identification of electro/magnetoencephalogram (EEG/MEG) generators, that is, the distribution of current sources inside the brain that generate the voltage -magnetic field measured over an array of sensors distributed on the scalp surface. This is known as the Inverse Problem (IP) of the EEG/ MEG.Much literature has been devoted to the solution of this problem. The main difficulty stems from its ill-posed character due to the nonuniqueness of the solution, which is caused by the existence of silent sources that cannot be measured over the scalp surface. Additional complications that arise when dealing with actual data are related to the limited number of sensors available, making the problem highly underdetermined, as well as to the numerical instability of the solution, given by its high sensitivity to measurement noise.The usual way to deal with these difficulties has been to include additional information or constraints about the physical and mathematical properties of the current sources inside the head, which limit the space of possible solutions. This has resulted in the emergence of a great variety of methods, each depending on the kind of information that has been introduced and resulting consequently in many different unique solutions.Some methods handle the many-to-one nature of the problem by characterizing the sources in terms of a limited number of current dipoles that are...

We study the generation of EEG rhythms by means of realistically coupled neural mass models. Previous neural mass models were used to model cortical voxels and the thalamus. Interactions between voxels of the same and other cortical areas and with the thalamus were taken into account. Voxels within the same cortical area were coupled (short-range connections) with both excitatory and inhibitory connections, while coupling between areas (long-range connections) was considered to be excitatory only. Short-range connection strengths were modeled by using a connectivity function depending on the distance between voxels. Coupling strength parameters between areas were defined from empirical anatomical data employing the information obtained from probabilistic paths, which were tracked by water diffusion imaging techniques and used to quantify white matter tracts in the brain. Each cortical voxel was then described by a set of 16 random differential equations, while the thalamus was described by a set of 12 random differential equations. Thus, for analyzing the neuronal dynamics emerging from the interaction of several areas, a large system of differential equations needs to be solved. The sparseness of the estimated anatomical connectivity matrix reduces the number of connection parameters substantially, making the solution of this system faster. Simulations of human brain rhythms were carried out in order to test the model. Physiologically plausible results were obtained based on this anatomically constrained neural mass model.

The remarkable capabilities displayed by humans in making sense of an overwhelming amount of sensory information cannot be explained easily if perception is viewed as a passive process. Current theoretical and computational models assume that to achieve meaningful and coherent perception, the human brain must anticipate upcoming stimulation. But how are upcoming stimuli predicted in the brain? We unmasked the neural representation of a prediction by omitting the predicted sensory input. Electrophysiological brain signals showed that when a clear prediction can be formulated, the brain activates a template of its response to the predicted stimulus before it arrives to our senses.