This paper concerns temporal filtering in fMRI time-series analysis. Whitening serially correlated data is the most efficient approach to parameter estimation. However, if there is a discrepancy between the assumed and the actual correlations, whitening can render the analysis exquisitely sensitive to bias when estimating the standard error of the ensuing parameter estimates. This bias, although not expressed in terms of the estimated responses, has profound effects on any statistic used for inference. The special constraints of fMRI analysis ensure that there will always be a misspecification of the assumed serial correlations. One resolution of this problem is to filter the data to minimize bias, while maintaining a reasonable degree of efficiency. In this paper we present expressions for efficiency (of parameter estimation) and bias (in estimating standard error) in terms of assumed and actual correlation structures in the context of the general linear model. We show that: (i) Whitening strategies can result in profound bias and are therefore probably precluded in parametric fMRI data analyses. (ii) Band-pass filtering, and implicitly smoothing, has an important role in protecting against inferential bias.
This paper describes an application of hierarchical or empirical Bayes to the distributed source reconstruction problem in electro- and magnetoencephalography (EEG and MEG). The key contribution is the automatic selection of multiple cortical sources with compact spatial support that are specified in terms of empirical priors. This obviates the need to use priors with a specific form (e.g., smoothness or minimum norm) or with spatial structure (e.g., priors based on depth constraints or functional magnetic resonance imaging results). Furthermore, the inversion scheme allows for a sparse solution for distributed sources, of the sort enforced by equivalent current dipole (ECD) models. This means the approach automatically selects either a sparse or a distributed model, depending on the data. The scheme is compared with conventional applications of Bayesian solutions to quantify the improvement in performance.
In this technical note, we derive two MCMC (Markov chain Monte Carlo) samplers for dynamic causal models (DCMs). Specifically, we use (a) Hamiltonian MCMC (HMC-E) where sampling is simulated using Hamilton's equation of motion and (b) Langevin Monte Carlo algorithm (LMC-R and LMC-E) that simulates the Langevin diffusion of samples using gradients either on a Euclidean (E) or on a Riemannian (R) manifold. While LMC-R requires minimal tuning, the implementation of HMC-E is heavily dependent on its tuning parameters. These parameters are therefore optimised by learning a Gaussian process model of the time-normalised sample correlation matrix. This allows one to formulate an objective function that balances tuning parameter exploration and exploitation, furnishing an intervention-free inference scheme. Using neural mass models (NMMs)-a class of biophysically motivated DCMs-we find that HMC-E is statistically more efficient than LMC-R (with a Riemannian metric); yet both gradient-based samplers are far superior to the random walk Metropolis algorithm, which proves inadequate to steer away from dynamical instability.
An important characteristic of spontaneous brain activity is the anticorrelation between the core default network (cDN) and the dorsal attention network (DAN) and the salience network (SN). This anticorrelation may constitute a key aspect of functional anatomy and is implicated in several brain disorders. We used dynamic causal modeling to assess the hypothesis that a causal hierarchy underlies this anticorrelation structure, using resting-state fMRI of healthy adolescent and young adults (N = 404). Our analysis revealed an asymmetric effective connectivity, such that the regions in the SN and DAN exerted an inhibitory influence on the cDN regions; whereas the cDN exerted an excitatory influence on the SN and DAN regions. The relative strength of efferent versus afferent connections places the SN at the apex of the hierarchy, suggesting that the SN modulates anticorrelated networks with descending hierarchical connections. In short, this study of directed neuronal coupling reveals a causal hierarchical architecture that generates or orchestrates anticorrelation of brain activity. These new findings shed light on functional integration of intrinsic brain networks at rest and speak to future dynamic causal modeling studies of large-scale networks.
Consciousness, Dreams, and Inference : The Cartesian Theatre Revisited. This paper considers the Cartesian theatre as a metaphor for the virtual reality models that the brain uses to make inferences about the world. This treatment derives from our attempts to understand dreaming and waking consciousness in terms of free energy minimization. The idea here is that the Cartesian theatre is not observed by an internal (homuncular) audience but furnishes a theatre in which fictive narratives and fantasies can be rehearsed and tested against sensory evidence. We suppose the brain is driven by the imperative to infer the causes of its sensory samples ; in much the same way as scientists are compelled to test hypotheses about experimental data. This recapitulates Helmholtz’s notion of unconscious inference and Gregory’s treatment of perception as hypothesis testing. However, we take this further and consider the active on our virtual reality. The ensuing picture of consciousness (or active inference) resolves a number of seemingly hard problems in consciousness research and is internally consistent with current thinking in systems neuroscience and theoretical neurobiology. In this formalism, there is a dualism that distinguishes between the (conscious) process of inference and the (material) process that entails inference. This separation is reflected by the distinction between beliefs (probability distributions over hidden world states or res cogitans) and the physical brain states (sufficient statistics or res extensa) that encode them. This formal approach allows us to appeal to simple but fundamental theorems in information theory and statistical thermodynamics that dissolve some of the mysterious aspects of consciousness.
Convolutional and Recurrent, deep neural networks have been successful in machine learning systems for computer vision, reinforcement learning, and other allied fields. However, the robustness of such neural networks is seldom apprised, especially after high classification accuracy has been attained. In this paper, we evaluate the robustness of three recurrent neural networks to tiny perturbations, on three widely used datasets, to argue that high accuracy does not always mean a stable and a robust (to bounded perturbations, adversarial attacks, etc.) system. Especially, normalizing the spectrum of the discrete recurrent network to bound the spectrum (using power method, Rayleigh quotient, etc.) on a unit disk produces stable, albeit highly non-robust neural networks. Furthermore, using the $\epsilon$-pseudo-spectrum, we show that training of recurrent networks, say using gradient-based methods, often result in non-normal matrices that may or may not be diagonalizable. Therefore, the open problem lies in constructing methods that optimize not only for accuracy but also for the stability and the robustness of the underlying neural network, a criterion that is distinct from the other.
Neurotransmitter receptor expression and neuronal population dynamics show regional variability across the human cortex. However, currently there is an explanatory gap regarding how cortical microarchitecture and mesoscopic electrophysiological signals are mechanistically related, limiting our ability to exploit these measures of brain (dys)function for improved treatments of brain disorder; e.g., epilepsy. To bridge this gap, we leveraged dynamic causal modelling (DCM) and fitted biophysically informed neural mass models to a normative set of intracranial EEG data. Subsequently, using a hierarchical Bayesian modelling approach, we evaluated whether model evidence improved when information about regional neurotransmitter receptor densities is provided. We then tested whether the inferred constraints - furnished by receptor density - generalise across different electrophysiological recording modalities. The neural mass models explained regionally specific intracranial EEG spectra accurately, when fitted independently. Incorporating prior information on receptor distributions, further improved model evidence, indicating that variability in receptor density explains some variance in cortical population dynamics. The output of this modelling provides a cortical atlas of neurobiologically informed intracortical synaptic connectivity parameters that can be used as empirical priors in future - e.g., patient specific - modelling, as demonstrated in a worked example (a single-subject mismatch negativity study). In summary, we show that molecular cortical characteristics (i.e., receptor densities) can be incorporated to improve generative, biophysically plausible models of coupled neuronal populations. This work can help to explain regional variations in human electrophysiology, may provide a methodological foundation to integrate multi-modal data, and might serve as a normative resource for future DCM studies of electrophysiology.
n/a.
Spontaneous fluctuations in resting state activity can change in response to experience-dependent plasticity and learning. Visual learning is fast and can be elicited in an MRI scanner. Here, we showed that a random dot motion coherence task can be learned within one training session. While the task activated primarily visual and parietal brain areas, learning related changes in neural activity were observed in the hippocampus. Crucially, even this rapid learning affected resting state dynamics both immediately after the learning and 24h later. Specifically, the hippocampus changed its coupling with the striatum, in a way that was best explained as a consolidation of early learning related changes. Our findings suggest that long-lasting changes in neuronal coupling are accompanied by changes in resting state activity.
Abstract To exhibit social intelligence, animals have to recognize who they are communicating with. One way to make this inference is to select among multiple internal generative models of each conspecific. This induces an interesting problem: when receiving sensory input generated by a particular conspecific, how does an animal know which internal model to update? We consider a theoretical and neurobiologically plausible solution that enables inference and learning under multiple generative models by integrating active inference and (online) Bayesian model selection. This scheme fits sensory inputs under each generative model. Model parameters are then updated in proportion to the probability it could have generated the current input (i.e., model evidence). We show that a synthetic bird who employs the proposed scheme successfully learns and distinguishes (real zebra finch) birdsongs generated by several different birds. These results highlight the utility of having multiple internal models to make inferences in complicated social environments.
To behave adaptively, an organism must balance the accurate maintenance of information stored in working memory with the ability to update that information when the context changes. This trade-off between fidelity and flexibility may depend upon the anticipated likelihood that updating will be necessary. To address the neurobiological basis of anticipatory optimization, we acquired functional magnetic resonance imaging data, while healthy human subjects performed a modified delayed-response task. This task used cues that predicted memory updating, with high or low probability, followed by a contingent updating or maintenance event. This enabled us to compare behavior and neuronal activity during conditions in which updating was anticipated with high and low probability, and measure responses to expected and unexpected memory updating. Based on the known role of dopamine in cognitive flexibility and working memory updating, we hypothesized that differences in anticipatory set would be manifest in the dopaminergic midbrain and striatum. Consistent with our predictions, we identified sustained activation in the dopaminergic midbrain and the striatum, associated with anticipations of high versus low updating probability. We also found that this anticipatory factor affected neural responses to subsequent updating processes, which suppressed, rather than elevated, midbrain and striatal activity. Our study addresses for the first time an important and hitherto understudied aspect of working memory.
