Functional integration and inference in the brain.

TLDR: It will be shown that functional neuroimaging can be used to test for interactions between bottom-up and top-up inputs to an area and the prevalence of top-down influences and the plausibility of generative models of sensory brain function are pointed toward.

About: This article is published in Progress in Neurobiology.The article was published on 2002-10-01 and is currently open access. It has received 302 citations till now. The article focuses on the topics: Functional integration (neurobiology) & Perceptual learning.

Figures

Fig. 5. Top: Examples of the stimuli presented to subjects. During the measurement of brain responses only degraded stimuli where shown (e.g. the right hand picture). In half the scans the subject was given the underlying cause of these stimuli, through presentation of the original picture (e.g. left)before scanning. This priming induced a profound difference in perceptual set for the primed, relative to non-primed, stimuli,Right: Activity observed in a right infero-temporal region, as a function of (mean corrected) PPC activity. This region showed the most significant interaction between the presence of faces in visually presented stimuli and activity in a reference location in the posterior medial parietal cortex (PPC). This analysis can be thought of as finding those areas that are subject to top–down modulation of face-specific responses by medial parietal activity. The crosses correspond to activity whilst viewing non-face stimuli and the circles to faces. The essence of this effect can be seen by noting that this region differentiates between faces and non-faces when, and only when, medial parietal activity is high. The lines correspond to the best second-order polynomial fit. These data were acquired from six subjects using PET.Left: Schematic depicting the underlying conceptual model in which driving afferents from ventral form areas (here designated as V4) excite infero-temporal (IT) responses, subject to permissive modulation by PPC projections.

Fig. 6. Left: Brain regions and connections comprising the model.Right: Characterisation of the effects of V2 inputs on V5 and their modulation by posterior parietal cortex (PPC). The broken lines represent estimates of V5 responses when PPC activity is zero, according to a second-order Volterra model of effective connectivity with inputs to V5 from V2, PPC and the pulvinar (PUL). The solid curves represent the same response when PPC activity is one standard deviation of its variation over conditions. It is evident that V2 has an activating effect on V5 and that PPC increases the responsivenes of V5 to these inputs. The insert shows all the voxels in V5 that evidenced a modulatory effect (P < 0.05 uncorrected). These voxels were identified by thresholding a SPM (Friston et al., 1995b) of the F statistic testing for the contribution of second-order kernels involving V2 and PPC (treating all other terms as nuisance variables). The data were obtained with fMRI under identical stimulus conditions (visual motion subtended by radially moving dots) whilst manipulating the attentional component of the task (detection of velocity changes).

Fig. 3. Schematic illustrating the role of priors in biasing towards one representation of an input or another.Left: The word ‘event’ is selected as the most likely cause of the visual input.Right: The word ‘went’ is selected as the most likely word that is: (i) a reasonable explanation for the sensory input; and (ii) conforms to prior expectations induced by semantic context.

Fig. 2. Schematic depicting a hierarchical extension to the predictive coding architecture, using the same format asFig. 1. Here hierarchical arrangements within the model serve to provide predictions or priors to representations in the level below. The open circles are the error units and the filled circles are the representations of causes in the environment. These representations change to minimise both the discrepancy between their predicted value and the mismatch incurred by their own prediction of the representations in the level below. These two constraints correspond to prior and likelihood potentials, respectively (see main text).

Fig. 7. (a)Top: These renderings illustrate the extent of cerebral infarcts in four patients, as identified by voxel-based morphometry. Regions of reduced grey matter (relative to neurologically normal controls) are shown in white on the left hemisphere. The SPMs (Friston et al., 1995b) were thresholded at P < 0.001 uncorrected. All patients had damage to Broca’s area. The first (upper left) patient’s left middle cerebral artery infarct was most extensive encompassing temporal and parietal regions as well as frontal and motor cortex. (b)Bottom: SPMs illustrating the functional imaging results with regions of significant activation shown in black on the left hemisphere. Results are shown for: (i) normal subjects reading words (left); (ii) activations common to normal subjects and patients reading words using a conjunction analysis (middle-top); (iii) areas where normal subjects activate significantly more than patients reading words, using the group times condition interaction (middle lower); and (iv) the first patient activating normally for a semantic task. Context-sensitive failures to activate are implied by the abnormal activations in the first patient, for the implicit reading task, despite a normal activation during a semantic task.

Fig. 1. Schematic illustrating the architectures implied by supervised, information theory-based approaches and predictive coding. The circles represent nodes in a network and the arrows represent a few of the connections. See the main text for an explanation of the equations and designation of the variables each set of nodes represents. The light grey boxes encompass connections and nodes within the model. Connection strengths are determined by the free parameters of the modelφ (forward connections) andθ (backward connections). Nonlinear effects are implied when one arrow connects with another. Nonlinearities can be construed as the modulation of responsiveness to one input by another (see Box 1 for a more formal account). The broken arrow in the lower panel denotes connections that convey an error signal to the higher level from the input level.

Frequently asked questions

Q1. What have the authors contributed in "Functional integration and inference in the brain" ?

This review describes the two classes of models and their implications for the functional anatomy of sensory cortical hierarchies in the brain. The authors then consider how empirical evidence can be used to disambiguate between architectures that are sufficient for perceptual learning and synthesis. Using the notion of empirical Bayes, the authors show that these assumptions are not necessary and that priors can be learned in a hierarchical context. Furthermore, the authors try to show that learning can be implemented in a biologically plausible way. The main point made in this review is that backward connections, mediating internal or generative models of how sensory inputs are caused, are essential if the process generating inputs can not be inverted. This review summarises approaches to integration in terms of effective connectivity and proceeds to address the question posed by the theoretical considerations above. This enforces an explicit parameterisation of generative models ( i. e. backward connections ) to enable approximate recognition and suggests that feedforward architectures, on their own, are not sufficient.