Dynamic causal modelling of precision and synaptic gain in visual perception - an EEG study

Abstract

Estimating the precision or uncertainty associated with sensory signals is an important part of perception. Based on a previous computational model, we tested the hypothesis that increasing visual contrast increased the precision encoded in early visual areas by the gain or excitability of superficial pyramidal cells. This hypothesis was investigated using electroencephalography and dynamic causal modelling (DCM); a biologically constrained modelling of the cortical processes underlying EEG activity. Source localisation identified the electromagnetic sources of visually evoked responses and DCM was used to characterise the coupling among these sources. Bayesian model selection was used to select the most likely connectivity pattern and contrast-dependent changes in connectivity. As predicted, the model with the highest evidence entailed increased superficial pyramidal cell gain in higher-contrast trials. As predicted theoretically, contrast-dependent increases were reduced at higher levels of the hierarchy. These results demonstrate that increased signal-to-noise ratio in sensory signals produce (or are represented by) increased superficial pyramidal cell gain, and that synaptic parameters encoding statistical properties like sensory precision can be quantified using EEG and dynamic causal modelling.

Fig. 1

The Craik-O'Brien-Cornsweet (CBC) illusion. Upper panel: A demonstration of the CBC illusion. The two side panels have identical luminance. Close to the shared edge, there is a ramp of increasing luminance on the left and decreasing luminance on the right, which gives rise to the illusory percept that the panels have constant luminance and that there is a luminance step between them. Occluding the luminance ramps destroys this effect. Lower panel: psychophysical and simulated data from Brown and Friston (in submission). The black points are psychophysical data from a behavioural matching paradigm, in which the contrast of the stimulus luminance was varied and stimuli were matched to a real luminance step. The red points are simulated responses to the same stimuli, using a generalised predictive coding scheme; where the simulated and real psychophysical results have been scaled to match as closely as possible. Gamma values correspond to the log-precision of the (simulated) sensory input. Increasing the precision of sensory input reproduces the expression of the CBC illusion in human observers as visual contrast increases.

Fig. 2

Source localisation. Sources located by source reconstruction using multiple sparse priors and group constraints. The figures show absolute source activity averaged across subjects; the maxima were used as source locations for DCM. Four locations emerged bilaterally: the inferior occipital gyrus, the inferior parietal cortex, the superior occipital gyrus and the superior orbital gyrus.

Fig. 3

Results of fixed-effects Bayesian Model Selection. Upper panel: out of the different extrinsic connectivity models, Model 5, a serial hierarchy with interhemispheric connections, had the most evidence. This model was used for subsequent analyses.

Fig. 4

Prediction error in the cortical hierarchy. This figure shows the activity reconstructed at each of the sources used for DCM analysis (based on a DCM of the grand average event related potentials over subjects). These responses can be taken to be a rough proxy for prediction error, since superficial pyramidal cells contribute most of the EEG signal. The difference in signal between high-contrast and low-contrast clearly reduces as the hierarchy is ascended, reflecting the decreasing differences in the precision of prediction error.

Fig. 5

Model fits. The fits of the three models of contrast-dependent effects to event related potentials in sensor-space for an illustrative subject; these responses are summarised with the first two principal components or modes. The modes are used for data reduction — the data are projected onto the principal eigenvectors of the prior covariance of the data. In this paper, eight modes are used in total. The dashed lines show the data modes and the solid lines the model predictions. In the best-fitting model (centre) these are almost superimposed, whereas in the less well-fitting models, substantial differences are evident.

Fig. 6

Contrast-dependent changes in the game of superficial pyramidal cells. These are the average parameters, over subjects, controlling the contrast dependent changes in negative self-inhibition (gain) under the winning model of the previous figures. Note the progressive decrease in contrast-dependent effects at higher levels of the hierarchy. This is predicted theoretically, because we have manipulated the precision of prediction errors at the lowest (sensory level) through experimental manipulations of visual contrast.

Fig. 7

CBC stimuli used for this study. These stimuli were created by applying a bandpass filter to white noise to create a random blob pattern with a fundamental frequency of 67 blobs/image (1 cycle/degree). This pattern was thresholded and convolved with a 2-D Laplacian-of-Gaussian filter to produce a CBC stimulus. Stimuli were scaled to have 90% (top), 25% (middle) or 10% (bottom) of the maximum contrast supported by the monitor. The stimuli subtended approximately 32° of visual angle. The central 2° of visual angle were left blank. Stimuli were presented against a grey background on a gamma-corrected monitor.

Fig. 8

Model selection. Initial model selection was carried out to identify the extrinsic connectivity pattern on the sources identified. Only plausible models were tested; these models had the inferior orbital gyrus at the bottom of the hierarchy and the superior orbital gyrus at the top (Felleman and Van Essen, 1991). These models are distinguished by the deployment of forward and backward extrinsic connections, as determined by their level in the hierarchy. Here, this level corresponds to vertical position.