Precision and neuronal dynamics in the human posterior parietal cortex during evidence accumulation

Abstract

Primate studies show slow ramping activity in posterior parietal cortex (PPC) neurons during perceptual decision-making. These findings have inspired a rich theoretical literature to account for this activity. These accounts are largely unrelated to Bayesian theories of perception and predictive coding, a related formulation of perceptual inference in the cortical hierarchy. Here, we tested a key prediction of such hierarchical inference, namely that the estimated precision (reliability) of information ascending the cortical hierarchy plays a key role in determining both the speed of decision-making and the rate of increase of PPC activity. Using dynamic causal modelling of magnetoencephalographic (MEG) evoked responses, recorded during a simple perceptual decision-making task, we recover ramping-activity from an anatomically and functionally plausible network of regions, including early visual cortex, the middle temporal area (MT) and PPC. Precision, as reflected by the gain on pyramidal cell activity, was strongly correlated with both the speed of decision making and the slope of PPC ramping activity. Our findings indicate that the dynamics of neuronal activity in the human PPC during perceptual decision-making recapitulate those observed in the macaque, and in so doing we link observations from primate electrophysiology and human choice behaviour. Moreover, the synaptic gain control modulating these dynamics is consistent with predictive coding formulations of evidence accumulation.

Fig. 1

Illustration of candidate neuronal mechanisms underlying behavioural and neuronal response time variability. Our model contains three neuronal populations, input receiving spiny stellate cells, pyramidal cells that send extrinsic forward connections to other cortical regions, and inhibitory interneurons (Jansen and Rit, 1995; David and Friston, 2003). The strength (gain) of connections between these populations is parameterised by γ_{1 − 4}. Based on predictive coding, we hypothesised that response time variability would be driven by changes in the precision of ascending sensory information, which is operationalised in our model as the gain (γ₂) on the connection between input receiving stellate cells and the pyramidal cell population (left). We compared models in which the strength of intrinsic gain changed with response time, compared to changes in extrinsic forward connections between regions (right), and a null model in which response time had no effect (not shown).

Fig. 2

A. Top left panel: Illustration of a single trial of the direction discrimination task. A red fixation cross appeared, surrounded by a field of stationary dots. After an interval, jittered between 1000 and 1500 ms, the dots began to move. A subset of the dots moved coherently, and the rest moved randomly. Subjects had 3000 ms to make their decision; indicating whether they thought the direction of coherent motion was to the left or to the right. Dots disappeared from the screen once the decision was made, and a blank screen was shown for the remainder of 3000 ms, plus a further 1000 ms. B. Top right panel: Bar plot showing the mean coherence level of trials in each of the three reaction time bins used to analyse the behavioural data. Reaction time and motion coherence level showed a strong negative correlation. (Red: RT bin 1 (fastest). Green: RT bin 2 (intermediate). Blue: RT bin 3 (slowest)) (error bars indicate bootstrapped 95% confidence intervals). C. Bottom left panel: Average reaction time for each coherence level averaged across subjects. Reaction times showed a strong negative relationship with coherence level. (Error bars indicate bootstrapped 95% confidence intervals). D. Bottom right panel: Choice behaviour averaged across subjects. Both direction of motion and coherence level strongly affected choice behaviour, which was well fit with a sigmoid function. (Error bars indicate bootstrapped 95% confidence intervals).

Fig. 3

A. Top row: Topoplots illustrating grand mean evoked fields, averaged across all time bins, at 50, 250 and 450 ms. B. Bottom row: Topoplots illustrating the effect of response speed (faster responses minus slower) on grand mean evoked fields at 50, 250 and 450 ms. Response speed shows a similar topography to evoked fields averaged across conditions, consistent with its correlations with stronger (and faster rising) activity in the same cortical areas.

Fig. 4

A. Left panel: Active sources from MSP reconstruction of evoked responses from 0 to 500 ms (the image shows the 200 most activated voxels). Three bilateral sources were found, in the early visual cortex (VC), the middle temporal area (MT) and the posterior parietal cortex (PPC). B. Right panel: Winning network structure from our dynamic causal modelling analysis. This reveals a plausible hierarchy in which PPC sits at the top, VC at the bottom, and MT in between.

Fig. 5

Models used for initial network selection. We tested 10 anatomically plausible models, in linear and parallel hierarchies, with and without lateral connections. Regions at the same level in the figure are connected by lateral connections; regions at different levels are connected by both forward connections (running from lower to higher levels) and backward connections (running from higher to lower).

Fig. 6

Normalised pyramidal cell activity averaged across subjects for each of the sources in the winning DCM. Posterior parietal cortex (rightmost figure) showed slow ramping activity modulated by response time, as predicted for a putative decision network, but this was not observed in any other region. (Red: RT bin 1. Green: RT bin 2. Blue: RT bin 3. Dotted lines indicate bootstrapped 95% confidence intervals).

Fig. 7

A: Top panel. Sensitivity analysis. To provide a qualitative illustration of the different predictions of gain and connectivity model families, we simulated the effect of small changes in model parameters on activity in the left PPC (all parameters other than those being perturbed were set to their posterior expectations). Altering the gain on the intrinsic connection within the PPC produced changes that increase smoothly (solid line), whereas altering the strength of the forward connection from MT to PPC produces an initial dip followed by a faster increase in activity which rapidly reaches a plateau (dotted line) (similar patterns were observed in right PPC). B: Bottom panel. Model fits. The fits of the gain (left) and forward connection (right) model families in sensor space for an illustrative subject. Responses are summarised by the first two principal spatial modes (used for data reduction — we used eight modes in total). Solid lines indicate predicted responses, dotted lines indicate observed data. The gain model family provided a significantly better fit to the data, as illustrated here by a closer correlation between predicted and observed responses (Red: RT bin 1. Green: RT bin 2. Blue: RT bin 3).