Abstract
Numerosity refers to the set size of a group of items. We previously showed that numerosity-selective neuronal populations are organized in topographic maps in parietal cortex (Harvey et al., 2013). Recently, we demonstrated that visual population receptive field tuning varies systematically across cortical depth, similar to the signal processing between cortical layers (Fracasso, et al., 2016). Here, we combine both methods to investigate cortical depth-dependent processing in higher-order cortex, specifically investigating whether numerosity-selective populations in parietal cortex follow a similar signal processing sequence across cortical depth.
Participants (n=7) viewed dot patterns containing 1-7 or 20 dots that systematically varied over time while recording sub-millimeter fMRI responses from a numerosity map in the parietal cortex (Harvey et al., 2013). We estimated preferred numerosity (numerosity eliciting the highest response) and tuning width (numerosity range that elicit responses) for each voxel using computational modelling (Dumoulin, Wandell, 2008; Harvey et al., 2013).
Even in sub-millimeter fMRI data, our model explains 28% (mean of best 20% of fits, range = 9-50%) variance of the data. We find topographic numerosity maps at all depths. Moreover, we show that tuning width increases with preferred numerosity, in line with previous findings (Harvey & Dumoulin, 2017; Harvey et al., 2013), and that this pattern is present at all cortical depths. We show that on average, tuning width follows an inverted U-shaped profile across cortical depth, indicating a sharpening of numerosity responses with increased laminar processing.
Our results suggest that information processing across cortical depth in higher-order cortex is analogous to information processing in early visual cortex, but with a sharpening in tuning width across cortical depth in the parietal cortex instead of broadening. Despite the apparent discrepancy, we speculate that in both cases the change in tuning width reflects specific processing and signal pooling to extract more abstract features.