Near-optimal Integration of Magnitude in the Human Parietal Cortex

Abstract

Abstract■ Humans are often observed to make optimal sensorimotor decisions but to be poor judges of situations involving explicit estimation of magnitudes or numerical quantities. For example, when drawing conclusions from data, humans tend to neglect the size of the sample from which it was collected. Here, we asked whether this sample size neglect is a general property of human decisions and investigated its neural implementation. Participants viewed eight discrete visual arrays (samples) depicting variable numbers of blue and pink balls. They then judged whether the samples were being drawn from an urn in which blue or pink predominated. A participant who neglects the sample size will integrate the ratio of balls on each array, giving equal weight to each sample. However, we found that human behavior resembled that of an optimal observer, giving more credence to larger sample sizes. Recording scalp EEG signals while participants performed the task allowed us to assess the decision information that was computed during integration. We found that neural signals over the posterior cortex after each sample correlated first with the sample size and then with the difference in the number of balls in either category. Moreover, lateralized beta-band activity over motor cortex was predicted by the cumulative difference in number of balls in each category. Together, these findings suggest that humans achieve statistically near-optimal decisions by adding up the difference in evidence on each sample, and imply that sample size neglect may not be a general feature of human decision-making. ■

Optional Stopping in a Heteroscedastic World

“…9a (parietal) and 9b (frontal). Encoding of | | exhibited a characteristic negative-positive deflection at ~200/450ms over parietal electrodes (red trace; upper panels), and the sign reversed pattern in frontal electrodes (lower panels) that closely replicates the encoding of decision information described previously (Tickle et al, 2016;Wyart et al, 2012;Wyart et al, 2015).…”

Selective Integration during Sequential Sampling in Posterior Neural Signals

“…In the SI model, samples are first weighted according to their relative value, and then integrated into a cumulative signal that eventually determines choice. Previous studies have observed in both perceptual and value-based decision making tasks that central EEG signals build up in concert with the strength of accumulated evidence (Donner et al, 2009;van Vugt et al, 2012;Gluth et al, 2013;Wyart et al, 2015;Tickle et al, 2016;Pisauro et al, 2017;von Lautz et al, 2019). We used a time-window several hundreds of milliseconds before the response onset to test for encoding of the accumulated evidence, as predicted by the SI model.…”

Humans treat unreliable filled-in percepts as more real than veridical ones

“…This has been best demonstrated in cue-integration tasks in which humans perform as if they were weighting the different sources of information according to their respective reliabilities (Ernst and Banks, 2002; Alais and Burr, 2004; Körding and Wolpert, 2004; Tickle et al, 2016). This form of statistical inference has also been demonstrated for cortical neurons of the monkey brain, with patterns of activity at the population level that are consistent with the implementation of a probabilistic population code (Gu et al, 2008; Fetsch et al, 2011).…”