Humans routinely make judgments about complex stimuli, such as faces, which requires an elaborate representation of the stimulus statistics. For instance, when attempting to recognize recently seen faces, we can make fine distinctions of facial features and can also use these representations to recognize fine changes, such as emotional characteristics, based on changes of these facial features. These challenging tasks cannot be solved based on the actual stimulus alone but require the support of internal representations. These internal representations gauge the likely and less likely combinations of facial features and use this knowledge to effectively asses actual stimuli.
While the presence of internal representations of stimulus statistics have strong support in the literature, demonstrating the defining features of such representations in humans remained elusive. The main challenge of an experimenter is to link stimulus features to the judgements of the participants. The most reliable information to assess internal representations of humans in this respect is to collect the decisions of participants from binary or ternary tasks. This information, however, is limited and ‘reading the minds’ from mere button presses proved to be a challenge.
We developed ‘cognitive tomography’ to gauge the characteristics of internal representations. Focusing on the fine details, which we call the statistics, of the patterns in human judgements we were able to reveal subject-specific internal representations. Since this representation could merely be a quantity loosely related to the true internal representation, our goal was to demonstrate that we are indeed reading the ‘mind’ of our subjects rather than a subject-specific mapping from stimulus to response. Fascinatingly, we were able to use cross-task prediction to demonstrate that ‘cognitive tomography’ indeed reveals the relevant internal model. In cross-task prediction we learn from the judgements of a subject in one task and use this knowledge to predict her judgements in the other task.
Related reading
Houlsby NMT, Huszár F, Ghassemi MM, Orbán G, Wolpert DM, Lengyel M (2013)
Cognitive tomography reveals complex task-independent mental representations.
Current Biololgy, 10.1016/j.cub.2013.09.012.
Török B, Nagy DG, Kiss M, Janacsek K*, Németh D*, Orbán G* (2022)
Tracking the contribution of inductive bias to individualised internal models
PLoS Computational Biology, 18(6): e1010182
*equal contributions