Brain representations separate stored relationships from the demands of a task
Brain scans from 27 people suggest that learned relationships can retain a shared structure while other representations emphasize the information a decision needs. Participants compared faces ranked by two attributes. The peer-reviewed reanalysis paired imaging with artificial neural networks, where researchers could alter contextual input. Those model experiments test a possible mechanism rather than demonstrating causation in human brains.
Science··Night
A colour cue changes which ranking matters
A decision can require one aspect of knowledge while another aspect remains irrelevant. In a single peer-reviewed reanalysis, 27 young participants learned how 16 faces ranked for competence and popularity. Training showed neighbouring pairs; later tests required comparisons between faces not previously paired. A colour signalled the relevant attribute. Participants were never told that the relationships formed a two-dimensional grid, allowing researchers to examine how learned relationships were represented during inference.[1]
Shared maps coexist with selective representations
Functional magnetic resonance imaging, which tracks changes associated with brain activity, identified different representational patterns. The hippocampus and entorhinal cortex, regions involved in memory, and parts of the prefrontal cortex retained information about both ranking dimensions. Other frontal and parietal regions emphasized the relevant dimension while compressing the irrelevant one. Comparisons became harder when the two attributes called for conflicting responses. Differences in map geometry were associated with that behavioural difficulty.[1]
Network interventions test the role of context
Artificial recurrent neural networks learned the same ranking task and developed comparable representational patterns. Altering their contextual input changed learning and map geometry, providing an intervention inside the model. The human imaging results remained associative. Training also differed: people learned across separate days, whereas the network encountered interleaved tasks. The small sample and constructed face-ranking exercise constrain how far the findings extend to decisions outside the experiment.[1]
