Sorting based on partial information: behavior, computational modeling, and neural evidence

Poster Session E - Monday, April 15, 2024, 2:30 – 4:30 pm EDT, Sheraton Hall ABC

Dongning Liu^1,2,3, Muzhi Wang^1,2,3, Huan Luo^1,2,3; ¹School of Psychological and Cognitive Sciences, Peking University, ²PKU-IDG/McGovern Institute for Brain Research, Peking University, ³Beijing Key Laboratory of Behavior and Mental Health, Peking University

Our daily lives are full of sorting activities, from assigning ranks to job candidates to filtering information by relevance. Meanwhile, information is always limited and partial, and how human subjects infer full-scope ranks from fragmented inputs remains unknown. Here we developed a new behavioral paradigm combined with magnetoencephalography (MEG) recordings to examine the underlying computational and neural mechanisms. Participants were instructed to learn the artificial "popularity" ranking of eight films (denoted by eight images) from brief exposure to eight pairwise comparisons (e.g., Film A is two units more popular than Film B). Crucially, the eight pairs are mostly non-adjacent along the rank score and thus only offer partial information (8 out of 28 pairs). Subjects were asked to infer the full-scale popularity of eight films after partial training. First, items close to the lowest or highest popularity score showed higher ranking accuracy (U-curve). Second, pairs with larger rank distances had higher sorting accuracy than those with closer ranks. Both the U-curve properties and distance effect could be well characterized using the Rescorla-Wagner model, in which participants incrementally learn the value of each film from observed pairs. Finally, MEG recordings show that, after training partial image pairs, evoked responses occurring in the right parietal region with approximately 980 ms latency carry the learned rank information. Overall, our behavioral and MEG studies consistently support that sorting with incomplete information can be achieved by a simple reinforcement learning model incorporating implicit value representations and value-updating rules to achieve rapid inference on ranks.

Topic Area: THINKING: Problem solving