Abstract
Relational cognition — the ability to infer relationships that generalize to novel combinations of objects — is fundamental to human and animal intelligence. Despite this importance, it remains unclear how relational cognition is implemented in the brain due in part to a lack of hypotheses and predictions at the levels of collective neural activity and behavior. Here we discovered, analyzed, and experimentally tested neural networks (NNs) that perform transitive inference (TI), a classic relational task (if A > B and B > C, then A > C). We found NNs that (i) generalized perfectly, despite lacking overt transitive structure prior to training, (ii) generalized when the task required working memory (WM), a capacity thought essential to inference in the brain, (iii) emergently expressed behaviors long observed in living subjects, in addition to a novel order-dependent behavior, and (iv) adopted different task solutions yielding alternative behavioral and neural predictions. Further, in a large-scale experiment, we found that human subjects performing WM-based TI showed behavior that matched an unorthodox subset of the NNs. These findings provide neural insights into a classical relational ability, with wider implications for how the brain realizes relational cognition.
Competing Interest Statement
The authors have declared no competing interest.
Footnotes
- An important error in the original version has been corrected: item 1 vs. 2 identity was erroneously reversed in plots and analyses. Thus, in the updated version, all order-dependent findings are reversed, though basic findings (e.g. TI generalization, behaviors not dependent on item order, structure of subtractive solution) are the same. All plots and analyses have been corrected in the updated version. - Several additional main results have been added: results providing an explanation of the order-dependent behavior (Fig. 7), and behavioral data from human subjects performing delay TI (Fig. 8). - Results from RNNs trained on extended and variable delay periods are now reported (Table 5, Fig. S3). - Analysis of activity geometry now includes examples of activity configurations and trajectories (Fig. 6 and supplemental figures), in addition to examples and discussion of 'V' shaped activity geometry. - Larger numbers of model instances have been run and analyzed (up to 200 / model variant). - A simplified and further clarified version of the population-level 'subtractive' solution is now presented in Fig. 5e and a Supplemental Diagram. - Main predictions and implications (Table 4) have been both further clarified and updated. - Text and figures throughout the study have been refined and reorganized for greater clarity. - Various fixes and clarifications have been made in the Methods.
