It remains controversial whether the visual system encodes abstract relational roles such as Agent and Patient in visual events. The present experiment tested whether abstract role bindings induce priming effects across consecutive events. Each trial included a static target image preceded either by a brief silent video of a priming event or by an audio-visual presentation of an English sentence describing the same event. Example sentence: “The red goat on the left knocked down the blue goat on the right.” 64 videos counterbalanced 4 event types: launching, deforming, breaking, and a relationally ambiguous control. The set of static targets were the final frames of the same videos. The role bindings were either repeated, switched, or ambiguous across the target and prime. The dependent variable was the latency on a color-localization task (e.g., whether the red animal was on the left or on the right). Whereas the linguistic primes had no statistically significant effect on the latency of the visual task, the role bindings of the video primes did have an effect: The latency on unambiguous trials (which required role binding) was significantly greater than that on ambiguous trials (on which at least one component lacked clear relational roles). This suggests the visual system is sensitive to (the ambiguity of) the role bindings of abstract relations.

We investigated whether the representation of relational categories is different from that of featural categories. Earlier work has suggested an extreme-value hypothesis: when a category is defined in terms of a relation, exemplars with exaggerated values along this stimulus dimension are judged as better members of the category. Featural categories, on the other hand, are not exaggerated. To test this hypothesis, we trained participants to categorize two fictional diseases defined either by a deterministic relation or a deterministic feature. After the categorization task was mastered up to a predefined learning criterion, we provided a graphical user interface that enabled participants to construct good examples of the acquired categories by adjusting the stimulus attributes. We constructed a novel index of relational exaggeration based on residual deviations from a non-exaggerated response strategy. These results supported the extreme-value hypothesis. This replicates and extends an earlier quasi-experimental study (Du et al., 2021).

The mental representation of relation-based concepts is different from that of feature-based concepts. In the present experiment, participants learned to categorize two fictional diseases that were defined either by a feature (e.g., short cells) or an ordinal relation (e.g., diseased cells being shorter than healthy cells). After the participants learned the categorization task to criterion, their strategies were probed in transfer task in which features and relations were pitted against one another. Finally, participants engaged in a stimulus reconstruction task. The results supported the prediction that participants who had adopted a feature-based strategy on a stimulus dimension, as identified by transfer data, tended to reconstruct values close to the means presented during training. By contrast, participants who had adopted a relation-based strategy tended to exaggerate that dimension away from the mean of the training examples and in the direction of the category-defining comparative relation. These data add to the growing literature suggesting that, unlike featural categories, relational categories are not represented in terms of the category’s central tendency.

In a classic paper, Miller (1956) summarized findings showing that people can only identify a limited number of distinctstimuli at a time. One puzzling aspect of this capacity limitation is that it is approximately invariant to range. Thatis, the number of accurately identifiable stimuli is approximately the same regardless of how far apart the stimuli arespaced. Models of this phenomenon have suggested that people operate in a context-coding mode when performing thesetasks, effectively carrying out a form of contextual normalization, but why such normalization might take place is unclear.Here, we propose an explanation by appealing to a tradeoff with generalization. Specifically, we implement contextualnormalization in a recurrent neural network and show that this normalization enables stronger generalization in a relationalreasoning task, but also results in a perceptual capacity limitation which captures many of these classic phenomena.

Human analogical ability involves the re-use of abstract, struc-tured representations within and across domains. Here, wepresent a generative neural network that completes analogiesin a 1D metric space, without explicit training on analogy.Our model integrates two key ideas. First, it operates overrepresentations inspired by properties of the mammalian En-torhinal Cortex (EC), believed to extract low-dimensional rep-resentations of the environment from the transition probabil-ities between states. Second, we show that a neural networkequipped with a simple predictive objective and highly generalinductive bias can learn to utilize these EC-like codes to com-pute explicit, abstract relations between pairs of objects. Theproposed inductive bias favors a latent code that consists ofanti-correlated representations. The relational representationslearned by the model can then be used to complete analogiesinvolving the signed distance between novel input pairs (1:3:: 5:? (7)), and extrapolate outside of the network’s trainingdomain. As a proof of principle, we extend the same architec-ture to more richly structured tree representations. We suggestthat this combination of predictive, error-driven learning andsimple inductive biases offers promise for deriving and utiliz-ing the representations necessary for high-level cognitive func-tions, such as analogy.