Search

Our interaction with others largely hinges on how we semantically organize the social world. The organization of such conceptual information is not static—as we age, our experiences and ever-changing anatomy alter how we represent and arrange semantic information. How does semantic distance between concepts affect this organization, particularly for those with pathological deficits in semantic knowledge? Using triplet judgment responses collected from healthy participants, we compute an ordinal similarity embedding for a set of social words and images that vary in the dimensions of age and gender. We compare semantic distances between items in the space to patterns of error in a word-picture matching task performed by patients with semantic dementia (SD). Error patterns reveal that SD patients retain gender information more robustly than age information, and that age-related errors are a function of linear distance in age from a concept word. The distances between probed and exemplar items in the resulting conceptual map reflect error patterns in SD patient responses such that items semantically closer to a probed concept—in gender category or in linear age—are more likely to be erroneously chosen by patients in a word-picture matching task. To our knowledge, this is the first triplet embedding work to embed representations of words and images in a unified space, and to use this space to explain patterns of behavior in patients with impaired social semantic cognition.

Recent work using neural networks and crowd-sourced perceptual judgements has shown that human figure drawings contain latent structure that can predict many characteristics of the artist including parent-reported motor function and perceived gender. We extend these approaches to two-dimensional renderings of three-dimensional cubes, assessing whether latent structure in these cube drawings likewise predicts demographic characteristics and motor function measured via a paper-folding task. Drawings produced with marker and paper showed a large predictive relationship with paper-folding (accounting for 59% of the offset variance, 62% of the angle variance, ps < .01). We also observed a complex interaction with gender: better cube-drawings predicted better paper-folding for male-identifying participants, but this relationship was reversed for female-identifying participants, who demonstrated better paper folding abilities overall. The results suggest that cube drawings contain richer structure than previously recognized and can provide a useful nonverbal metric for characterizing aspects of cognitive and motor abilities.

In visual cognitive neuroscience, there are two main theories about the function of the ventral visual system. One suggests that it serves to classify objects (H1); the other suggests that it generates intermediate representations from which people can generate verbal descriptions, actions, and other kinds of information (H2). To adjudicate these, we trained two deep convolutional AlexNet models on 330,000 images belonging to 86 classes, representing the intersection of Ecoset images and the semantic norms collected by the Leuven group. One model was trained to produce category labels (H1) , the other to generate all of an item's semantic features (H2). The two models learned very different representational geometries throughout the network. The representations acquired by the feature-generating model aligned better with human-perceived similarities amongst images, and better predicted human judgments in a triadic comparison task. The results thus support H2.

Human groups are able to converge to more accurate beliefs through deliberation, even in the presence of polarization and partisan bias --- a phenomenon known as the ``wisdom of partisan crowds.'' Large Language Models (LLMs) are increasingly being used to simulate human collective behavior, yet few benchmarks exist for evaluating their dynamics against the behavior of human groups. In this paper, we examine the extent to which the wisdom of partisan crowds emerges in groups of LLM-based agents that are prompted to role-play as partisan personas (e.g., Democrat or Republican). We find that they not only display human-like partisan biases, but also converge to more accurate beliefs through deliberation, as humans do. We then identify several factors that interfere with convergence, including the use of chain-of-thought prompting and lack of details in personas. Conversely, fine-tuning on human data appears to enhance convergence. These findings show the potential and limitations of LLM-based agents as a model of human collective intelligence.

Accurately simulating human opinion dynamics is crucial for understanding a variety of societal phenomena, including polarization and the spread of misinformation. However, the agent-based models (ABMs) commonly used for such simulations often over-simplify human behavior. We propose a new approach to simulating opinion dynamics based on populations of Large Language Models (LLMs). Our findings reveal a strong inherent bias in LLM agents towards producing accurate information, leading simulated agents to consensus in line with scientific reality. This bias limits their utility for understanding resistance to consensus views on issues like climate change. After inducing confirmation bias through prompt engineering, however, we observed opinion fragmentation in line with existing agent-based modeling and opinion dynamics research. These insights highlight the promise and limitations of LLM agents in this domain and suggest a path forward: refining LLMs with real-world discourse to better simulate the evolution of human beliefs.

Neural network models of language have long been used asa tool for developing hypotheses about conceptual representationin the mind and brain. For many years, such use involvedextracting vector-space representations of words andusing distances among these to predict or understand humanbehavior in various semantic tasks. In contemporary languageAIs, however, it is possible to interrogate the latent structureof conceptual representations using methods nearly identicalto those commonly used with human participants. The currentwork uses two common techniques borrowed from cognitivepsychology to estimate and compare lexical-semantic structurein both humans and a well-known AI, the DaVinci variant ofGPT-3. In humans, we show that conceptual structure is robustto differences in culture, language, and method of estimation.Structures estimated from AI behavior, while individuallyfairly consistent with those estimated from human behavior,depend much more upon the particular task used to generatebehavior responses–responses generated by the very samemodel in the two tasks yield estimates of conceptual structurethat cohere less with one another than do human structure estimates.The results suggest one important way that knowledgeinhering in contemporary AIs can differ from human cognition.

How are our conceptual representations organized, and how does this semantic structure differ in individuals or groups? We included 36 faces from the Chicago Face Database (Ma et al., 2015) varying in race, gender, and facial expression in a triplet judgement task, where participants were showed a target face and asked which of two other faces is most similar to the target face expression. Each of our 71 participants completed 1,012 triplet judgements. Using their judgements, we fit individualized embeddings capturing distances between faces in semantic space, resulting in 71 embeddings. We calculated the Procrustes similarity between all pairs of embeddings and clustered similarities, revealing four clusters of respondents. Each cluster organized faces differently, indicating group-wide differences in semantic organization of expression. At the individual level, we find that each participant’s responses for held-out items are better predicted by their own embedding than by other participants’ embeddings, p < .00001.

Cognitive models of learning often propose that belief formation is rational: new experience is combined with prior belief to support statistically optimal inferences. Why then are false beliefs so ubiquitous and difficult to correct? Prior work using a novel category-learning paradigm suggests that false beliefs may arise from the way that learners, when presented with contradictory information, choose to weight different information sources. When at least one source provides evidence highly consistent with the learner’s current belief, other sources are strongly discounted and little belief change is observed; but when evidence from the most agreeable source is more discrepant with the learner, other sources receive proportionally more weight and effect larger belief change. The current work assesses whether a similar model explains patterns of belief change, not in a novel lab-based task, but for propositional real-world beliefs held by participants prior to coming into the lab. In a paradigm analogous to the previous category-learning study, participants known to hold a polarized opinion on a controversial topic read arguments varying in agreement and/or opposition to that opinion. The discrepancy between the participant’s existing belief and that expressed by the arguments was experimentally varied. The pattern of belief change was remarkably similar to that observed in the novel category-learning task, suggesting that a similar underlying model can explain belief elasticity even for existing real-world beliefs.

Drawing is a powerful cognitive technology for creating external representations of thought. While drawings have long provided inspiration to researchers in many areas of cognitive science, including psychology, machine learning, and neuroscience, these communities have not generally had opportunities to interact and share insights. The goal of this workshop is to bring together perspectives from multiple disciplines to explore the question of how humans use drawings to communicate knowledge, catalyzing new opportunities for multidisciplinary collaboration. We are introducing a novel ``flipped'' format wherein we will hold three virtual seminars in the weeks leading up to CogSci 2022, each highlighting insights from Machine Learning, Neuroscience, and Developmental Science, respectively.Holding these thematic seminars in advance will enable us to attract a broader audience for our event and focus on promoting informal interaction among workshop attendees at the on-site event.

The hippocampus replays experiences during quiet rest periods, and this replay benefits subsequent memory. A critical open question is how memories are prioritized for this replay. We used functional magnetic resonance imaging (fMRI) pattern analysis to track item-level replay in the hippocampus during an awake rest period after participants studied 15 objects and completed a memory test. Objects that were remembered less well were replayed more during the subsequent rest period, suggesting a prioritization process in which weaker memories-memories most vulnerable to forgetting-are selected for replay. In a second session 12 hours later, more replay of an object during a rest period predicted better subsequent memory for that object. Replay predicted memory improvement across sessions only for participants who slept during that interval. Our results provide evidence that replay in the human hippocampus prioritizes weakly learned information, predicts subsequent memory performance, and relates to memory improvement across a delay with sleep.

In STEM domains, students are expected to acquire domainknowledge from visual representations that they may not yetbe able to interpret. Such learning requires perceptual flu-ency, or the ability to intuitively and rapidly see the underlyingconcepts in visuals and to translate between them. Perceptualfluency is acquired via nonverbal, implicit learning processes.Thus far, we have lacked a principled approach for identify-ing a sequence of perceptual fluency problems that promoterobust learning. Here, we describe how a novel machine learn-ing technique can generate an optimal sequence of perceptualfluency problems. In a human experiment, we show that amachine-generated sequence outperforms both a random se-quence and a sequence generated by a human domain expert.Interestingly, the machine-generated sequence resulted in sig-nificantly lower accuracy during training, but higher posttestaccuracy. This suggests that the machine-generated sequenceinduced desirable difficulties. To our knowledge, our study isthe first to show that machine learning can yield desirable dif-ficulties for perceptual learning

Semantic memory encompasses knowledge about both the properties that typify concepts (e.g. robins, like all birds, have wings) as well as the properties that individuate conceptually related items (e.g. robins, in particular, have red breasts). We investigate the impact of sleep on new semantic learning using a property inference task in which both kinds of information are initially acquired equally well. Participants learned about three categories of novel objects possessing some properties that were shared among category exemplars and others that were unique to an exemplar, with exposure frequency varying across categories. In Experiment 1, memory for shared properties improved and memory for unique properties was preserved across a night of sleep, while memory for both feature types declined over a day awake. In Experiment 2, memory for shared properties improved across a nap, but only for the lower-frequency category, suggesting a prioritization of weakly learned information early in a sleep period. The increase was significantly correlated with amount of REM, but was also observed in participants who did not enter REM, suggesting involvement of both REM and NREM sleep. The results provide the first evidence that sleep improves memory for the shared structure of object categories, while simultaneously preserving object-unique information.

In a categorization task involving both labeled and unlabeleddata, it has been shown that humans make use of the underlyingdistribution of the unlabeled examples. It has also been shownthat humans are sensitive to shifts in this distribution, and willchange predicted classifications based on these shifts. It is notimmediately obvious what causes these shifts – what specificproperties of these distributions humans are sensitive to. Assuminga parametric model of human categorization learning,we can ask which parameters or sets of parameters humans fixafter exposure to labeled data and which are adjustable to fitsubsequent unlabeled data. We formulate models to describedifferent parameter sets which humans may be sensitive to anda dataset which optimally discriminates among these models.Experimental results indicate that humans are sensitive to allparameters, with the closest model fit being an unconstrainedversion of semi-supervised learning using expectation maximization.

Previous studies on numeric cognition have focused primarilyon magnitude, based on its role as a core feature of numberknowledge. In this paper, we report the results of threeexperiments investigating adults’ sensitivity to properties ofnumber apart from magnitude. In Experiment 1, we use atriadic judgment task to replicate a classic study of numberproperties. In Experiment 2, we compare these representationsamong expert and non-expert groups. In Experiment 3, weexamine whether instruction can tune representation ofnumber properties. Results indicate that the triadiccomparison task is a reliable method of assessing sensitivityto number properties. We found that magnitude is difficult tosuppress among non-experts, who are primarily attuned tomagnitude and parity. Mathematically sophisticatedparticipants were sensitive to a range of number propertiescompared with the non-expert group. We discuss implicationsfor theories of number concepts and their relation to specialpopulations