Deep neural network representations of entities can serve as inputs to computational models of human mental representations to predict people's behavioral and physiological responses to those entities. Though increasingly successful in their predictive capabilities, the implicit notion of "human" that they rely upon often glosses over individual-level differences in beliefs, attitudes, and associations, as well as group-level cultural constructs. In this paper, we model shared representations of food healthiness by aligning learned word representations with the consensus among a group of respondents. To do so, we extend Cultural Consensus Theory to include latent constructs structured as fine-tuned word representations. We then apply the model to a dataset of people's judgments of food healthiness. We show that our method creates a robust mapping between learned word representations and culturally constructed representations that guide consumer behavior.

Online communities rely on their members to understand and follow community norms, which they learn by observing others and the consequences of their behavior, seeing codes of conduct, and receiving feedback via moderation. Here, to determine the contribution of each source of learning to the preservation of a social norm, we extend cultural consensus theory, a mathematical framework for identifying the cultural consensus in a community. In particular, we extend the model to include learning from experience, centralized moderation, and decentralized moderation, three features commonly found in online communities. We then apply the extended model to data from an online community dedicated to preserving a norm related to the psychophysical scaling of intersubjective notions of beauty derived from facial aesthetics. We find that users' perceptual alignment with the norm before enculturation predicts involvement in the community and that experience in the community is an important indicator for group perceptual learning.

Collectible items, such as stamps, coins, paintings, and trading cards, are often valued for their rarity. A side effect of rarer items being more highly valued is that they are also more often traded, discussed, and displayed. A new collector's experience of the category defined by a set collectible items is thus heavily biased towards the rare items. Theories of category learning predict that these conditions make for a uniquely challenging environment in which to learn a category because rarity-based sampling can invert the distribution of associated attribute frequencies. Here, we show that under these conditions, the demand for rarity is self-defeating: when newcomers do not correct for the sampling bias present in their experience, they will have a distorted sense of the category and misunderstand which items are in fact rare, causing rarity to become devalued over time. We find evidence for this dynamic in the context of The Bored Ape Yacht Club (BAYC), a collection of 10,000 non-fungible tokens (NFTs), each with a set of attributes that vary in rarity. We demonstrate that, in line with our theory, over time the influx of newcomers learning about BAYC has been associated with a decrease in the demand for tokens with rare attributes.

High-dimensional deep neural network representations of images and concepts can be aligned to predict human annotations of diverse stimuli. However, such alignment requires the costly collection of behavioral responses, such that, in practice, the deep-feature spaces are only ever sparsely sampled. Here, we propose an active learning approach to adaptively sample experimental stimuli to efficiently learn a Bayesian matrix factorization model with deep side information. We observe a significant efficiency gain over a passive baseline. Furthermore, with a sequential batched sampling strategy, the algorithm is applicable not only to small datasets collected from traditional laboratory experiments but also to settings where large-scale crowdsourced data collection is needed to accurately align the high-dimensional deep feature representations derived from pre-trained networks. This provides cost-effective solutions for collecting behavioral data and generating high-quality predictions in large-scale behavioral and cognitive studies.

Working memory is a limited-capacity form of human mem-ory that actively holds information in mind. Which memoriesought to be maintained? We approach this question by showingan equivalence between active maintenance in working mem-ory and a Markov decision process in which, at each moment,a cognitive control mechanism selects a memory as the targetof maintenance. The challenge of remembering is then findinga maintenance policy well-suited to the task at hand. We com-pute the optimal policy under various conditions and defineplausible cognitive mechanisms that can approximate these op-timal policies. Framing the problem of maintenance in thisway makes it possible to capture in a single model many of theessential behavioral phenomena of memory maintenance, in-cluding directed forgetting and self-directed remembering. Fi-nally, we consider the case of imperfect metamemory — wherethe current state of memory is only partially observable — andshow that the fidelity of metamemory determines the effective-ness of maintenance.

Generative models of human identity and appearance havebroad applicability to behavioral science and technology, butthe exquisite sensitivity of human face perception means thattheir utility hinges on the alignment of the model’s representa-tion to human psychological representations and the photoreal-ism of the generated images. Meeting these requirements is anexacting task, and existing models of human identity and ap-pearance are often unworkably abstract, artificial, uncanny, orbiased. Here, we use a variational autoencoder with an autore-gressive decoder to learn a face space from a uniquely diversedataset of portraits that control much of the variation irrele-vant to human identity and appearance. Our method generatesphotorealistic portraits of fictive identities with a smooth, navi-gable latent space. We validate our model’s alignment with hu-man sensitivities by introducing a psychophysical Turing testfor images, which humans mostly fail. Lastly, we demonstratean initial application of our model to the problem of fast searchin mental space to obtain detailed “police sketches” in a smallnumber of trials.

The proliferation of web-connected devices has presentedsignificant opportunities and challenges to cognitive sci-ence — opportunities in that cognitive scientists can collectdata relevant to human cognition orders of magnitude fasterthan before, addressing questions that were otherwise impos-sible to address; and challenges in that cognitive scientistsrequire new infrastructure to collect these data and new meth-ods to analyze them once collected.This workshop brings together cognitive scientists who areat the forefront of these opportunities and challenges in scal-ing cognitive science, along with cognitive scientists whowould like to be, to engage in a day of interactive exchangeand development of ideas related to scaling cognitive science.The workshop is a full day. Each presentation addresses adifferent opportunity or challenge. One set of presentationshighlights opportunities. A second set highlights challengesof statistical analysis and data collection. In a break-out ses-sion, attendees address these opportunities and challengeshead on.

Much of our knowledge is transmitted socially rather thanthrough firsthand experience. Even our memories depend onrecollections of those around us. Surprisingly, when people re-call memories with others, they do not reach the potential num-ber of items they could have recalled alone. This phenomenonis called collaborative inhibition. Recently, Luhmann and Ra-jaram (2015) analyzed the dynamics of collaborative inhibitionat scale with an agent-based model, extrapolating from previ-ous small-scale laboratory experiments. We tested their modelagainst human data collected in a large-scale experiment andfound that participants demonstrate non-monotonicities not ev-ident in these predictions. We next analyzed memory transmis-sion beyond directly interacting agents by placing agents intonetworks. Contrary to model predictions, we observed highsimilarity only within directly interacting pairs. By comparingbehavior to model predictions in large-scale experiments, wereveal unexpected results that motivate future work in eluci-dating the algorithms underlying collaborative memory.

Understanding how people represent categories is a core prob-lem in cognitive science. Decades of research have yieldeda variety of formal theories of categories, but validating themwith naturalistic stimuli is difficult. The challenge is that hu-man category representations cannot be directly observed andrunning informative experiments with naturalistic stimuli suchas images requires a workable representation of these stimuli.Deep neural networks have recently been successful in solvinga range of computer vision tasks and provide a way to com-pactly represent image features. Here, we introduce a methodto estimate the structure of human categories that combinesideas from cognitive science and machine learning, blendinghuman-based algorithms with state-of-the-art deep image gen-erators. We provide qualitative and quantitative results as aproof-of-concept for the method’s feasibility. Samples drawnfrom human distributions rival those from state-of-the-art gen-erative models in quality and outperform alternative methodsfor estimating the structure of human categories.

The quiet hum of interpersonal coordination that runs through-out social communication and interaction shows how individu-als can subtly influence one another’s behaviors, thoughts, andemotions over time. While the majority of research on co-ordination studies face-to-face interaction, recent advances incrowdsourcing afford the opportunity to conduct large-scale,real-time social interaction experiments. We take advantageof these tools to explore interpersonal coordination in a “min-imally interactive context,” distilling the richness of naturalcommunication into a tightly controlled setting to explore howpeople become coupled in their perceptual and memory sys-tems while performing a task together. Consistent with previ-ous work on postural sway and gaze, we found that individualsbecome coupled to one another’s cognitive processes withoutneeding to be co-located or fully interactive with their partner;interestingly, although participants had no information abouttheir partner and no means of direct communication, we alsofound hints that social forces can shape minimally interactivecontexts, similar to effects observed in face-to-face interaction.