Search

Representation matters and contributes to student engagement, sense of belonging, and overall success. This is especially true for our most vulnerable student populations, which include transgender and non-binary students who do not often see themselves represented within the structure of the academic institution. While university-wide initiatives to engage students such as first year seminars are important, so too are smaller scale efforts such as creating representation through the use of teaching materials and exercises in courses and information literacy sessions. When discussing representation for the LGBTQ + community, and especially transgender and non-binary parts of that community, it is important to acknowledge that within this larger group, there are any number of intersectional identities impacting individuals' sense of belonging, support, and access to success, all of which cannot be addressed in the space of this column or in the time it takes to run an information literacy session. However, that does not mean we, as educators, should not try. [ABSTRACT FROM AUTHOR]

Aphantasia, namely the inability to voluntarily form visual mental imagery, does not, counterintuitively, impair the affected from successfully performing mental imagery tasks. One way of explaining this finding is to posit that aphantasics, despite their claim to the contrary, can form visual imagery, a position here referred to as aphantasia skepticism. This article outlines and rejects two types of aphantasia skepticism and argues that the position results from what is coined the visualizer's fallacy, namely the false belief that visual mental imagery is necessary to carry out mental imagery tasks. Furthermore, it is argued that the visualizer's fallacy and the resulting aphantasia skepticism are not only potentially harmful to aphantasics but may also lead to an impoverished view of the dynamics of cognition in general.

Spontaneous recovery of fear after extinction is a well-established behavioral phenomenon. Different theories in psychology account for spontaneous recovery by proposing that it may result from temporal weighting, reduced processing of stimuli over time, enhanced salience of adverse events or return of the acquisition context. We propose a novel mechanism of spontaneous recovery: selective maintenance of adverse events, and ground this mechanism in a computational model of latent cause inference. To investigate the proposed mechanism, we collected behavioral data with an aversive conditioning and extinction task (N=280) and fit the data with computational models formalizing our and others' theories. Quantitative and qualitative model comparisons indicated that selective maintenance of adverse events accounts for spontaneous recovery better than alternative theories. As spontaneous recovery of fear after extinction can serve as a model of relapse after exposure therapy, we use this mechanistic understanding of spontaneous recovery to propose and simulate the effect of add-on interventions to prevent relapse after exposure therapy.

Counterfactual reasoning is at the centre of human daily life and plays a key role in shaping our moral and social judgments. Its effect on moral judgment in adulthood, such as justifying immoral behavior (e.g., “If you had not left your phone on the table, it would not have been stolen.”), has been studied for years. However, we still know very little about when counterfactual reasoning starts to affect humans' moral judgments. To test this, we examined the effect of better and worse counterfactual alternatives on 5-6-year-olds' (N = 91) moral and happiness judgments. We found that children judged social exclusion (e.g., a new kid has to play alone while other children play together) as less morally acceptable after imagining how it could have been better (e.g., the new kid and other children at the school could have played all together), but, contrary to past work with adults, they did not justify it after imagining how it could have been even worse (e.g., the other children could have broken the new kid's toy). However, children's happiness judgments showed the opposite effect: they reported feeling happier about reality after imagining a worse counterfactual alternative compared to children who only thought about what actually happened.Keywords: counterfactuals; moral judgment; children; happiness judgment

Spatial cognition is fundamental to human behavior, but people differ in how they remember spatial relations, variably using body-based (egocentric) and environment-based (allocentric) spatial reference frames. Despite decades of study, the causes of this variation and flexibility in spatial memory remain unclear. Here we show that people spontaneously use different reference frames on different spatial axes at the same time. When remembering the placement of a target object in a 2-dimensional array, Indigenous Tsimane' adults preferentially used allocentric space to determine lateral placement and egocentric space to determine sagittal placement in the same action. This effect of axis was also significant among US university students, whose overall preference for egocentric space was stronger on the sagittal than lateral axis. These findings support a novel account of spatial cognitive diversity and suggest that people across cultures habitually integrate egocentric and allocentric spatial reference frames into the same action.

An important aspect of language comprehension is learning and generalizing complex lexical relations. For instance, having learned that the phrase preserve cucumbers predicts vinegar and that preserve berries predicts dehydrator, one should be able to infer that the novel phrase preserve peppers is more compatible with vinegar, because pepper is more similar to cucumber. We studied the ability to perform such (compositional) generalization in distributional models trained on an artificial corpus with strict semantic regularities. We found that word-encoding models failed to learn the multi-way lexical dependencies. Recurrent neural networks learned those dependencies but struggled to generalize to novel combinations. Only mini GPT-2, a minified version of the Transformer GPT-2, succeeded in both learning and generalization. Because successful generalization in our tasks requires capturing the relationship between a phrase and a word, we argue that mini GPT-2 acquired hierarchical representations that approximate phrase structure. Our results show that, compared to older models, Transformers are architecturally advantaged to perform compositional generalization.

Decoding naturalistic stimuli from neural recordings is a significant challenge in systems neuroscience, primarily due to the high-dimensional and nonlinear nature of stimulus-response interactions, and is further exacerbated by the limited availability and noisiness of neural data. While contemporary approaches that incorporate generative models, such as Generative Adversarial Networks (GANs), attempt to address these issues by mapping neural responses to latent representations, they do not fully overcome these obstacles. In this work, we present a novel paradigm of differential neural decoding (dicoding) that focuses on the relative changes in response patterns, which not only expands the neural training data quadratically but also inherently denoises it. To determine the corresponding stimulus changes, this method leverages the Euclidean and feature-disentangled properties of the underlying latents through vector arithmetic. As such, we not only effectively exploit the latent space but also achieve semantically meaningful latent offsets in the context of the stimuli, resulting in improved sample efficiency. We trained a decoder to predict changes in latent vectors based on the corresponding changes in neural responses. The absolute latent vector itself was derived by vector addition of the predicted latent change (indicative of stimulus shift) to a reference latent, which was fed to the generator for the reconstruction of the perceived stimulus. Our results show that this geometrically principled approach facilitates more effective reconstruction of naturalistic stimuli from noisy and limited neural data.

Advancements in deep neural networks have led to significant progress in computer vision and natural language processing. These networks, trained on real-world stimuli, develop high-level feature representations of stimuli. It is hypothesized that these representations, stemming from different inputs, should converge into similar conceptual systems, as they reflect various perspectives of the same underlying reality. This paper examines the degree to which different conceptual systems can be aligned in an unsupervised manner, using feature-based representations from deep neural networks. Our investigation centers on the alignment between the image and word representations produced by diverse neural networks, emphasizing those trained via self-supervised learning methods. Subsequently, to probe comparable alignment patterns in human learning, we extend this examination to models trained on developmental headcam data from children. Our findings reveal a more pronounced alignment in models trained through self-supervised learning compared to supervised learning, effectively uncovering higher-level structural connections among categories. However, this alignment was notably absent in models trained with limited developmental headcam data, suggesting more data, more inductive biases, or more supervision are needed to establish alignment from realistic input.

While Vector Symbolic Architectures (VSAs) are promising for modelling spatial cognition, their application is currently limited to artificially generated images and simple spatial queries. We propose VSA4VQA – a novel 4D implementation of VSAs that implements a mental representation of natural images for the challenging task of Visual Question Answering (VQA). VSA4VQA is the first model to scale a VSA to complex spatial queries. Our method is based on the Semantic Pointer Architecture (SPA) to encode objects in a hyperdimensional vector space. To encode natural images, we extend the SPA to include dimensions for object's width and height in addition to their spatial location. To perform spatial queries we further introduce learned spatial query masks and integrate a pre-trained vision-language model for answering attribute-related questions. We evaluate our method on the GQA benchmark dataset and show that it can effectively encode natural images, achieving competitive performance to state-of-the-art deep learning methods for zero-shot VQA.

This study explores human capabilities in distinguishing and recognizing entities that change over time from those that do not. We specifically investigate the linguistic distinction between "individual-level predicates" (ILPs) and "stage-level predicates" (SLPs). Our empirical approach focuses on how humans visually distinguish these two types. We performed a corpus analysis, in which a set of image captions were randomly extracted and annotated by experts with either SLP or ILP labels. The findings indicated a predominance of SLPs over ILPs in the image captions. We then performed automatic annotation on a large dataset of image captions and conducted a machine-learning experiment on image classification based on ILSs and SLPs. Our results demonstrated that SLPs were identified with high accuracy, while ILPs were identified with about chance level, substantially lower than human capabilities. Given the analyses, we discuss what features of the image contribute to distinguishing between ILPs and SLPs.

In term series problems where multiple mental models can be constructed, partial-order models can be created as mental representations, which make it easier to perceive the symmetry of the terms. To test these hypotheses, we categorized multi-model (indeterminate) term series problems according to the patterns of partial-order models that could be constructed, and analyzed the reasoning performance for each pattern. These results suggest that reasoners tend to use the symmetry of terms to reduce the cognitive load of reasoning. Analysis of the patterns of incorrect answers also suggests that attempts to exploit the symmetry of the term may be biased, leading to errors in reasoning.

We test whether the support one holds about an event is influenced by other hypotheses. We addressed this by examining context effects in subjective probabilities (SPs) when forecasting NCAA men's basketball team rankings. A challenge in investigating context effects with naturalistic stimuli is the need to model the different representations of the options. To do so, we adapted the Spatial Arrangement method to capture individual representations and developed an algorithm to select stimuli. We asked participants steeped in basketball knowledge to create spatial maps for 50 teams. They were then presented with customized triplets of teams and asked to estimate their SP that one team would outrank the others. The study uncovered context effects in SPs, and moderators of the effects. Our findings suggest that similar cognitive processes may govern the construction of belief and preference and highlight the importance of modeling mental representations to understand forecasting scenarios.

Construal Level Theory (CLT) suggests that we represent objects close to us in a concrete and modal fashion, and that representations become more abstract and amodal with increasing distance from ourselves. Evidence for such an association of abstraction level and distance comes from the Implicit Association Test (IAT), where participants are faster when pressing one key for “near” and “concrete” and another key for “far” and “abstract” targets (congruent), than when “near” is paired with “abstract” and “far” with “concrete” (incongruent). However, previous experiments might have confounded distance and abstraction by employing inherently near and far targets (e.g., CHAIR vs. SUN) that might also differ in their abstractness. Here, we thus experimentally induced different distances in a learning phase before a subsequent IAT task. Even with this controlled distance manipulation, a pronounced congruency effect emerged, providing further support for an association of distance and abstraction level as suggested by CLT.

This study investigates whether homophone competitors are activated during typewriting and to which extent such activation is modulated by syntactic category. In two experiments, we compared the typewriting of homophone pairs in high vs. low conflict sentences (i.e., both homophones vs. only one homophone in the sentence, respectively) in a sentence dictation task (Experiment 1) and in a question-answering task (Experiment 2). The homophone pairs either belonged to the same or different syntactic categories. In Experiment 1, we found a homophone interference effect in accuracy, independent of conflict and syntactic category. In Experiment 2, this effect was replicated, but in addition, participants were slower to type homophones in a high vs. a low conflict context. Our results show a robust, lexically-situated homophone interference effect, regardless of conflict and syntactic category, but when deeper processing of the sentence is involved, conflict starts to play a role.

How do humans decide what to look at and when to stop looking? The Rational Action, Noisy Choice for Habituation (RANCH) model formulates looking behaviors as a rational information acquisition process. RANCH instantiates a hypothesis about the perceptual encoding process using a neural network-derived embedding space, which allows it to operate on raw images. In this paper, we show that the model not only captures key looking time patterns such as habituation and dishabituation, but also makes fine-grained, out-of-sample predictions about magnitudes of dishabituation to previously unseen stimuli. We validated those predictions experimentally with a self-paced looking time task in adults (N = 468). We also show that model fits are robust across parameters, but that assumptions about the perceptual encoding process, the learning process and the decision process are all critical for predicting human performance.

Large scale neural models show impressive performance across a wide array of linguistic tasks. Despite this they remain, largely, black-boxes - learning vector-representations of their input that prove difficult to interpret. This limits our ability to understand what they learn, and when the learn it, or characterise why they often fail to generalise systematically. To address this we introduce a novel approach to interpretability that looks at the mapping a model learns from sentences to representations as a kind of language in its own right. In doing so we introduce a set of information-theoretic measures that quantify how structured a model's representations are with respect to its input, and when during training that structure arises. Our measures are fast to compute, grounded in linguistic theory, and can predict which models will generalise best based on their representations. We use these measures to describe two distinct phases of training a transformer: an initial phase of in-distribution learning which reduces task loss, then a second stage where representations becoming robust to noise. Generalisation performance begins to increase during this second phase, drawing a link between generalisation and robustness to noise. Finally we look at how model size affects the structure of the representational space, showing that larger models ultimately compress their representations more than their smaller counterparts.

Generalization to new tasks requires learning of task representations that accurately reflect the similarity structure of the task space. Here, we argue that episodic memory (EM) plays an essential role in this process by stabilizing task representations, thereby supporting the accumulation of structured knowledge. We demonstrate this using a neural network model that infers task representations that minimize the current task's objective function; crucially, the model can retrieve previously encoded task representations from EM and use these to initialize the task inference process. With EM, the model succeeds in learning the underlying task structure; without EM, task representations drift and the network fails to learn the structure. We further show that EM errors can support structure learning by promoting the activation of similar task representations in tasks with similar sensory inputs. Overall, this model provides a novel account of how EM supports the acquisition of structured task representations.

We propose that kinds relate to particular things either constitutively or contingently. Taxonomic categories of animals and artifacts constitutively relate their members: DOG and CAR group things by aspects of the forms of their matter; the forms that make them things instead of stuff. Categories of things in roles or with diseases contingently relate to their members: LAWYER and DIABETIC group things by forms other than the forms that make them things. We confirm this distinction in five experiments with American adults.

Speech perception involves storing and integrating sequentially presented items. Recent work in cognitive neuroscience has identified temporal and contextual characteristics in humans' neural encoding of speech that may facilitate this temporal processing. In this study, we simulated similar analyses with representations extracted from a computational model that was trained on unlabelled speech with the learning objective of predicting upcoming acoustics. Our simulations revealed temporal dynamics similar to those in brain signals, implying that these properties can arise without linguistic knowledge. Another property shared between brains and the model is that the encoding patterns of phonemes support some degree of cross-context generalization. However, we found evidence that the effectiveness of these generalizations depends on the specific contexts, which suggests that this analysis alone is insufficient to support the presence of context-invariant encoding.

One feature of mental time travel is the ability to recursively embed temporal perspectives across different times: humans can remember how we anticipated the future and anticipate how we will remember the past. This recursive structure might be formalised in terms of a “grammar” that is reflective of but more general than linguistic notions of absolute and relative tense. Here I provide a foundation for this grammatical framework, emphasising a bounded (rather than unbounded) role of recursion in supporting mental time travel to a limited temporal depth and to actual and possible scenarios. Anticipated counterfactual thinking, for instance, entails three levels of mental time travel to a possible scenario (“in the future I will reflect on how my past self could have taken a different future action”) and is implicated in complex human decision-making. This perspective calls for further research into the nature and origins of recursive mental time travel.

Across three experiments, we show that 15-month-old infants understand that arbitrary objects can be used as symbols. Experiment 1 shows that infants map geometric shapes (e.g., a triangle) onto familiar discourse referents (e.g., a duck) based on labeling (e.g., “Look, a duck!”). Experiment 2 shows that infants do not generalize these mappings to a new speaker. This rules out the alternative hypothesis that infants interpret the labeling events literally. Experiment 3 shows that infants are sensitive to the conceptual identity of the discourse referent. After being told that one shape represents an agent (e.g., a duck) and another shape represents a patient (e.g., a cup), infants attend differentially when the agent symbol moves towards the patient symbol than the opposite. This rules out the alternative hypothesis that infants interpret the labeling events as referential pacts. The findings jointly indicate that symbolic relations are easily activated and available early in human development.

When do children understand that number words (such as ‘five') refer to exact quantities and that the same number word can be used to label two sets whose items correspond 1-to-1 (e.g., if each bunny has a hat, and there are five hats, then there are five bunnies)? Two studies with English-speaking 2- to 5-year-olds revealed that children who could accurately count large sets (CP knowers) were able to infer that sets exhibiting 1-to-1 correspondence share the same number word, but not children who could not accurately count large sets (subset knowers). However, not all CP knowers made this inference, suggesting that learning to construct and label large sets is a critical but insufficient step in discovering that numbers represent exact quantities. CP knowers also failed to identify 1-to-1 corresponding sets when faced with sets that had an off-by-one difference, suggesting that children who could accurately count large sets used approximate magnitude to establish set equality, rather than 1-to-1 correspondence. These results suggest that children's initial intuitions about numerical and set equality are based on approximation, not 1-to-1 correspondence, and that this occurs well after they have learned to count and construct large sets.

Humans learn and generate languages, music, games, and seemingly limitless varieties of other structures across domains. Unlike many AI systems, we often do so from little data. How do we learn such large varieties of richly structured representations so efficiently? One possibility is that people ``learn by programming,'' synthesizing data-generating algorithms to explain what we observe. We examine the nature and origins of this learning mechanism in adults, children, and nonhuman primates (macaque monkeys), using a highly unconstrained sequence prediction task. Although adults and children quickly learn many richly structured sequences, monkeys learn only the simplest sequences (e.g. lines). We test multiple learning models, finding that adults are best explained by a ``Language of Thought''-like program-learning model and monkeys by a simpler extrapolation strategy. Children exhibit varied learning strategies but are best fit in aggregate by an intermediately expressive model. Paper available at https://sites.google.com/view/patternlearning.

Languages tend to describe who is doing what to whom by placing subjects before objects. This bias for agents is reflected in event cognition: agents capture more attention than patients in human adults and infants. We investigated whether this agent preference is unique to humans. We presented Guinea baboons (Papio papio, N = 13) with a change detection paradigm with chasing animations. The baboons had to respond to a colour change which was applied to either the chaser/agent or the chasee/patient. They were faster to detect a change to the chaser than to the chasee, which cannot be explained by low-level features in our stimuli. Our study suggests that baboons show an agent preference similar to human infants and adults. This may be an evolutionarily old mechanism that is shared between humans and other primates, which could have become externalised in language as a tendency to place the subject first.

One of the core properties of human language is compositionality: the meaning of a sentence can be understood by the meaning of individual words and the rules for combining them (Szabó, 2020). We investigate the development of conceptual compositionality (the combination of concepts). In our study, 6- to 9-year-old children (N = 40) were shown a card with two objects (e.g., a car and a star). Participants were introduced to two characters (a robot and a wizard) that used their powers to change the objects in different ways (e.g., turning one object pink). In the test trials, participants were asked to predict what a card would look like after both characters used their powers on the same card. All participants successfully learned the characters' powers, but only participants 7.5 years and older succeeded in the compositionality test trials. Our findings suggest that by age 7.5 children can successfully compose functions.

How do people perform general-purpose physical reasoning across a variety of scenarios in everyday life? Across two studies with seven different physical scenarios, we asked participants to predict whether or where two objects will make contact. People achieved high accuracy and were highly consistent with each other in their predictions. We hypothesize that this robust generalization is a consequence of mental simulations of noisy physics. We designed an "intuitive physics engine'' model to capture this generalizable simulation. We find that this model generalized in human-like ways to unseen stimuli and to a different query of predictions. We evaluated several state-of-the-art deep learning and scene feature models on the same task and found that they could not explain human predictions as well. This study provides evidence that human's robust generalization in physics predictions are supported by a probabilistic simulation model, and suggests the need for structure in learned dynamics models.

Research suggests that the meanings of iconic signs are not easily guessable by sign-naive people; however, some signs' meanings are more easily guessed than others'. What causes some signs to be more easily guessable (more transparent) than others is not well-understood. In our previous research, we showed that signs whose form is based on more cross-linguistically common underlying motivations were chosen as "better suited" to a meaning—that is, they are more transparent—than signs based on less common underlying motivations (Tkachman, Sadlier-Brown, Lo, & Hudson Kam, 2023). In the current study, we ask whether, in addition, iconicity affects a sign's transparency. We asked sign-naive English speakers to rate all the signs from our previous study for how iconic they are. We then reanalyzed the data from our previous study in light of the obtained iconicity ratings. Results show that when people are asked to choose between an attested sign for a given animal label and an unattested one (i.e., a sign for a different animal), iconicity ratings did not affect participants' preferences: attested signs are preferred regardless of how iconic they are. However, when participants are asked to choose between two attested signs with the same meaning (e.g., two signs for 'cat' from different sign languages), iconicity does appear to affect participants' choices: participants were more likely to pick the more cross-linguistically common sign if the difference in iconicity ratings between the two signs was bigger. These results shed additional light on the ongoing debate on the connection between iconicity and transparency: iconicity by itself does not make a sign transparent, but it can enhance transparency under certain conditions.

Video game playing is an extremely structured domain where algorithmic decision-making can be tested without adverse real-world consequences. While prevailing methods rely on image inputs to avoid the problem of hand-crafting state space representations, this approach systematically diverges from the way humans actually learn to play games. In this paper, we design object-based input representations that generalize well across a number of video games. Using these representations, we evaluate an agent's ability to learn games similar to an infant - with limited world experience, employing simple inductive biases derived from intuitive representations of physics from the real world. Using such biases, we construct an object category representation to be used by a Q-learning algorithm and assess how well it learns to play multiple games based on observed object affordances. Our results suggest that a human-like object interaction setup capably learns to play several video games, and demonstrates superior generalizability, particularly for unfamiliar objects. Further exploring such methods will allow machines to learn in a human-centric way, thus incorporating more human-like learning benefits.

The multimodal deep neural networks, represented by CLIP, have generated rich downstream applications owing to their excellent performance, thus making understanding the decision-making process of CLIP an essential research topic. Due to the complex structure and the massive pre-training data, it is often regarded as a black-box model that is too difficult to understand and interpret. Concept-based models map the black-box visual representations extracted by deep neural networks onto a set of human-understandable concepts and use the concepts to make predictions, enhancing the transparency of the decision-making process. However, these methods involve the datasets labeled with fine-grained attributes by expert knowledge, which incur high costs and introduce excessive human prior knowledge and bias. In this paper, we observe the long-tail distribution of concepts, based on which we propose a two-stage Concept Selection Model (CSM) to mine core concepts without introducing any human priors. The concept greedy rough selection algorithm is applied to extract head concepts, and then the concept mask fine selection method performs the extraction of core concepts. Experiments show that our approach achieves comparable performance to end-to-end black-box models, and human evaluation demonstrates that the concepts discovered by our method are interpretable and comprehensible for humans.

The literature on conceptual combination has thus far been limited to research at the aggregate level investigating adjective-noun and noun-noun combinations. One well-established phenomenon within this literature is that of concept dominance (Hampton, 1988), which is the finding that the relative contribution of constituent concepts (for example sport or game) to their conjunction (sport that is also a game) is often very unequal. This exploratory study investigated individual differences in how people understand adjective-adjective-noun combinations, such as long blue coat. Participants rated images of coats varying along the perceptual dimensions of length and color for typicality in two different conjunctions, namely long blue coat and long purple coat. We used multidimensional scaling (MDS) to construct an aggregate coat space from similarity data collected with the Spatial Arrangement Method (SpAM). Using external unfolding, we modeled participants' typicality judgements by representing their individual typicality data as vectors within the aggregate MDS space, such that orthogonal projections from the coats onto the vectors represent their perceived typicality in the conjunctions. We did not find strong evidence for concept dominance at the aggregate level; however, we did find evidence for concept dominance at the individual level, with marked individual differences in the extent of dominance and which dimension was dominant. The validation of external unfolding for research into conceptual combination comes with new research possibilities, several of which are proposed.

Existing models of the basal ganglia assume the existence of separate channels of neuron populations for representing each available action. This type of localist mapping limits models to small, discrete action spaces, since additional actions require additional channels, costing neural resources and imposing new connective tracts. In contrast, evidence suggests that the basal ganglia plays a role in the selection of both discrete action units, and continuously-valued action kinematics. In this work, we model the basal ganglia with distributed action representations, using high-dimensional vectors. This method lends itself to representing both discrete and continuous action spaces. Vectors that represent actions are weighted by a scalar value (their salience to the current task), and bundled together to form a single input vector. This paper provides an overview of the encoding method and network structure, as well as a demonstration of the model solving an action selection task using spiking neurons.

Many aspects of human learning have been proposed as a process of constructing mental programs: from acquiring symbolic number representations to intuitive theories about the world. In parallel, there is a long-tradition of using information processing to model human cognition through Rate Distortion Theory (RDT). Yet, it is still poorly understood how to apply RDT when mental representations take the form of programs. In this work, we adapt RDT by proposing a three way trade-off among rate (description length), distortion (error), and computational costs (search budget). We use simulations on a melody task to study the implications of this trade-off, and show that constructing a shared program library across tasks provides global benefits. However, this comes at the cost of sensitivity to curricula, which is also characteristic of human learners. Finally, we use methods from partial information decomposition to generate training curricula that induce more effective libraries and better generalization.

While meaning variation in common words across language and culture is well established, only a few studies have explicitly quantified how general such differences are and whether differences reflect slight variations in meaning or could be considered to map onto entirely distinct concepts for different groups. The present study aims to investigate the extent to which common words can be interpreted differently between groups of English-proficient native Chinese speakers and native English speakers. This was done through a free judgment of associative strength (JAS) task using 42 cue English nouns. Our findings revealed language-specific meanings across all 42 cue words, with strong evidence for language-specific meaning in nearly 95\% of nouns. To determine whether these words map onto entirely distinct language-specific concepts, we measured conceptual diversity using Latent Profile Analysis (LPA). The results of the LPA showed that nearly 69% of the cue words could be mapped onto more than one concept across all participants. Importantly, language differences were related to conceptual diversity in nearly 64% of words featuring multiple concepts. In sum, we found robust evidence of word meanings and conceptual variations among individuals across distinct linguistic and cultural backgrounds, even for common English words.

The dominant mental timelines of native Chinese speakers (Exp1) and Mandarin learners of near-native proficiency (Exp2) was examined with the spontaneous gesture task. The results demonstrated that (1) both groups produced horizontal, vertical, sagittal, fused horizontal and vertical, and fused horizontal and sagittal gestures for all kinds of Chinese temporal words, indicating a strong preference for horizontal over vertical gestures. (2) Negligible correlations between immediate spatio-temporal metaphors and the mental timelines were observed, with an almost non-existent difference in gesture distribution across metaphorical types between the two groups. The findings indicate that (1) the horizontal mental timeline is the dominant timeline for two groups; (2) visuospatial experience exerts a greater influence on temporal cognition; (3) mental timelines formed by the long-term effects of language may operate beyond the immediate metaphors, similar to the horizontal gestures. A unified model proposing embodied experience as the mechanism for activating mental timelines is presented.

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.

This study investigates whether linguistic influences can affect the manifestation of lateral space-valence mappings in people's minds. Although most oral languages and cultures of the world have expressions and conventions that associate the good with the right space, this association seems to be body- specific: while right-handers associate positive concepts with the right side and negative concepts with the left side, left- handers have the oppositive association, and the size of the effect of the body specificity does not vary with linguistic and cultural conventions. Thus, it is widely believed that this conceptual metaphor only depends on the body. However, sign languages do not seem to have any conventional association between lateral space and valence, and a recent study has shown that signers do not associate valence with lateral space, opening the possibility of a causal influence of language. The present study set to replicate this surprising and controversial finding by comparing a sign language group, consisting of Spanish and Chinese Sign Language users, and an oral Spanish control group on the widely applied “Bob” task in this field. Supporting prior findings, Spanish language participants associated the “good” with their dominant side of space, closely matching the anticipated proportion, but signers did not. This pattern of results can be explained by a strong linguistic influence on the formation of lateral associations of emotional valence, but we discuss some alternative possibilities.

Although even infants appear to consider multiple possibilities, preschoolers often fail tasks that require reasoning about mutually exclusive alternatives. We review two explanations for this failure: (1) children have a minimal representation of possibility and fail to distinguish necessary from merely possible outcomes; and (2) children are sensitive to this distinction, but competing motivations (e.g., the tendency to explore) can lead to apparent failures. To test these hypotheses, we assessed 3- and 4-year-olds on a novel search task. Here, children searched for an object that was dropped from either a transparent (one necessary location) or opaque (two possible locations) set of inverted Y-shaped tubes. In Exp. 1, we found that children spent less time searching the first location when there were two possible candidates. Exp. 2 replicates these results in a digital task that does not require manual search.

In logical theories of meaning, threshold and prototype models are two distinctive formal approaches. In cognitive science literature, however, where the two models are operationalized, there is support for the use of a threshold model in categorization (Schmidt, Goodman, Barner, & Tenenbaum, 2009; Ramotowska, Haaf, Van Maanen, & Szymanik, 2022) as well as sup- port for the prototype model (Douven, 2016; Douven, Wenmackers, Jraissati, & Decock, 2017), and in many cases the two models are used interchangeably (Kruschke, 2008). We test for the case of relative gradable adjectives whether a) there is a difference between predicted degrees of membership from the two models when relying on explicit reports of threshold and prototype values, and b) which of the models better pre- dicts behavioral data from categorization tasks. Results suggest that prototype and threshold models are highly predictive of behaviour in a categorization task and that the two models yield similar results with a slight advantage of the threshold model.

As children enter elementary school, their understanding of the ordinal structure of numbers transitions from a memorized count list of the first 50-100 numbers to knowing the successor function and understanding the countably infinite. We investigate this developmental change in two neural network models that learn the successor function on the pairs (N, N+1) for N in (0, 98). The first uses a one-hot encoding of the input and output values and corresponds to children memorizing a count list, while the second model uses a place-value encoding and corresponds to children learning the language rules for naming numbers. The place-value model showed a predicted drop in representational similarity across tens boundaries. Analysis of the latent representation shows that counting across a tens boundary can be understood as a vector operation in 2D space, where the numbers with the same tens place are organized in a linearly separable manner, whereas those with the same ones place are grouped together. A curriculum learning simulation shows that, in the expanding numerical environment of the developing child, representations of smaller numbers continue to be sharpened even as larger numbers begin to be learned. These models set the stage for future work using recurrent architectures to move beyond learning the successor function to simulating the counting process more generally, and point towards a deeper understanding of what it means to understand the countably infinite.

Humans can flexibly bind familiar functional roles to novel entities in their environment. For example, children who have the concept of ``goal posts'' can bind this abstract role to two hats placed on the street. In doing so, they can port over existing expectations of ``goal posts" for the duration of the game. In this paper, we seek to explore artificial agents' ability to perform flexible role binding and rebinding. To this end, we designed a Gridworld navigation game and tested a popular CNN-based agent which has had success in other tasks involving visual and spatial state spaces (e.g. Atari or Minigrid). To our surprise, we found that while this architecture was capable of overfitting to the training set, it was not able to learn flexible role binding without intervention. We ultimately show that with carefully engineered data augmentation techniques, our artificial agent is able to learn the task. This suggests that the diversity of the training dataset was a limiting factor.

Having a correct mental model of a technical system facilitates interaction and problem solving. To assess such mental models of system functioning, appropriate methods are needed. We tested whether concept mapping with a focus on means-ends relations leads to valid assessments of participants' mental models of system functioning. Automotive and utility vehicle apprentices constructed concept maps of two simple, everyday systems (bike, traffic) and one complex, domain-specific system (fuel temperature control). However, only one group of participants had previously covered the complex system in class. Aspects of participants' concept maps regarding content (correct functional propositions) and structure (intersection over union) were assessed and related to respective reference maps. Results indicated that group differences in knowledge about the complex system were represented by concept map content, but not structure. We argue that the applied structural reference might need to be adapted to match typical requirements of the domain and task.

In the remote triad task, participants judge the relatedness between randomly chosen words in a three-alternative choice triadic judgement task. While most word pairs in these triads are weakly related, humans agree on which to choose. This is theoretically interesting as it contradicts previous claims that suggest that the notion of similarity is unconstrained in principle (e.g., Goodman, 1972}. Here, we present new evidence from GPT-4, showing that context-aware LLMs provide excellent predictions of this task. Moreover, the strength of this effect was even larger than that found for basic-level comparisons, which involve highly similar items. Together, this implies that the similarity of human representations is highly structured at every scale, even in tasks with limited context. Follow-up analysis provides insights into how LLMs are successful in this task. Further implications of the ability to compare words at every scale are discussed.

The human cognitive system exhibits remarkable flexibility and generalization capabilities, partly due to its ability to form low-dimensional, compositional representations of the environment. In contrast, standard neural network architectures often struggle with abstract reasoning tasks, overfitting, and requiring extensive data for training. This paper investigates the impact of the relational bottleneck—a mechanism that focuses processing on relations among inputs—on the learning of factorized representations conducive to compositional coding and the attendant flexibility of processing. We demonstrate that such a bottleneck not only improves generalization and learning efficiency, but also aligns network performance with human-like behavioral biases. Networks trained with the relational bottleneck developed orthogonal representations of feature dimensions latent in the dataset, reflecting the factorized structure thought to underlie human cognitive flexibility. Moreover, the relational network mimics human biases towards regularity without pre-specified symbolic primitives, suggesting that the bottleneck fosters the emergence of abstract representations that confer flexibility akin to symbols.

An important aspect of mathematical and computational thinking is algorithmic thinking––the analysis of systems, algorithms, and natural processes. A fundamental skill in algorithmic thinking is estimating the growth of functions with increasing input size. In this study, we asked 178 participants to estimate values of seven common functions in algorithmic analysis [log(n), sqrt(n), nlog(n), n^2, n^3, 2^n, n!] to understand their intuitive perception of their growth. Their estimates were fit against the actual values for all functions. Participants showed a linearization bias: sublinear functions were best fit by a linear function, and superlinear functions were best fit by a cubic (i.e., polynomial) function, even those that grow much faster (e.g., n!). In addition, participants estimated logarithmic functions least accurately. These results provide insight into how people perceive the growth of functions and set the stage for future studies of how to best improve people's reasoning about functions more generally.

Semanticists have debated the extent to which modality and causation are related in natural language. This paper aims to promote a theory in which overt causatives share core components of meaning with deontic modals. We report a sentence recall experiment that suggests that priming can be used to target the high-level semantic representations shared between two syntactically distinct linguistic expressions. Our results show that it's possible to prime the production of the deontic modal 'had to' (e.g., ``George 'had to' go to the store"), with causative 'made' (e.g., ``Jane 'made' George go to the store''), suggesting that the two expressions share a component of their meaning. Our results contribute to the methodological development in experimental semantics by establishing the utility of the priming effect to target meaning.

As mental representations are standardly thought to underlie all cognitive processes, a major goal of cognitive science has been to uncover representations. Methods for representation learning from behavioral data often model choice or reaction time data alone, but not jointly, leaving out potentially useful information. Here we develop two models of choice and RT in the odd-one-out task, including one based on the Linear Ballistic Accumulator. Parameter recovery simulations show joint modeling of choice and RT with LBA recovers representations more accurately than modeling choice alone with softmax. However, on two empirical datasets of images and words, joint models performed no better than choice-only models, despite a significant correlation of reaction time with two measures of similarity and choice difficulty in both datasets. We speculate on reasons for the unrealized promise of joint modeling of RT and choice in representation learning.

Index systems are central to our everyday and intellectual lives. Their ubiquity and diversity make them an important class of cognitive artifacts, the study of which has implications for our understanding of representational systems in general. This paper builds schema-theoretic network models of the nature of the memory structures, that underpin the interpretation of indexing systems. We identify four common classes of index systems. Using Representation Interpretation Structure Theory, we ex-plain how the four basic classes can be responsible for the substantial diversity among index systems.

In similarity comparisons, people often show a preference for one direction over the other. Bowdle and Gentner (1997) proposed the base systematicity advantage account to explain this—namely, that people prefer similarity comparisons in which the more systematic item serves as the base. Results from a series of studies supported this account. However, the studies only covered literal similarity comparisons. The question of whether analogical comparison follows the base advantage pattern remained untested. Therefore, the present study investigated this question for analogical comparisons. We tested the prediction that a comparison will be preferred when the more systematic item serves as the base. This prediction was supported. We also found support for a further prediction: namely, that inferences were projected from the systematic to the less systematic passage. Further, these inferences spontaneously arose even when not requested. The overall results from these processes are consistent with the base systematicity advantage account.

Despite the exceptional performance of large language models (LLMs) on a wide range of tasks involving natural language processing and reasoning, there has been sharp disagreement as to whether their abilities extend to more creative human abilities. A core example is the interpretation of novel metaphors. Given the enormous and non-curated text corpora used to train LLMs, a serious obstacle to designing tests is the need to obtain novel yet high-quality metaphors that are unlikely to have been included in the training data. Here we assessed the ability of GPT-4, a state-of-the-art large language model, to provide natural-language interpretations of novel literary metaphors drawn from Serbian poetry and translated into English. Human judges—blind to the fact that an AI model was involved—rated metaphor interpretations generated by GPT-4 as superior to those provided by a group of college students. In interpreting reversed metaphors, GPT-4, as well as humans, exhibited signs of sensitivity to the Gricean cooperative principle. These results indicate that LLMs such as GPT-4 have acquired an emergent ability to interpret literary metaphors.