Search

Throughout their lives, humans seem to learn a variety of rules for things like applying category labels, following procedures, and explaining causal relationships. These rules are often algorithmically rich but are nonetheless acquired with minimal data and computation. Symbolic models based on program learning successfully explain rule-learning in many domains, but performance degrades quickly as program complexity increases. It remains unclear how to scale symbolic rule-learning methods to model human performance in challenging domains. Here we show that symbolic search over the space of metaprograms-programs that revise programs-dramatically improves learning efficiency. On a behavioral benchmark of 100 algorithmically rich rules, this approach fits human learning more accurately than alternative models while also using orders of magnitude less search. The computation required to match median human performance is consistent with conservative estimates of human thinking time. Our results suggest that metaprogram-like representations may help human learners to efficiently acquire rules.

Program induction is an appealing model for human concept learning, but faces scaling challenges in searching the massive space of programs. We propose a computational model capturing two key aspects of human concept learning – our ability to judge how promising a vague, partial hypothesis is, and our ability to gradually refine these vague explanations of observations to precise ones. We represent hypotheses as probabilistic programs with randomness in place of unresolved programmatic structure. To model the evaluation of partial hypotheses, we implement a novel algorithm for efficiently computing the likelihood that a probabilistic program produces the observations. With this, we guide a search process whereby high-entropy, coarse programs are iteratively refined to introduce deterministic structure. Preliminary synthesis results on list manipulation and formal grammar learning tasks show improvements in sample efficiency when leveraging likelihood guidance, and a preliminary human study explores how model intermediate hypotheses compare to those of participants.

Reducing inequality, mitigating climate change, and responding to public health crises are large-scale goals that require the cooperation and coordination of many individuals. These goals cannot be achieved by one individual alone, and contributing is not always beneficial to each individual. And yet, individuals must contribute in order to make a difference. How do we hold individuals and groups responsible for collective action?

We offer a computational framework for modeling explanation as cooperative rational communication. Under our framework, when an explainer is faced with a ``why?'' question, they reason about the question-asker's current mental model, and intervene on that mental model in order to maximize the listener's future utility. We instantiate our framework in a planning domain, and show that our framework can model human explanations about plans across a wide variety of scenarios.

There has been exciting recent progress in computational modeling of moral cognition. Work in this area tends to describe the cognitive mechanisms of human moral judgment using symbolic models, which are interpretable and written in terms of representations that carry meaning. However, these models fall short of capturing the full human capacity to make moral judgments in that they fail to process natural language inputs but instead rely on formal problem specifications. The inability to interface with natural language also limits the usefulness of symbolic models. Meanwhile, there have been steady advances in conversational AI systems built using large language models (LLMs) that interface with natural language. However, these systems fall short as models of human reasoning, particularly in the domain of morality. In this paper we explore the possibility of building neuro-symbolic models of human moral cognition that use the strengths of LLMs to interface with natural language (specifically, to extract morally relevant features from it) and the strengths of symbolic approaches to reason over representations. Our goal is to construct a model of human moral cognition that interfaces with natural language, predicts human moral judgment with high accuracy, and does so in a way that is transparent and interpretable.

One hallmark of human language is its combinatoriality---reusing a relatively small inventory of building blocks to create a far larger inventory of increasingly complex structures. In this paper, we explore the idea that combinatoriality in language reflects a human inductive bias toward representational efficiency in symbol systems. We develop a computational framework for discovering structure in a writing system. Built on top of state-of-the-art library learning and program synthesis techniques, our computational framework discovers known linguistic structures in the Chinese writing system and reveals how the system evolves towards simplification under pressures for representational efficiency. We demonstrate how a library learning approach, utilizing learned abstractions and compression, may help reveal the fundamental computational principles that underlie the creation of combinatorial structures in human cognition, and offer broader insights into the evolution of efficient communication systems.

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.

We can evaluate features of problems and their potential solutions well before we can effectively solve them. When considering a game we have never played, for instance, we might infer whether it is likely to be challenging, fair, or fun simply from hearing the game rules, prior to deciding whether to invest time in learning the game or trying to play it well. Many studies of game play have focused on optimality and expertise, characterizing how people and computational models play based on moderate to extensive search and after playing a game dozens (if not thousands or millions) of times. Here, we study how people reason about a range of simple but novel connect-n style board games. We ask people to judge how fair and how fun the games are from very little experience: just thinking about the game for a minute or so, before they have ever actually played with anyone else, and we propose a resource-limited model that captures their judgments using only a small number of partial game simulations and almost no lookahead search. For more information about this project, see https://sites.google.com/view/intuitive-game-theory

We readily get intuitions about a problem's complexity, how much thinking it will require to solve, and how long it should take, both for ourselves and for others. These intuitions allow us to make inferences about other people's mental processing---like whether they are thinking hard, remembering, or merely mind-wandering. But where do these intuitions come from? Prior work suggests that people try solving problems themselves so as to draw inferences about another person's thinking. If we use our own thinking to build up expectations about other people, does this introduce biases into our judgments? We present a behavioral experiment testing for effects of first-person thinking speed on judgments about another person's thinking in the puzzle game Rush Hour. Although participants overwhelmingly reported solving the puzzles themselves, we found no evidence for participants' thinking speeds influencing their judgments about another person's thinking, suggesting that people can correct for first-person biases.

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.

Questions combine our mastery of language with our remarkable facility for reasoning about uncertainty. How do people navigate vast hypothesis spaces to pose informative questions given limited cognitive resources? We study these tradeoffs in a classic grounded question-asking task based on the board game Battleship. Our language-informed program sampling (LIPS) model uses large language models (LLMs) to generate natural language questions, translate them into symbolic programs, and evaluate their expected information gain. We find that with a surprisingly modest resource budget, this simple Monte Carlo optimization strategy yields informative questions that mirror human performance across varied Battleship board scenarios. In contrast, LLM-only baselines struggle to ground questions in the board state; notably, GPT-4V provides no improvement over non-visual baselines. Our results illustrate how Bayesian models of question-asking can leverage the statistics of language to capture human priors, while highlighting some shortcomings of pure LLMs as grounded reasoners.

People sometimes display moral flexibility by deciding that a commonly accepted moral norm ought not to apply in particular circumstances. But how? We explore this question in the context of queuing. We show that people's judgements about the moral permissibility of queue-cutting can be explained through cognitive processes related to moral contractualism: universalization, virtual bargaining, and functional thinking. Participants were presented vignettes depicting prospective queue-cutters, and asked whether it was morally permissible to queue-cut in those circumstances. We model these judgements with reference to the existence of a game-theoretic equilibrium supporting queue cutting in a repeated game, and to considerations of whether queue cutting would subvert or enact the function of a queue: if you pay the waiting cost, you should get the reward. These results support the notion that moral flexibility is in part related to contractualist moral principles.

Children know a lot about counting, even before they can count; for instance, even toddlers know that the counting routine involves establishing one-to-one correspondence between number words and items counted. Here we varied the size, numerosity, density, and layout of elements of sets, and asked children which set was easier to count in pair-wise comparisons across twelve trials. We also asked children themselves to count 5 to 15 items arranged in straight lines. Even children who could not count to 15 recognized that it was easier to count fewer than more dots and recognized that structured sets were easier than random arrays; however, they failed to recognize that some layouts made tracking easier than others. This suggests that children's meta-knowledge about counting precedes their ability to count for some but not all properties of sets.

Much of social cognition involves reasoning about others'minds: predicting their reactions, inferring their feelings, andexplaining their behavior. By representing mental contents likebeliefs, desires, and emotions, Bayesian theory of mind mod-els have made progress in capturing how humans manage thesecognitive feats. But social life is not merely observation: hu-mans must also plan to intervene on these same mental con-tents. The present work models how people choose interven-tions to influence others' emotions. Building on a prior modelof people's intuitive theory of emotions, we model how peopleuse their intuitive theory to evaluate and simulate the effectsof different interventions. We apply our model to data frombehavioral experiments requiring counterfactual and joint in-terventions, and show a close alignment with human choices.Our results provide a step towards a potentially unifying expla-nation for emotion prediction and intervention, suggesting thatthey could arise from the same underlying generative model.

Despite the fact that beliefs are mental states that cannot be directly observed, humans talk about each others' beliefs on a regular basis, often using rich compositional language to describe what others think and know. What explains this capacity to interpret the hidden epistemic content of other minds? In this paper, we take a step towards an answer by grounding the semantics of belief statements in a Bayesian theory-of-mind: By modeling how humans jointly infer coherent sets of goals, beliefs, and plans that explain an agent's actions, then evaluating statements about the agent's beliefs against these inferences via epistemic logic, our framework provides a functional role semantics for belief, explaining the gradedness and compositionality of human belief attributions, as well as their intimate connection with goals and plans. We evaluate this framework by studying how humans attribute goals and evaluate belief sentences while watching an agent solve a doors-and-keys gridworld puzzle that requires instrumental reasoning about hidden objects. In contrast to pure logical deduction, non-mentalizing baselines, and mentalizing that ignores the role of instrumental plans, our model provides a much better fit to human goal and belief attributions, demonstrating the importance of theory-of-mind for modeling how humans understand language about beliefs.

The space of human goals is tremendously vast; and yet, from just a few moments of watching a scene or reading a story, we seem to spontaneously infer a range of plausible motivations for the people and characters involved. What explains this remarkable capacity for intuiting other agents' goals, despite the infinitude of ends they might pursue? And how does this cohere with our understanding of other people as approximately rational agents? In this paper, we introduce a sequential Monte Carlo model of open-ended goal inference, which combines top-down Bayesian inverse planning with bottom-up sampling based on the statistics of co-occurring subgoals. By proposing goal hypotheses related to the subgoals achieved by an agent, our model rapidly generates plausible goals without exhaustive search, then filters out goals that would be irrational given the actions taken so far. We validate this model in a goal inference task called Block Words, where participants try to guess the word that someone is stacking out of lettered blocks. In comparison to both heuristic bottom-up guessing and exact Bayesian inference over hundreds of goals, our model better predicts the mean, variance, efficiency, and resource rationality of human goal inferences, achieving similar accuracy to the exact model at a fraction of the cognitive cost, while also explaining garden-path effects that arise from misleading bottom-up cues. Our experiments thus highlight the importance of uniting top-down and bottom-up models for explaining the speed, accuracy, and generality of human theory-of-mind.

We tailor the explanations we give depending on the person asking for them – you would explain why an event happened differently depending on which of the contributing causes the listener already knows. While significant prior work focuses on how causal structure in the world influences explanation, we focus on how explanation production is modulated by listener belief. We propose a computational model framing explanation as rational communication about causal events, using a recursive theory-of-mind and language production framework to choose amongst possible explanatory utterances that minimize the divergence between speaker and listener belief about a why an event happened. We evaluate our model using some partial observer stimuli, which manipulate the listener's stated prior knowledge about an event, and find that our model well-predicts human judgements about which of several contributing causes is the best explanation for a speaker to provide by modeling their communicative value to the listener.

Human ecological success stems from our ability to absorb and build upon cultural knowledge, a process we aim to model computationally by integrating individual and cultural learning from language — one of the main vehicles of cultural transmission (e.g. instructions, explanations, stories). In simple video games, our model infers game rules from both interaction data (individual learning) and partial causal models extracted from game descriptions (cultural learning). Given exhaustive descriptions (either hand-written or generated by a model given access to oracle data), models leveraging the two learning sources induce more accurate game rules from limited data than both the individual- and cultural-only controls. Interestingly, descriptions from human game players do not consistently yield better rule induction. We hypothesize that players may preferentially communicate information that will be essential to the others' future decision-making and we aim to investigate cultural transmission by integrating individual and cultural learning with both causal understanding and decision-making.

BackgroundDeep learning has shown potential in various scientific domains but faces challenges when applied to complex, high-dimensional multi-omics data. Alzheimer's Disease (AD) is a neurodegenerative disorder that lacks targeted therapeutic options. This study introduces the Circular-Sliding Window Association Test (c-SWAT) to improve the classification accuracy in predicting AD using serum-based metabolomics data, specifically lipidomics.MethodsThe c-SWAT methodology builds upon the existing Sliding Window Association Test (SWAT) and utilizes a three-step approach: feature correlation analysis, feature selection, and classification. Data from 997 participants from the Alzheimer's Disease Neuroimaging Initiative (ADNI) served as the basis for model training and validation. Feature correlations were analyzed using Weighted Gene Co-expression Network Analysis (WGCNA), and Convolutional Neural Networks (CNN) were employed for feature selection. Random Forest was used for the final classification.FindingsThe application of c-SWAT resulted in a classification accuracy of up to 80.8% and an AUC of 0.808 for distinguishing AD from cognitively normal older adults. This marks a 9.4% improvement in accuracy and a 0.169 increase in AUC compared to methods without c-SWAT. These results were statistically significant, with a p-value of 1.04 × 10ˆ-4. The approach also identified key lipids associated with AD, such as Cer(d16:1/22:0) and PI(37:6).InterpretationOur results indicate that c-SWAT is effective in improving classification accuracy and in identifying potential lipid biomarkers for AD. These identified lipids offer new avenues for understanding AD and warrant further investigation.FundingThe specific funding of this article is provided in the acknowledgements section.

ObjectivesClinical artificial intelligence and machine learning (ML) face barriers related to implementation and trust. There have been few prospective opportunities to evaluate these concerns. System for High Intensity EvaLuation During Radiotherapy (NCT03775265) was a randomised controlled study demonstrating that ML accurately directed clinical evaluations to reduce acute care during cancer radiotherapy. We characterised subsequent perceptions and barriers to implementation.MethodsAn anonymous 7-question Likert-type scale survey with optional free text was administered to multidisciplinary staff focused on workflow, agreement with ML and patient experience.Results59/71 (83%) responded. 81% disagreed/strongly disagreed their workflow was disrupted. 67% agreed/strongly agreed patients undergoing intervention were high risk. 75% agreed/strongly agreed they would implement the ML approach routinely if the study was positive. Free-text feedback focused on patient education and ML predictions.ConclusionsRandomised data and firsthand experience support positive reception of clinical ML. Providers highlighted future priorities, including patient counselling and workflow optimisation.

Human physical scene understanding requires more than simply localizing and recognizing objects Ñ we can quickly adapt our predictions about how a scene will unfold by incorporating objects' latent physics properties, such as the masses of the objects in the scene. What are the underlying computational mechanisms that allow humans to infer these physical properties and adapt their physical predictions so efficiently from visual inputs? One hypothesis is that general intuitive physics knowledge can be learned from enough raw data, instantiated as computational models that predict future video frames in large datasets of complex scenes. To test this hypothesis, we evaluate existing state-of-the-art video models. We measured both model and human performance on Physion++, a novel dataset and benchmark that rigorously evaluates visual physical prediction in humans and machines, under circumstances where accurate physical prediction relies on accurate estimates of the latent physical properties of objects in the scene. Specifically, we tested scenarios where accurate prediction relied on accurate estimates of objects' mechanical properties, including masses, friction, elasticity and deformability, and the values of these mechanical properties could only be inferred by observing how these objects moved and interacted with other objects and/or fluids. We found that models that encode objectness and physical states tend to perform better, yet there is still a huge gap compared to human performance. We also found most models' predictions correlate poorly with that made by humans. These results show that current deep learning models that succeed in some settings nevertheless fail to achieve human-level physical prediction in other cases, especially those where latent property inference is required.

Human language richly invokes our intuitive physical knowledge. We talk about physical objects, scenes, properties, and events; and we can ask questions and answer them with predictions and inferences about physical worlds described entirely in language. How does language construct meanings that connect to our general physical reasoning? In this paper, we propose PiLoT, a computational model that maps language into a probabilistic language of thought—meanings are constructed as probabilistic programs, which provide a formal basis for probabilistic and physical reasoning. Our model uses a large language model (LLM) to map from language to meanings and a probabilistic physics engine to support inferences over scenes described in language. We conduct a linguistic reasoning experiment based on prior psychophysics studies that requires reasoning about physical outcomes based on linguistic descriptions. We show that PiLoT well predicts human judgments across this experiment and outperforms baseline models which use the LLM to directly perform the same task.

Many models of intuitive physical reasoning posit some kindof mental simulation mechanism, yet everyday environmentsfrequently contain far more objects than people could plau-sibly represent with their limited cognitive capacity. Whatdetermines which objects are actually included in our repre-sentations? We asked participants to predict how a ball willbounce through a complex field of obstacles, and probed work-ing memory for objects in the scene that were more and lesslikely to be relevant to the ball’s trajectory. We evaluate differ-ent accounts of relevance and find that successful object mem-ory is best predicted by how frequently a ball’s trajectory isexpected to contact that object under a probabilistic simulationmodel. This suggests that people construct representations formental simulation efficiently and dynamically, on the fly, byadding objects “just in time”: only when they are expected tobecome relevant for the next stage of simulation.

Humans can successfully interpret images even when significant image transformations have distorted them. Such images aid in differentiating existing computational models for perception because models which predict similar results for typical stimuli may diverge when confronted with atypical stimuli. We propose an explanation of people’s ability to perceive a specific class of degraded stimuli that require extreme generalization capabilities: Mooney, or two-tone images of faces. We provide a parametric and controllable method to generate Mooney faces and hypothesize that the human visual system generalizes under extreme cases because it reverses the naturalistic image generation process. We show that Mooney faces are approximated under extreme illumination conditions; a computer graphics perspective of the Mooney formation process suggests inverse graphics as a model for perceptual generalization. With a number of demonstrations, we argue that Mooney faces are important to understand the role of generative models and inverse graphics in human visual processing.

If we want to know whether a stack of blocks will fall, we might form our judgment by using mental simulation or simple heuristics. How do we know which strategy to use? Here, we test whether people’s strategy selection is “resource rational,” balancing their utilities and costs. We had participants judge how a balance beam will fall – a task that relies on both simulation and heuristics – and imposed visuospatial cognitive load to increase the cost of mental simulation, expecting the use of simulation to then decrease and the use of heuristics to increase. When cognitive load varied between participants, higher load led to less simulation. However, when load varied across trials for an individual participant, they instead used simulation at about the same rate as high-load-only participants. Thus strategy selection is sensitive to cognitive costs but it is also sticky, not accounting for rapidly fluctuating costs.

Research in psychology and artificial intelligence has sought to ground information-seeking behavior in rational terms, typically assuming that people or agents prefer more informative data over less informative data. While this seems reasonable on its surface, it assumes that informativeness is only a property of the data, rather than a joint property of the data and a (potentially bounded) learner. That is, to the extent that it is hard to draw the right inferences from data that are theoretically "high information," the data will not actually be highly informative to the learner. Here, we investigate active learning in humans using the code-breaking game Mastermind, which requires deductive reasoning from evidence. We find that people make queries that are less informative than random guesses, challenging standard rational or resource-rational accounts of information-seeking. We then show that people make queries are informative to them assuming they have a bounded capacity to draw inferences. We also find that participants prefer queries that provide easily-interpretable information over queries that provide more information but are less interpretable. Our results suggest that people are aware of their own cognitive limitations and seek information that they can use.

The relationship between communicated language and intended meaning is often probabilistic and sensitive to context. Numerous strategies attempt to estimate such a mapping, often leveraging recursive Bayesian models of communication. In parallel, LLMs have been increasingly applied to semantic parsing applications, tasked with inferring logical representations from natural language. While existing LLM explorations have been largely restricted to literal language use, in this work, we evaluate the capacity of LLMs to infer the meanings of pragmatic utterances. Specifically, we explore the case of threshold estimation on the gradable adjective “strong”, contextually conditioned on a strength prior, then extended to composition with qualification, negation, polarity inversion, and class comparison. We find that LLMs can derive context-grounded, human-like distributions over the interpretations of several complex pragmatic utterances, yet struggle composing with negation. These results inform the inferential capacity of statistical language models, and their use in pragmatic and semantic parsing applications.

In science, we use common graphical representations to indicate changes in events over time, independent of domain. Are children also sensitive to abstract patterns in the ways events change over time? In a series of four experiments, we show that young children (range: 48-84 months) distinguish different function families (Exp 1). Children can also distinguish specific function types within function families (e.g., between linear and sigmoid monotonic functions; Exp 2); nonetheless, they will group different function types within a family together, rather than with functions in a different family (Exp 3). Finally, we show that children’s sensitivity to functions is abstract, allowing them to match observable causes to verbal descriptions of their effects (Exp 4). These results suggest that although some aspects of function understanding, like learning how to interpret graphs, requires formal education, the ability to identify abstract functional relationships is intuitive and early emerging.

How do we know when it's OK to break moral rules? We propose that — alongside well-studied outcome-based measures of welfare and harm — people sometimes use universalization, asking "What if everyone felt at liberty to ignore the rule?'' We develop a virtual environment where agents stand in line to gather water. Subjects judge agents who get out of line to try to get water more quickly. If subjects use universalization, they would need to imagine all agents getting out of line and going straight for the water in each environment. To test this prediction, we model an action's universalizability by simulating what would happen if every agent tried to follow a path directly to the water, then evaluating the effects. We also investigate the role of several outcome-based measures, including welfare aggregation and harm-based measures. We find that universalizability plays a important role in rule-breaking judgments alongside outcome-based concerns.

Great storytelling takes us on a journey the way ordinary reality rarely does. But what exactly do we mean by a "journey"? Recently, literary theorist Kukkonen (2014) proposed that storytelling is "probability design": the art of giving an audience pieces of information bit by bit, to craft the *journey of their changing beliefs* about the fictional world. A good "probability design" choreographs a delicate dance of certainty and surprise in the reader's mind as the story unfolds from beginning to end. In this paper, we computationally model this conception of storytelling. Building on the classic Bayesian inverse planning model of human social cognition, we treat storytelling as *inverse* inverse planning: the task of choosing actions to manipulate an inverse planner's inferences, and therefore a human audience's beliefs. First, we use an inverse inverse planner to depict social and physical situations, and present behavioral studies indicating that inverse inverse planning produces more expressive behavior than ordinary "naïve planning." Then, through a series of examples, we demonstrate how inverse inverse planning captures many storytelling elements from first principles: character, narrative arcs, plot twists, irony, flashbacks, and deus ex machina are all naturally encoded in the flexible language of probability design.

Bayesian models of theory of mind (ToM) have been successful in explaining how humans infer goals from the actions of other agents. However, they have typically been limited to small and fixed sets of possible goals specified in advance by the modeler, leaving open the question of how spontaneous goal inference occurs in rich and complex environments. To address this question, we posit that people are guided by informative, context-specific goal priors and proposals when making inferences about others. As proxies for these informed priors, we make use of context-conditioned large language models (LLMs) and integrate them into a Bayesian inverse planning framework. We find that LLMs can serve as usefully informative priors and proposals over goals compared to a structural baseline prior, allowing them to be used as models of the learned statistical knowledge that humans bring to bear in their inferences about others' goals.

Confidence is an important concept in cognitive science, as it integrates seamlessly with our beliefs, goals, and decisions. Humans naturally represent and express degrees of confidence in beliefs and predictions that reflect their accuracy. However, the dynamics of how our underlying beliefs about the world relate to explicitly represented confidence over those beliefs is yet not well understood. In this work, we make progress on this question in the domain of \textit{concept learning}. Specifically, we analyze how confidence and beliefs jointly evolve in the absence of explicit feedback. We evaluate some leading computational accounts of confidence in the present literature, and we find that these accounts do not accurately predict confidence in the context of our task. We advocate for caution in making claims about the generalizability of such accounts across tasks or domains and propose some new model-based measurements for predicting humans' confidence judgments. Of these measurements, we find that ones taking individual-level response patterns into account perform the best. We close by suggesting promising future directions for the study of confidence in concept learning.

Theory of Mind enables us to attribute mental states to others. But we not only make inferences about mental states (like what someone believes or wants), but about mental processes (like if someone is distracted or whether they remember something). Here, we present a computational formalization of these kinds of inferences. We propose that inferences about mental processes are structured around a principle of rational mental effort: the expectation that other people allocate mental resources rationally so as to minimize thinking costs incurred while pursuing their goals. We develop this theory into a computational model in the context of the Rush Hour puzzle game. In two behavioral experiments testing different inferences about mental processing, we find that our model predicts participant judgments. This work advances our understanding of the richness of the human mind’s ability to think about other minds, and even about thinking itself.

BackgroundDisparity in mental health care among cancer patients remains understudied.MethodsA large, retrospective, single tertiary-care institution cohort study was conducted based on deidentified electronic health record data of 54,852 adult cancer patients without prior mental health diagnosis (MHD) diagnosed at the University of California, San Francisco between January 2012 and September 2019. The exposure of interest was early-onset MHD with or without psychotropic medication (PM) within 12 months of cancer diagnosis and primary outcome was all-cause mortality.ResultsThere were 8.2% of patients who received a new MHD at a median of 197 days (interquartile range, 61-553) after incident cancer diagnosis; 31.0% received a PM prescription; and 3.7% a mental health-related visit (MHRV). There were 62.6% of patients who were non-Hispanic White (NHW), 10.8% were Asian, 9.8% were Hispanic, and 3.8% were Black. Compared with NHWs, minority cancer patients had reduced adjusted odds of MHDs, PM prescriptions, and MHRVs, particularly for generalized anxiety (Asian odds ratio [OR], 0.66, 95% CI, 0.55-0.78; Black OR, 0.60, 95% CI, 0.45-0.79; Hispanic OR, 0.72, 95% CI, 0.61-0.85) and selective serotonin-reuptake inhibitors (Asian OR, 0.43, 95% CI, 0.37-0.50; Black OR, 0.51, 95% CI, 0.40-0.61; Hispanic OR, 0.79, 95% CI, 0.70-0.89). New early MHD with PM was associated with elevated all-cause mortality (12-24 months: hazard ratio [HR], 1.43, 95% CI, 1.25-1.64) that waned by 24 to 36 months (HR, 1.18, 95% CI, 0.95-1.45).ConclusionsNew mental health diagnosis with PM was a marker of early mortality among cancer patients. Minority cancer patients were less likely to receive documentation of MHDs or treatment, which may represent missed opportunities to identify and treat cancer-related mental health conditions.

The way two people choose to share food reveals how closetheir relationship is. Very close relationships alleviate the discomfort of exchanging saliva. We measure human inferencesabout relationships from observed food sharing actions withvariable risks of saliva exchange; and then use a formal modelof inverse planning to quantitatively capture these inferences.The model that best fits human judgments construes food sharing as a rational communicative social action, according towhich actions are chosen both to maximize comfort given arelationship, and to communicate about the relationship itself.

Humans remarkably can perceive 3D surfaces that are literally invisible – when their shapes are implied through dynamic physical interaction with various materials (liquid, granular, elastic). To study this ability and its interaction with perception of material properties, we produced videos simulating unfamiliar non-rigid or non-solid materials interacting with differently shaped rigid objects and containers. Crucially, videos rendered only the materials, while the rigids remained invisible. Observers correctly identified the underlying surface shape from two alternatives in most, but not all cases; observers also simultaneously inferred the internal properties of the material and identified a matching sample from how it interacted with novel surfaces. In fact, judgments about materials were comparable to when observers saw fully visible renderings of all objects in the scenes. Our results highlight the role of an internal physics model in the joint perception of shape and material properties, and in the loop of perception more generally.

Human language offers a powerful window into our thoughts -- we tell stories, give explanations, and express our beliefs and goals through words. Abundant evidence also suggests that language plays a developmental role in structuring our learning. Here, we ask: how much of human-like thinking can be captured by learning statistical patterns in language alone? We first contribute a new challenge benchmark for comparing humans and distributional large language models (LLMs). Our benchmark contains two problem-solving domains (planning and explanation generation) and is designed to require generalization to new, out-of-distribution problems expressed in language. We find that humans are far more robust than LLMs on this benchmark. Next, we propose a hybrid Parse-and-Solve model, which augments distributional LLMs with a structured symbolic reasoning module. We find that this model shows more robust adaptation to out-of-distribution planning problems, demonstrating the promise of hybrid AI models for more human-like reasoning.

How do humans find their way in new environments, so quickly and efficiently? Humans reuse old knowledge to build new concepts in many non-spatial domains, such as language and drawing. Could people learn maps by a similar process, that extracts common structure to speed up learning, and generalize across maps? Understanding the computational cognitive mechanisms that support this efficiency can advance the study of the human mind and enable more efficient exploration algorithms. We hypothesize that human map learning relies on inferences over the structure of unobserved spaces, based on spatial priors informed by previous experience. We model this by combining Program Induction with a Hierarchical Bayesian framework that explicitly reasons about uncertainty through strong spatial priors. Using a new behavioral Map Induction Task, we demonstrate that this computational framework explains human exploration behavior better than non-inductive models and outperforms state-of-the-art planning algorithms in a realistic spatial navigation domain.

When deciding whether and how to punish, people consider not only the potential direct consequences, but also, how their choice will affect observers’ judgements about the values and motives underlying the choice. We formalize the decision to punish as a rational communicative social action (RCSA). The model generates rational decisions to punish, incorporating anticipated observers’ judgements obtained from a recursive model of inference using an intuitive theory of mind. Using this model, we synthesize patterns of human punishment from recently published papers. RCSA thus offers a formal model of the cognitive process that humans use to balance preferences for how they are perceived, with other goals for punishing.

Rules undoubtedly guide our moral lives. Simple moral rules prohibit lying, cheating, and stealing, for instance. But the moral mind is more flexible than a theory based only on rule-adherence can account for. In this paper, we look at one particular kind of flexibility: the ability to figure out when it is okay to break a moral rule. We elicit judgments of the moral acceptability of breaking a simple rule: it's wrong to cut in line. We created a video game environment in which agents attempt to gather water -- and sometimes must stand in line behind others to do so. Subjects watch clips of the game being played and make judgments of the moral acceptability of cutting in line across a wide range of spatio-temporally varied and dynamic scenarios. Our data suggests that subjects make judgments by using a generative understanding of the underlying function of the rule about waiting in line. We further show that our data cannot be accounted for by either 1) simple rule adherence or 2) utility maximization.

We introduce a new benchmark dataset based on classic methods of studying 3D shape perception from texture and motion, inspired by earlier work on Gestalt principles of perceptual organization and the ecological (Gibsonian) approach to perception of structure in moving displays. The dataset consists of parametric 3D shapes (superquadrics) with procedurally generated textures rotating and translating against a similarly textured backdrop. We expect these stimuli to be challenging for current computer vision models, as they depart from the statistics of real-world or realistically rendered stimuli. We test a variety of models’ ability to segment textured stimuli across three training conditions: pre-trained on naturalistic stimuli, pre-trained+fine-tuned on textured stimuli, trained on textured stimuli. While no models generalize to segment textured stimuli without fine-tuning, performance improves with fine-tuning and training on textured stimuli. We will discuss how this benchmark can guide models of scene perception towards more human-like robustness and generality.

Causal relationships can generate many different kinds of correlations among variables. However, research on children’s causal reasoning has focused almost exclusively on just one kind of regularity: the temporal covariation between candidate causes and effects, and in particular, the covariation between interventions and outcomes. Here we show that young children recognize more abstract correlations – in the ways that object properties are distributed over sets, or change over time – and constrain their causal hypotheses accordingly. Specifically, we show that children (range: 48-84 months) distinguish candidate causes based on correlations in the distribution of discrete (set size, arity, and proportion) and continuous (mean and mode) properties of sets (Experiment 1), and also within monotonic, quadratic, and periodic functions (Experiment 2).Keywords: children; causal reasoning; abstract concepts; sets; functions

Our understanding of the visual world goes beyond naming objects, encompassing our ability to parse objects into meaningful parts, attributes, and relations. In this work, we leverage natural language descriptions for a diverse set of 2K procedurally generated objects to identify the parts people use and the principles leading these parts to be favored over others.We formalize our problem as search over a space of program libraries that contain different part concepts, using tools from machine translation to evaluate how well programs expressed in each library align to human language. By combining naturalistic language at scale with structured program representations, we discover a fundamental information-theoretic tradeoff governing the part concepts people name: people favor a lexicon that allows concise descriptions of each object, while also minimizing the size of the lexicon itself.

Accounts of human and machine concept learning face a fundamental challenge. Some approaches, notably deep learning, frequently achieve human-level performance on specific tasks but lack consistently human-like—i.e. generalizable, composable, and explainable—solutions. Others, including classic symbolic accounts, produce human-like hypotheses but scale poorly. We present a model of learning which is both human-level and human-like. It represents concepts as program-like expressions formed by applying a series of higher-order inferences that iteratively revise preexisting concepts into novel target concepts. Learning seeks the best combination of revisions under a Bayesian score. This model predicts learning behavior in 392 humans over 100 computationally sophisticated concepts more accurately than alternative models (Enumerate, Metagol, RobustFill, Fleet) while using three orders of magnitude less computation. This work shows how humans plausibly construct sophisticated algorithmic representations, a necessity for compelling human-like artificial intelligence.

Our goal is to summarize the collective experience of 15 organizations in dealing with uncoordinated efforts that result in unnecessary delays in understanding, predicting, preparing for, containing, and mitigating the COVID-19 pandemic in the US. Response efforts involve the collection and analysis of data corresponding to healthcare organizations, public health departments, socioeconomic indicators, as well as additional signals collected directly from individuals and communities. We focused on electronic health record (EHR) data, since EHRs can be leveraged and scaled to improve clinical care, research, and to inform public health decision-making. We outline the current challenges in the data ecosystem and the technology infrastructure that are relevant to COVID-19, as witnessed in our 15 institutions. The infrastructure includes registries and clinical data networks to support population-level analyses. We propose a specific set of strategic next steps to increase interoperability, overall organization, and efficiencies.

ObjectivesExpert abstraction of acute toxicities is critical in oncology research but is labor-intensive and variable. We assessed the accuracy of a natural language processing (NLP) pipeline to extract symptoms from clinical notes compared to physicians.Materials and methodsTwo independent reviewers identified present and negated National Cancer Institute Common Terminology Criteria for Adverse Events (CTCAE) v5.0 symptoms from 100 randomly selected notes for on-treatment visits during radiation therapy with adjudication by a third reviewer. A NLP pipeline based on Apache clinical Text Analysis Knowledge Extraction System was developed and used to extract CTCAE terms. Accuracy was assessed by precision, recall, and F1.ResultsThe NLP pipeline demonstrated high accuracy for common physician-abstracted symptoms, such as radiation dermatitis (F1 0.88), fatigue (0.85), and nausea (0.88). NLP had poor sensitivity for negated symptoms.ConclusionNLP accurately detects a subset of documented present CTCAE symptoms, though is limited for negated symptoms. It may facilitate strategies to more consistently identify toxicities during cancer therapy.

Our goal is to summarize the collective experience of 15 organizations in dealing with uncoordinated efforts that result in unnecessary delays in understanding, predicting, preparing for, containing, and mitigating the COVID-19 pandemic in the US. Response efforts involve the collection and analysis of data corresponding to healthcare organizations, public health departments, socioeconomic indicators, as well as additional signals collected directly from individuals and communities. We focused on electronic health record (EHR) data, since EHRs can be leveraged and scaled to improve clinical care, research, and to inform public health decision-making. We outline the current challenges in the data ecosystem and the technology infrastructure that are relevant to COVID-19, as witnessed in our 15 institutions. The infrastructure includes registries and clinical data networks to support population-level analyses. We propose a specific set of strategic next steps to increase interoperability, overall organization, and efficiencies.