Search

A major goal of linguistics and cognitive science is to understand what class of learning systems can acquire natural language. Until recently, the computational requirements of language have been used to argue that learning is impossible without a highly constrained hypothesis space. Here, we describe a learning system that is maximally unconstrained, operating over the space of all computations, and is able to acquire many of the key structures present in natural language from positive evidence alone. We demonstrate this by providing the same learning model with data from 74 distinct formal languages which have been argued to capture key features of language, have been studied in experimental work, or come from an interesting complexity class. The model is able to successfully induce the latent system generating the observed strings from small amounts of evidence in almost all cases, including for regular (e.g., an , [Formula: see text], and [Formula: see text]), context-free (e.g., [Formula: see text], and [Formula: see text]), and context-sensitive (e.g., [Formula: see text], and xx) languages, as well as for many languages studied in learning experiments. These results show that relatively small amounts of positive evidence can support learning of rich classes of generative computations over structures. The model provides an idealized learning setup upon which additional cognitive constraints and biases can be formalized.

Temporal connectives are function words that relate events intime. Despite their ubiquity and utility, children acquire themeanings of temporal connective words late in development.Experimental work has uncovered patterns in the acquisitionof temporal connectives that clarify the learning challenge thatthese words pose to children. In particular, developmentalstudies have identified differing acquisition trajectories acrossconnective types, asymmetries in learning within pairs of re-lated connectives, and monotonic increases in comprehensionwith age. Expanding on prior theoretical accounts, we formal-ize temporal connective acquisition in a computational wordlearning framework. We demonstrate that each of the empir-ically determined acquisition patterns emerges in the learningbehavior of our computational model. Finally, we discuss ourfindings in relation to earlier theories and to general learnabil-ity concerns in language acquisition.

From early in life, people implicitly associate time, number,and other abstract conceptual domains with space. Accord-ing to the Generalized Magnitude System proposal, these men-tal mappings reflect a common neural system for represent-ing various magnitudes, and share a common spatial organiza-tion. In a test of this proposal, here we measured mappings ofsize, time, and number in the Tsimane’, an indigenous Ama-zonian group with few of the cultural practices (like readingand math) that spatialize size, time, and number in the expe-rience of industrialized adults. On three spatial axes, the Tsi-mane’ systematically arranged imagistic stimuli according totheir magnitudes, but they showed no directional preferencesoverall and individuals often mapped different domains in op-posite directions. The results are inconsistent with predictionsof the Generalized Magnitude System proposal but can be ex-plained by Hierarchical Mental Metaphor Theory, accordingto which mental mappings initially reflect a set of correlationsobservable in the natural world.

Probabilities can be described by a numerator and a de-nominator and students and decision-makers are not in-different to numerical values of the components. For in-stance, when people compare two equal ratios their choicesgravitate to the option with larger number, even if theyknow both ratios are equal. To the date, however, it is un-clear if whole number biases are present in other cultures.We tested a farming-foraging group living in the Bolivianrain forest in a simple 2AFC ratio comparison task. Af-ter appropriate training, the Tsimane were highly accu-rate in this task, confirming that visual proportional rea-soning is present across cultures. Importantly, they hada strong tendency to favor large numbers in equal ratiocomparisons, similar to what is found in educated popu-lations. Even though our sample size is moderate (n=76),the whole number bias we found occurred under good pro-portional reasoning. The bias may be a general feature ofcognition, rather than a cultural or education artifact, thatmay help humans solve ambiguous situations.

BackgroundSeasons affect many social, economic, and biological outcomes, particularly in low-resource settings, and some studies suggest that birth season affects child growth.AimTo study a predictor of stunting that has received limited attention: birth season.Subjects and methodsThis study uses cross-sectional data collected during 2008 in a low-resource society of horticulturists-foragers in the Bolivian Amazon, Tsimane'. It estimates the associations between birth months and height-for-age Z-scores (HAZ) for 562 girls and 546 boys separately, from birth until age 11 years or pre-puberty, which in this society occurs ∼13-14 years.ResultsChildren born during the rainy season (February-May) were shorter, while children born during the end of the dry season and the start of the rainy season (August-November) were taller, both compared with their age-sex peers born during the rest of the year. The correlations of birth season with HAZ were stronger for boys than for girls. Controlling for birth season, there is some evidence of eventual partial catch-up growth, with the HAZ of girls or boys worsening until ∼ age 4-5 years, but improving thereafter. By age 6 years, many girls and boys had ceased to be stunted, irrespective of birth season.ConclusionThe results suggest that redressing stunting will require attention to conditions in utero, infancy and late childhood.

BackgroundSeasons affect many social, economic, and biological outcomes, particularly in low-resource settings, and some studies suggest that birth season affects child growth.AimTo study a predictor of stunting that has received limited attention: birth season.Subjects and methodsThis study uses cross-sectional data collected during 2008 in a low-resource society of horticulturists-foragers in the Bolivian Amazon, Tsimane'. It estimates the associations between birth months and height-for-age Z-scores (HAZ) for 562 girls and 546 boys separately, from birth until age 11 years or pre-puberty, which in this society occurs ∼13-14 years.ResultsChildren born during the rainy season (February-May) were shorter, while children born during the end of the dry season and the start of the rainy season (August-November) were taller, both compared with their age-sex peers born during the rest of the year. The correlations of birth season with HAZ were stronger for boys than for girls. Controlling for birth season, there is some evidence of eventual partial catch-up growth, with the HAZ of girls or boys worsening until ∼ age 4-5 years, but improving thereafter. By age 6 years, many girls and boys had ceased to be stunted, irrespective of birth season.ConclusionThe results suggest that redressing stunting will require attention to conditions in utero, infancy and late childhood.

When learning about the world, we develop mental represen-tations or concepts for things we have never seen. At the sametime, we also develop representations for things that are similarto what we have experienced. Traditionally, similarity-basedand rule-based systems have been used as distinct models tocapture conceptual representation. However, it seems implau-sible that we do not flexibly deploy both systems. Whetherboth systems can be used simultaneously to represent compo-nents of a single concept is an open empirical question. Oneexample suggesting that the use of both systems is possible isthe concept of a ZEBRA , which looks like a horse but striped.Using an artificial concept learning task, we test whether peo-ple can combine similarity and rules compositionally in orderto represent concepts. Our results suggest that people are ableto compose similarity and rules when mentally representing asingle concept.

Humans master complex systems of interrelated concepts likemathematics and natural language. Previous work suggestslearning these systems relies on iteratively and directly re-vising a language-like conceptual representation. We intro-duce and assess a novel concept learning paradigm calledMartha’s Magical Machines that captures complex relation-ships between concepts. We model human concept learning inthis paradigm as a search in the space of term rewriting sys-tems, previously developed as an abstract model of compu-tation. Our model accurately predicts that participants learnsome transformations more easily than others and that theylearn harder concepts more easily using a bootstrapping cur-riculum focused on their compositional parts. Our results sug-gest that term rewriting systems may be a useful model of hu-man conceptual representations.

In many domains, people are able to transfer abstract knowl-edge about objects, events, or contexts that are superficiallydissimilar, enabling striking new insights and inferences. Weprovide evidence that this ability is naturally explained as theaddition of new primitive elements to a compositional mentalrepresentation, such as that in the probabilistic Language OfThought (LOT). We conducted a transfer-learning experimentin which participants learned about two sequences, one afterthe other. We show that participants’ ability to learn the secondsequence is affected the first sequence they saw. We test twoprobabilistic models to evaluate alternative theories of how al-gorithmic knowledge is transferred from the first to second se-quence: one model rationally updates the prior probability ofthe primitive operations in the LOT based on what was used inthe first sequence; the other stores previously likely hypothesesas new primitives. Both models perform better than baselinesin explaining behavior, with the human subjects appearing totransfer entire hypotheses when they can, and otherwise updat-ing the prior on primitives.

A widely observed phenomenon in children’s word-extensionsand generalizations is the characteristic-to-defining shift,whereby young children initially generalize words based ontypical properties and gradually transition into generalizingwords using abstract, logical information. In this paper, wepropose a statistically principled model of conceptual devel-opment grounded in the trade-off between simplicity and fit tothe data. We run our model based on informant-provided fam-ily trees and the real-life characteristic features of people onthose trees. We demonstrate that the characteristic-to-definingshift does not necessarily depend on discrete change in rep-resentation or processes. Instead, the shift could fall out nat-urally from statistical inference over conceptual hypotheses.Our model finds that the shift occurs even when abstract logicalrelations are present from the outset of learning as long as char-acteristic features are informative but imperfect in their abilityto capture the underlying concept to be learned—a property ofour elicited features.

People have the capability to process text three times fasterthan they would naturally read it, yet many current theories ofsentence processing rely on natural reading times as a proxyfor processing difficulty. How can people read material soquickly in spite of information processing limitations sug-gested by sentence processing theories? One possibility is thatsurprisal effects on reading time, the hallmark of processingdifficulty under sentence processing theories, might arise fromperceptual processing, implying no relation between surprisaland sentence processing difficulty. In this paper, we conducteda novel self-paced rapid serial visual presentation (RSVP) ex-periment, which controlled perceptual processes to probe forsentence processing related surprisal effects. We further testedhow readers might compensate for information processing lim-its during RSVP. We find support for sentence processing re-lated surprisal effects, the pattern of which is consistent with aFirst-In, First-Out (FIFO) buffer model.

How do people rapidly learn rich, structured concepts fromsparse input? Recent approaches to concept learning havefound success by integrating rules and statistics. We describe ahierarchical model in this spirit in which the rules are stochas-tic, generative processes, and the rules themselves arise froma higher-level stochastic, generative process. We evaluate thisprobabilistic language-of-thought model with data from an ab-stract rule learning experiment carried out with adults. In thisexperiment, we find novel generalization effects, and we showthat the model gives a qualitatively good account of the exper-imental data. We then discuss the role of this kind of model inthe larger context of concept learning.

We apply Bayesian data analysis to a structured cognitivemodel in order to determine the priors that support humangeneralizations in a simple concept learning task. We mod-eled 250,000 ratings in a “number game” experiment wheresubjects took examples of a numbers produced by a program(e.g. 4, 16, 32) and rated how likely other numbers (e.g. 8vs. 9) would be to be generated. This paper develops a dataanalysis technique for a family of compositional “Language ofThought” (LOT) models which permits discovery of subjects’prior probability of mental operations (e.g. addition, multi-plication, etc.) in this domain. Our results reveal high cor-relations between model mean predictions and subject gener-alizations, but with some qualitative mismatch for a stronglycompositional prior.

We study the projection of cognitive representations intocontinuous motor (reaching) responses with a computationalmodel that unifies three influential approaches: accumulationof evidence, statistical inference, and optimalfeedback control. We modeled a number comparison taskthat asked participants to respond with a reaching gesturewhich of two side had more dots. The model successfullyreproduced subjects’ pattern of reach and performanceacross varying difficulties of numerical comparison.Our model parameterized several potentially relevantcognitive variables, including a threshold, memory decay,and mental sampling rate. Remarkably, a threshold formovement was not needed for modeling human behaviorwhen statistical inference is combined with optimal motorplanning. Overall, the model indicates that the motorsystempositions the effectors optimally, both biomechanicallythrough an optimal feedback controller, and cognitivelyby means of continuous statistical inference on theavailable evidence.

Context plays a ubiquitous role in language processing. Forthe most part, work in language processing investigates theeffects of context without investigating questions about whatdetermines a context. For example, interpretation of any referentialexpression must take into account the notion of a referentialdomain. Here we investigate the influence of perceptualcues in establishing a referential domain, or linguistic context.We demonstrate that people use perceptual cues to establish alinguistic context; the influence of perceptual cues is gradientwith respect to cue magnitude; and the contribution of a perceptualcue in constructing a linguistic context is not an effectof attention or salience. We provide these results as a first steptoward developing a formal model for the establishment of linguisticcontext.

Current models of word learning focus on the mapping betweenwords and their referents and remain mute with regardto conceptual representation. We develop a cross-situationalmodel of word learning that captures word-concept mappingby jointly inferring the referents and underlying concepts foreach word. We also develop a variant of our model that incorporatesrecursion, which entertains the idea that children canuse learned words to aid future learning. We demonstrate bothmodels’ ability to learn kinship terms and show that addingrecursion into the model speeds acquisition