Four- and six-year-old children track a single meaning with both familiar and unfamiliar referents when the referent is clear: More evidence for propose-but-verify.

In word learning, learners need to identify the referent of words by leveraging the fact that the same word may co-occur with different sets of objects. This raises the question, what do children remember from "in the moment" that they can use for cross-situational learning? Furthermore, do children represent pictures of familiar animals versus drawings of non-existent novel objects as potential referents differently? This study examined these questions by creating learning scenarios with only two potential referents, requiring the least amount of memory to represent all co-present objects. Across three experiments (n > 250) with 4- and 6-year-old children, children reliably selected the intended referent from learning at test, though the learning of novel objects was better than familiar objects. When asked for a co-present object, children of all ages in the study performed at chance in all of the conditions. We discuss the developmental differences in cross-situational word learning capabilities with regard to representing different stimuli as potential referents. Importantly, all children used a propose-but-verify procedure for learning novel words even in the simplest of the learning scenarios given repeated exposure.

An algorithmic approach to determine expertise development using object-related gaze pattern sequences.

Eye tracking (ET) technology is increasingly utilized to quantify visual behavior in the study of the development of domain-specific expertise. However, the identification and measurement of distinct gaze patterns using traditional ET metrics has been challenging, and the insights gained shown to be inconclusive about the nature of expert gaze behavior. In this article, we introduce an algorithmic approach for the extraction of object-related gaze sequences and determine task-related expertise by investigating the development of gaze sequence patterns during a multi-trial study of a simplified airplane assembly task. We demonstrate the algorithm in a study where novice (n = 28) and expert (n = 2) eye movements were recorded in successive trials (n = 8), allowing us to verify whether similar patterns develop with increasing expertise. In the proposed approach, AOI sequences were transformed to string representation and processed using the k-mer method, a well-known method from the field of computational biology. Our results for expertise development suggest that basic tendencies are visible in traditional ET metrics, such as the fixation duration, but are much more evident for k-mers of k > 2. With increased on-task experience, the appearance of expert k-mer patterns in novice gaze sequences was shown to increase significantly (p < 0.001). The results illustrate that the multi-trial k-mer approach is suitable for revealing specific cognitive processes and can quantify learning progress using gaze patterns that include both spatial and temporal information, which could provide a valuable tool for novice training and expert assessment.

Spotting Dalmatians: Children's ability to discover subordinate-level word meanings cross-situationally.

Even when children encounter a novel word in the situation of a clear and unique referent, they are nevertheless faced with the problem of semantic uncertainty: when "puziv" refers to a co-present spotted dog, does the word mean Fido, Dalmatian, dog, animal, or entity? Here we explored the extent to which children (3-5 years of age) can reason about a novel word's meaning from information they have gathered cross-situationally, from a series of simple ostensive labeling events ("I see a puziv!"). Of particular interest were the conditions under which children arrive at a subordinate level meaning (e.g., Dalmatian) rather than a basic level meaning (e.g., dog). Experiment 1 showed that children (N = 32) were capable of using lexical contrast and/or mutual exclusivity cross-situationally, such that they arrived at subordinate level meanings only when the words being learned contrasted at the subordinate level, otherwise they strongly preferred basic level meanings (e.g., dog) even when the word had previously referred to subordinate level exemplars (always Dalmatians). Experiment 2 showed that some children in this same age range (N = 20) can also arrive at subordinate level meanings cross-situationally when offered relatively minimal linguistic support ("It's a kind of dog."). The findings are interpreted with respect to current theories of cross-situational word learning, and suggest that word meanings rather than sets of referential exemplars are tracked and used for cross-situational comparison.

Successfully learning non-adjacent dependencies in a continuous artificial language stream.

Much of the statistical learning literature has focused on adjacent dependency learning, which has shown that learners are capable of extracting adjacent statistics from continuous language streams. In contrast, studies on non-adjacent dependency learning have mixed results, with some showing success and others failure. We review the literature on non-adjacent dependency learning and examine various theories proposed to account for these results, including the proposed necessity of the presence of pauses in the learning stream, or proposals regarding competition between adjacent and non-adjacent dependency learning such that high variability of middle elements is beneficial to learning. Here we challenge those accounts by showing successful learning of non-adjacent dependencies under conditions that are inconsistent with predictions of previous theories. We show that non-adjacent dependencies are learnable without pauses at dependency edges in a variety of artificial language designs. Moreover, we find no evidence of a relationship between non-adjacent dependency learning and the robustness of adjacent statistics. We demonstrate that our two-step statistical learning model can account for all of our non-adjacent dependency learning results, and provides a unified learning account of adjacent and non-adjacent dependency learning. Finally, we discussed the theoretical implications of our findings for natural language acquisition, and argue that the dependency learning process can be a precursor to other language acquisition tasks that are vital to natural language acquisition.

Language acquisition is model-based rather than model-free.

Christiansen & Chater (C&C) propose that learning language is learning to process language. However, we believe that the general-purpose prediction mechanism they propose is insufficient to account for many phenomena in language acquisition. We argue from theoretical considerations and empirical evidence that many acquisition tasks are model-based, and that different acquisition tasks require different, specialized models.

Word categorization from distributional information: frames confer more than the sum of their (Bigram) parts.

Grammatical categories, such as noun and verb, are the building blocks of syntactic structure and the components that govern the grammatical patterns of language. However, in many languages words are not explicitly marked with their category information, hence a critical part of acquiring a language is categorizing the words. Computational analyses of child-directed speech have shown that distributional information-information about how words pattern with one another in sentences-could be a useful source of initial category information. Yet questions remain as to whether learners use this kind of information, and if so, what kinds of distributional patterns facilitate categorization. In this paper we investigated how adults exposed to an artificial language use distributional information to categorize words. We compared training situations in which target words occurred in frames (i.e., surrounded by two words that frequently co-occur) against situations in which target words occurred in simpler bigram contexts (where an immediately adjacent word provides the context for categorization). We found that learners categorized words together when they occurred in similar frame contexts, but not when they occurred in similar bigram contexts. These findings are particularly relevant because they accord with computational investigations showing that frame contexts provide accurate category information cross-linguistically. We discuss these findings in the context of prior research on distribution-based categorization and the broader implications for the role of distributional categorization in language acquisition.

Perceptual intake explains variability in statistical word segmentation.

One of the first problems in language learning is to segment words from continuous speech. Both prosodic and distributional information can be useful, and it is an important question how the two types of information are integrated. In this paper, we propose that the distinction between input (the statistical properties of the syllable sequence), and intake (how learners perceptually represent the syllable sequence) is a useful framework to integrate different sources of information. We took a novel approach, observing how a large number of syllable sequences were segmented. These sequences had the same transitional probability information for finding word boundaries but different syllables in them. We found large variability in the performance of the segmentation task, suggesting that factors other than the statistical properties of sequences were at play. This variability was explored using the input/intake asymmetry framework, which predicted that factors that shaped the representation of different syllable sequences could explain the variability of learning. We examined two factors, the saliency of the rhythm in these syllable sequences and how familiar the novel word forms in the sequence were to the existing lexicon. Both factors explained the variance in the learnability of different sequences, suggesting that processing of the sequences shaped learning. The implications of these results to computational models of statistical learning and broader implications to language learning were discussed.

Statistical word segmentation succeeds given the minimal amount of exposure.

One of the first tasks in language acquisition is word segmentation, a process to extract word forms from continuous speech streams. Statistical approaches to word segmentation have been shown to be a powerful mechanism, in which word boundaries are inferred from sequence statistics. This approach requires the learner to represent the frequency of units from syllable sequences, though accounts differ on how much statistical exposure is required. In this study, we examined the computational limit with which words can be extracted from continuous sequences. First, we discussed why two occurrences of a word in a continuous sequence is the computational lower limit for this word to be statistically defined. Next, we created short syllable sequences that contained certain words either two or four times. Learners were presented with these syllable sequences one at a time, immediately followed by a test of the novel words from these sequences. We found that, with the computationally minimal amount of two exposures, words were successfully segmented from continuous sequences. Moreover, longer syllable sequences providing four exposures to words generated more robust learning results. The implications of these results are discussed in terms of how learners segment and store the word candidates from continuous sequences.

A domain-agnostic approach for characterization of lifelong learning systems.

Despite the advancement of machine learning techniques in recent years, state-of-the-art systems lack robustness to "real world" events, where the input distributions and tasks encountered by the deployed systems will not be limited to the original training context, and systems will instead need to adapt to novel distributions and tasks while deployed. This critical gap may be addressed through the development of "Lifelong Learning" systems that are capable of (1) Continuous Learning, (2) Transfer and Adaptation, and (3) Scalability. Unfortunately, efforts to improve these capabilities are typically treated as distinct areas of research that are assessed independently, without regard to the impact of each separate capability on other aspects of the system. We instead propose a holistic approach, using a suite of metrics and an evaluation framework to assess Lifelong Learning in a principled way that is agnostic to specific domains or system techniques. Through five case studies, we show that this suite of metrics can inform the development of varied and complex Lifelong Learning systems. We highlight how the proposed suite of metrics quantifies performance trade-offs present during Lifelong Learning system development - both the widely discussed Stability-Plasticity dilemma and the newly proposed relationship between Sample Efficient and Robust Learning. Further, we make recommendations for the formulation and use of metrics to guide the continuing development of Lifelong Learning systems and assess their progress in the future.

Being suspicious of suspicious coincidences: The case of learning subordinate word meanings.

How do learners acquire subordinate terms (such as Dalmatian) and overcome the bias that words have basic-level meanings (such as dog)? Xu and Tenenbaum [Xu, F., & Tenenbaum, J. B. (2007a). Word learning as Bayesian inference. Psychological Review, 114(2), 245-272] found that both children and adults can learn the subordinate meaning of a novel word when it is used ostensively to label multiple exemplars: learners appear to reason about sampling statistics, detecting the suspicious coincidence that, e.g., a random sample of dogs all happen to be Dalmatians. Crucially though, their experimental support did not come from cross-situational ostensive labeling contexts, but from single instances that presented all exemplars at once and included a co-present test array that likely highlighted the relevant semantic contrast. Here we find that adults do not use suspicious coincidences during cross-situational word learning. We only find effects of suspicious coincidences in adults under specific testing conditions similar to those used by Xu and Tenenbaum. We find that adults show a basic-level meaning preference even after encountering five subordinate-level exemplars cross-situationally, even when the first three exemplars were presented simultaneously and labeled ostensively. Instead, participants arrived at subordinate meanings only within settings that highlighted the relevant semantic contrast, i.e., when the target words had referents that belonged to the same basic-level category (e.g., two words referring to dogs, with one referring to Dalmatians and the other to non-Dalmatian dogs). Our findings are consistent with a "semantic contrast" account of word learning, in which learners evaluate which semantic contrasts are relevant in the local learning context and use that information to constrain word meaning.

Object-Gaze Distance: Quantifying Near- Peripheral Gaze Behavior in Real-World Applications.

Eye tracking (ET) has shown to reveal the wearer's cognitive processes using the measurement of the central point of foveal vision. However, traditional ET evaluation methods have not been able to take into account the wearers' use of the peripheral field of vision. We propose an algorithmic enhancement to a state-of-the-art ET analysis method, the Object- Gaze Distance (OGD), which additionally allows the quantification of near-peripheral gaze behavior in complex real-world environments. The algorithm uses machine learning for area of interest (AOI) detection and computes the minimal 2D Euclidean pixel distance to the gaze point, creating a continuous gaze-based time-series. Based on an evaluation of two AOIs in a real surgical procedure, the results show that a considerable increase of interpretable fixation data from 23.8 % to 78.3 % of AOI screw and from 4.5 % to 67.2 % of AOI screwdriver was achieved, when incorporating the near-peripheral field of vision. Additionally, the evaluation of a multi-OGD time series representation has shown the potential to reveal novel gaze patterns, which may provide a more accurate depiction of human gaze behavior in multi-object environments.

Explicit and implicit memory representations in cross-situational word learning.

What kind of memory representations do word learners use when they learn the meaning of words cross-situationally? This study leverages the measure of the relationship between confidence and performance to explore the nature of memory representations in word learning. In the recognition memory literature, studies have shown that explicit memory can be used when subjects can semantically encode the study material. However, when the study material is chosen to be unverbalizable, implicit memory is used but is presumed to be only detectable under certain experimental conditions. In the current paper, five cross-situational word learning experiments manipulated the type of word referents with varying experimental paradigms that were designed to probe different types of memory under an implicit learning paradigm. When word referents were line drawings of familiar concepts, memory in cross situational learning was explicit. Implicit memory was found where referents were objects that cannot be encoded semantically (e.g., unverbalizable images). These findings have implications for different theoretical perspectives on early word learning, which differ in the extent to which existing semantic category information, as opposed to perceptual information, contributes to the word meaning process.

Top-down grouping affects adjacent dependency learning.

A large body of research has demonstrated that humans attend to adjacent co-occurrence statistics when processing sequential information, and bottom-up prosodic information can influence learning. In this study, we investigated how top-down grouping cues can influence statistical learning. Specifically, we presented English sentences that were structurally equivalent to each other, which induced top-down expectations of grouping in the artificial language sequences that immediately followed. We show that adjacent dependencies in the artificial language are learnable when these entrained boundaries bracket the adjacent dependencies into the same sub-sequence, but are not learnable when the elements cross an induced boundary, even though that boundary is not present in the bottom-up sensory input. We argue that when there is top-down bracketing information in the learning sequence, statistical learning takes place for elements bracketed within sub-sequences rather than all the elements in the continuous sequence. This limits the amount of linguistic computations that need to be performed, providing a domain over which statistical learning can operate.

Crossing the Cleft: Communication Challenges Between Neuroscience and Artificial Intelligence.

Historically, neuroscience principles have heavily influenced artificial intelligence (AI), for example the influence of the perceptron model, essentially a simple model of a biological neuron, on artificial neural networks. More recently, notable recent AI advances, for example the growing popularity of reinforcement learning, often appear more aligned with cognitive neuroscience or psychology, focusing on function at a relatively abstract level. At the same time, neuroscience stands poised to enter a new era of large-scale high-resolution data and appears more focused on underlying neural mechanisms or architectures that can, at times, seem rather removed from functional descriptions. While this might seem to foretell a new generation of AI approaches arising from a deeper exploration of neuroscience specifically for AI, the most direct path for achieving this is unclear. Here we discuss cultural differences between the two fields, including divergent priorities that should be considered when leveraging modern-day neuroscience for AI. For example, the two fields feed two very different applications that at times require potentially conflicting perspectives. We highlight small but significant cultural shifts that we feel would greatly facilitate increased synergy between the two fields.

Statistical Learning of Unfamiliar Sounds as Trajectories Through a Perceptual Similarity Space.

In typical statistical learning studies, researchers define sequences in terms of the probability of the next item in the sequence given the current item (or items), and they show that high probability sequences are treated as more familiar than low probability sequences. Existing accounts of these phenomena all assume that participants represent statistical regularities more or less as they are defined by the experimenters-as sequential probabilities of symbols in a string. Here we offer an alternative, or possibly supplementary, hypothesis. Specifically, rather than identifying or labeling individual stimuli discretely in order to predict the next item in a sequence, we need only assume that the participant is able to represent the stimuli as evincing particular similarity relations to one another, with sequences represented as trajectories through this similarity space. We present experiments in which this hypothesis makes sharply different predictions from hypotheses based on the assumption that sequences are learned over discrete, labeled stimuli. We also present a series of simulation models that encode stimuli as positions in a continuous two-dimensional space, and predict the next location from the current location. Although no model captures all of the data presented here, the results of three critical experiments are more consistent with the view that participants represent trajectories through similarity space rather than sequences of discrete labels under particular conditions.

The role of reference in cross-situational word learning.

Word learning involves massive ambiguity, since in a particular encounter with a novel word, there are an unlimited number of potential referents. One proposal for how learners surmount the problem of ambiguity is that learners use cross-situational statistics to constrain the ambiguity: When a word and its referent co-occur across multiple situations, learners will associate the word with the correct referent. Yu and Smith (2007) propose that these co-occurrence statistics are sufficient for word-to-referent mapping. Alternative accounts hold that co-occurrence statistics alone are insufficient to support learning, and that learners are further guided by knowledge that words are referential (e.g., Waxman & Gelman, 2009). However, no behavioral word learning studies we are aware of explicitly manipulate subjects' prior assumptions about the role of the words in the experiments in order to test the influence of these assumptions. In this study, we directly test whether, when faced with referential ambiguity, co-occurrence statistics are sufficient for word-to-referent mappings in adult word-learners. Across a series of cross-situational learning experiments, we varied the degree to which there was support for the notion that the words were referential. At the same time, the statistical information about the words' meanings was held constant. When we overrode support for the notion that words were referential, subjects failed to learn the word-to-referent mappings, but otherwise they succeeded. Thus, cross-situational statistics were useful only when learners had the goal of discovering mappings between words and referents. We discuss the implications of these results for theories of word learning in children's language acquisition.

Learning nonadjacent dependencies embedded in sentences of an artificial language: When learning breaks down.

The structure of natural languages give rise to many dependencies in the linear sequences of words, and within words themselves. Detecting these dependencies is arguably critical for young children in learning the underlying structure of their language. There is considerable evidence that human adults and infants are sensitive to the statistical properties of sequentially adjacent items. However, the conditions under which learners detect nonadjacent dependencies (NADs) appears to be much more limited. This has resulted in proposals that the kinds of learning mechanisms learners deploy in processing adjacent dependencies are fundamentally different from those deployed in learning NADs. Here we challenge this view. In 4 experiments, we show that learning both kinds of dependencies is hindered in conditions when they are embedded in longer sequences of words, and facilitated when they are isolated by silences. We argue that the findings from the present study and prior research is consistent with a theory that similar mechanisms are deployed for adjacent and nonadjacent dependency learning, but that NAD learning is simply computationally more complex. Hence, in some situations NAD learning is only successful when constraining information is provided, but critically, that additional information benefits adjacent dependency learning in similar ways. (PsycINFO Database Record

Top-down structure influences learning of nonadjacent dependencies in an artificial language.

Because of the hierarchical organization of natural languages, words that are syntactically related are not always linearly adjacent. For example, the subject and verb in the child always runs agree in person and number, although they are not adjacent in the sequences of words. Since such dependencies are indicative of abstract linguist structure, it is of significant theoretical interest how these relationships are acquired by language learners. Most experiments that investigate nonadjacent dependency (NAD) learning have used artificial languages in which the to-be-learned dependencies are isolated, by presenting the minimal sequences that contain the dependent elements. However, dependencies in natural language are not typically isolated in this way. We report the first demonstration to our knowledge of successful learning of embedded NADs, in which silences do not mark dependency boundaries. Subjects heard passages of English with a predictable structure, interspersed with passages of the artificial language. The English sentences were designed to induce boundaries in the artificial languages. In Experiment 1 & 3 the artificial NADs were contained within the induced boundaries and subjects learned them, whereas in Experiment 2 & 4, the NADs crossed the induced boundaries and subjects did not learn them. We take this as evidence that sentential structure was "carried over" from the English sentences and used to organize the artificial language. This approach provides several new insights into the basic mechanisms of NAD learning in particular and statistical learning in general. (PsycINFO Database Record