Statistical computations over a speech stream in a rodent

Functional reorganization of brain regions supporting artificial grammar learning across the first half year of life

Pre-babbling infants can track nonadjacent dependencies (NADs) in the auditory domain. While this forms a crucial prerequisite for language acquisition, the neurodevelopmental origins of this ability remain unknown. We applied functional near-infrared spectroscopy in neonates and 6- to 7-month-old infants to investigate the neural substrate supporting NAD learning and detection using tone sequences in an artificial grammar learning paradigm. Detection of NADs was indicated by left prefrontal activation in neonates while by left supramarginal gyrus (SMG), superior temporal gyrus (STG), and inferior frontal gyrus activation in 6- to 7-month-olds. Functional connectivity analyses further indicated that the neonate activation pattern during the test phase benefited from a brain network consisting of prefrontal regions, left SMG and STG during the rest and learning phases. These findings suggest a left-hemispheric learning-related functional brain network may emerge at birth and serve as the foundation for the later engagement of these regions for NAD detection, thus, providing a neural basis for language acquisition.

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.

Long-Horizon Associative Learning Explains Human Sensitivity to Statistical and Network Structures in Auditory Sequences

Networks are a useful mathematical tool for capturing the complexity of the world. In a previous behavioral study, we showed that human adults were sensitive to the high-level network structure underlying auditory sequences, even when presented with incomplete information. Their performance was best explained by a mathematical model compatible with associative learning principles, based on the integration of the transition probabilities between adjacent and nonadjacent elements with a memory decay. In the present study, we explored the neural correlates of this hypothesis via magnetoencephalography (MEG). Participants (N = 23, 16 females) passively listened to sequences of tones organized in a sparse community network structure comprising two communities. An early difference (∼150 ms) was observed in the brain responses to tone transitions with similar transition probability but occurring either within or between communities. This result implies a rapid and automatic encoding of the sequence structure. Using time-resolved decoding, we estimated the duration and overlap of the representation of each tone. The decoding performance exhibited exponential decay, resulting in a significant overlap between the representations of successive tones. Based on this extended decay profile, we estimated a long-horizon associative learning novelty index for each transition and found a correlation of this measure with the MEG signal. Overall, our study sheds light on the neural mechanisms underlying human sensitivity to network structures and highlights the potential role of Hebbian-like mechanisms in supporting learning at various temporal scales.

Limited but specific engagement of the mature language network during linguistic statistical learning

Statistical learning (SL) is the ability to detect and learn regularities from input and is foundational to language acquisition. Despite the dominant role of SL as a theoretical construct for language development, there is a lack of direct evidence supporting the shared neural substrates underlying language processing and SL. It is also not clear whether the similarities, if any, are related to linguistic processing, or statistical regularities in general. The current study tests whether the brain regions involved in natural language processing are similarly recruited during auditory, linguistic SL. Twenty-two adults performed an auditory linguistic SL task, an auditory nonlinguistic SL task, and a passive story listening task as their neural activation was monitored. Within the language network, the left posterior temporal gyrus showed sensitivity to embedded speech regularities during auditory, linguistic SL, but not auditory, nonlinguistic SL. Using a multivoxel pattern similarity analysis, we uncovered similarities between the neural representation of auditory, linguistic SL, and language processing within the left posterior temporal gyrus. No other brain regions showed similarities between linguistic SL and language comprehension, suggesting that a shared neurocomputational process for auditory SL and natural language processing within the left posterior temporal gyrus is specific to linguistic stimuli.

Abstract processing of syllabic structures in early infancy

Syllables are one of the fundamental building blocks of early language acquisition. From birth onwards, infants preferentially segment, process and represent the speech into syllable-sized units, raising the question of what type of computations infants are able to perform on these perceptual units. Syllables are abstract units structured in a way that allows grouping phonemes into sequences. The goal of this research was to investigate 4-to-5-month-old infants' ability to encode the internal structure of syllables, at a target age when the language system is not yet specialized on the sounds and the phonotactics of native languages. We conducted two experiments in which infants were first familiarized to lists of syllables implementing either CVC (consonant-vowel-consonant) or CCV (consonant-consonant-vowel) structures, then presented with new syllables implementing both structures at test. Experiments differ in the degree of phonological similarity between the materials used at familiarization and test. Results show that infants were able to differentiate syllabic structures at test, even when test syllables were implemented by combinations of phonemes that infants did not hear before. Only infants familiarized with CVC syllables discriminated the structures at test, pointing to a processing advantage for CVC over CCV structures. This research shows that, in addition to preferentially processing the speech into syllable-sized units, during the first months of life, infants are also capable of performing fine-grained computations within such units.

Brain-imaging evidence for compression of binary sound sequences in human memory

According to the language-of-thought hypothesis, regular sequences are compressed in human memory using recursive loops akin to a mental program that predicts future items. We tested this theory by probing memory for 16-item sequences made of two sounds. We recorded brain activity with functional MRI and magneto-encephalography (MEG) while participants listened to a hierarchy of sequences of variable complexity, whose minimal description required transition probabilities, chunking, or nested structures. Occasional deviant sounds probed the participants' knowledge of the sequence. We predicted that task difficulty and brain activity would be proportional to the complexity derived from the minimal description length in our formal language. Furthermore, activity should increase with complexity for learned sequences, and decrease with complexity for deviants. These predictions were upheld in both fMRI and MEG, indicating that sequence predictions are highly dependent on sequence structure and become weaker and delayed as complexity increases. The proposed language recruited bilateral superior temporal, precentral, anterior intraparietal, and cerebellar cortices. These regions overlapped extensively with a localizer for mathematical calculation, and much less with spoken or written language processing. We propose that these areas collectively encode regular sequences as repetitions with variations and their recursive composition into nested structures.

Sensitivity to the sonority sequencing principle in rats (Rattus norvegicus)

Albeit diverse, human languages exhibit universal structures. A salient example is the syllable, an important structure of language acquisition. The structure of syllables is determined by the Sonority Sequencing Principle (SSP), a linguistic constraint according to which phoneme intensity must increase at onset, reaching a peak at nucleus (vowel), and decline at offset. Such structure generates an intensity pattern with an arch shape. In humans, sensitivity to restrictions imposed by the SSP on syllables appears at birth, raising questions about its emergence. We investigated the biological mechanisms at the foundations of the SSP, testing a nonhuman, non-vocal-learner species with the same language materials used with humans. Rats discriminated well-structured syllables (e.g., pras) from ill-structured ones (e.g., lbug) after being familiarized with syllabic structures conforming to the SSP. In contrast, we did not observe evidence that rats familiarized with syllables that violate such constraint discriminated at test. This research provides the first evidence of sensitivity to the SSP in a nonhuman species, which likely stems from evolutionary-ancient cross-species biological predispositions for natural acoustic patterns. Humans' early sensitivity to the SSP possibly emerges from general auditory processing that favors sounds depicting an arch-shaped envelope, common amongst animal vocalizations. Ancient sensory mechanisms, responsible for processing vocalizations in the wild, would constitute an entry-gate for human language acquisition.

Dynamic Motion and Human Agents Facilitate Visual Nonadjacent Dependency Learning

Many events that humans and other species experience contain regularities in which certain elements within an event predict certain others. While some of these regularities involve tracking the co-occurrences between temporally adjacent stimuli, others involve tracking the co-occurrences between temporally distant stimuli (i.e., nonadjacent dependencies, NADs). Prior research shows robust learning of adjacent dependencies in humans and other species, whereas learning NADs is more difficult, and often requires support from properties of the stimulus to help learners notice the NADs. Here, we report on seven experiments that examined NAD learning from various types of visual stimuli, exploring the effects of dynamic motion on adults' NAD learning from visual sequences involving human and nonhuman agents. We tested adults' NAD learning from visual sequences of human actions, object transformations, static images of human postures, and static images of an object in different postures. We found that dynamic motion aids the acquisition of NADs. We also found that learning NADs in sequences involving human agents is more robust compared to sequences involving nonhuman objects. We propose that dynamic motion and human agents both independently result in richer representations that provide a stronger signal for NAD learning.

Tracking transitional probabilities and segmenting auditory sequences are dissociable processes in adults and neonates

Since speech is a continuous stream with no systematic boundaries between words, how do pre-verbal infants manage to discover words? A proposed solution is that they might use the transitional probability between adjacent syllables, which drops at word boundaries. Here, we tested the limits of this mechanism by increasing the size of the word-unit to four syllables, and its automaticity by testing asleep neonates. Using markers of statistical learning in neonates' EEG, compared to adult behavioral performances in the same task, we confirmed that statistical learning is automatic enough to be efficient even in sleeping neonates. We also revealed that: (1) Successfully tracking transition probabilities (TP) in a sequence is not sufficient to segment it. (2) Prosodic cues, as subtle as subliminal pauses, enable to recover words segmenting capacities. (3) Adults' and neonates' capacities to segment streams seem remarkably similar despite the difference of maturation and expertise. Finally, we observed that learning increased the overall similarity of neural responses across infants during exposure to the stream, providing a novel neural marker to monitor learning. Thus, from birth, infants are equipped with adult-like tools, allowing them to extract small coherent word-like units from auditory streams, based on the combination of statistical analyses and auditory parsing cues. RESEARCH HIGHLIGHTS: Successfully tracking transitional probabilities in a sequence is not always sufficient to segment it. Word segmentation solely based on transitional probability is limited to bi- or tri-syllabic elements. Prosodic cues, as subtle as subliminal pauses, enable to recover chunking capacities in sleeping neonates and awake adults for quadriplets.

Differential activation of a frontoparietal network explains population-level differences in statistical learning from speech

People of all ages display the ability to detect and learn from patterns in seemingly random stimuli. Referred to as statistical learning (SL), this process is particularly critical when learning a spoken language, helping in the identification of discrete words within a spoken phrase. Here, by considering individual differences in speech auditory–motor synchronization, we demonstrate that recruitment of a specific neural network supports behavioral differences in SL from speech. While independent component analysis (ICA) of fMRI data revealed that a network of auditory and superior pre/motor regions is universally activated in the process of learning, a frontoparietal network is additionally and selectively engaged by only some individuals (high auditory–motor synchronizers). Importantly, activation of this frontoparietal network is related to a boost in learning performance, and interference with this network via articulatory suppression (AS; i.e., producing irrelevant speech during learning) normalizes performance across the entire sample. Our work provides novel insights on SL from speech and reconciles previous contrasting findings. These findings also highlight a more general need to factor in fundamental individual differences for a precise characterization of cognitive phenomena.

In the context of speech, statistical learning is thought to be an important mechanism for language acquisition. This study shows that language statistical learning is boosted by the recruitment of a fronto-parietal brain network related to auditory-motor synchronization and its interplay with a mandatory auditory-motor learning system.

Statistical Learning in Vision

Vision and learning have long been considered to be two areas of research linked only distantly. However, recent developments in vision research have changed the conceptual definition of vision from a signal-evaluating process to a goal-oriented interpreting process, and this shift binds learning, together with the resulting internal representations, intimately to vision. In this review, we consider various types of learning (perceptual, statistical, and rule/abstract) associated with vision in the past decades and argue that they represent differently specialized versions of the fundamental learning process, which must be captured in its entirety when applied to complex visual processes. We show why the generalized version of statistical learning can provide the appropriate setup for such a unified treatment of learning in vision, what computational framework best accommodates this kind of statistical learning, and what plausible neural scheme could feasibly implement this framework. Finally, we list the challenges that the field of statistical learning faces in fulfilling the promise of being the right vehicle for advancing our understanding of vision in its entirety. Expected final online publication date for the Annual Review of Vision Science, Volume 8 is September 2022. Please see http://www.annualreviews.org/page/journal/pubdates for revised estimates.

Sleeping neonates track transitional probabilities in speech but only retain the first syllable of words

Extracting statistical regularities from the environment is a primary learning mechanism that might support language acquisition. While it has been shown that infants are sensitive to transition probabilities between syllables in speech, it is still not known what information they encode. Here we used electrophysiology to study how full-term neonates process an artificial language constructed by randomly concatenating four pseudo-words and what information they retain after a few minutes of exposure. Neural entrainment served as a marker of the regularities the brain was tracking during learning. Then in a post-learning phase, evoked-related potentials (ERP) to different triplets explored which information was retained. After two minutes of familiarization with the artificial language, neural entrainment at the word rate emerged, demonstrating rapid learning of the regularities. ERPs in the test phase significantly differed between triplets starting or not with the correct first syllables, but no difference was associated with subsequent violations in transition probabilities. Thus, our results revealed a two-step learning process: neonates segmented the stream based on its statistical regularities, but memory encoding targeted during the word recognition phase entangled the ordinal position of the syllables but was still incomplete at that age.

Speech naturalness detection and language representation in the dog brain

Family dogs are exposed to a continuous flow of human speech throughout their lives. However, the extent of their abilities in speech perception is unknown. Here, we used functional magnetic resonance imaging (fMRI) to test speech detection and language representation in the dog brain. Dogs (n = 18) listened to natural speech and scrambled speech in a familiar and an unfamiliar language. Speech scrambling distorts auditory regularities specific to speech and to a given language, but keeps spectral voice cues intact. We hypothesized that if dogs can extract auditory regularities of speech, and of a familiar language, then there will be distinct patterns of brain activity for natural speech vs. scrambled speech, and also for familiar vs. unfamiliar language. Using multivoxel pattern analysis (MVPA) we found that bilateral auditory cortical regions represented natural speech and scrambled speech differently; with a better classifier performance in longer-headed dogs in a right auditory region. This neural capacity for speech detection was not based on preferential processing for speech but rather on sensitivity to sound naturalness. Furthermore, in case of natural speech, distinct activity patterns were found for the two languages in the secondary auditory cortex and in the precruciate gyrus; with a greater difference in responses to the familiar and unfamiliar languages in older dogs, indicating a role for the amount of language exposure. No regions represented differently the scrambled versions of the two languages, suggesting that the activity difference between languages in natural speech reflected sensitivity to language-specific regularities rather than to spectral voice cues. These findings suggest that separate cortical regions support speech naturalness detection and language representation in the dog brain.

Is there such a thing as a 'good statistical learner'?

A growing body of research investigates individual differences in the learning of statistical structure, tying them to variability in cognitive (dis)abilities. This approach views statistical learning (SL) as a general individual ability that underlies performance across a range of cognitive domains. But is there a general SL capacity that can sort individuals from 'bad' to 'good' statistical learners? Explicating the suppositions underlying this approach, we suggest that current evidence supporting it is meager. We outline an alternative perspective that considers the variability of statistical environments within different cognitive domains. Once we focus on learning that is tuned to the statistics of real-world sensory inputs, an alternative view of SL computations emerges with a radically different outlook for SL research.

Neural processes underlying statistical learning for speech segmentation in dogs

To learn words, humans extract statistical regularities from speech. Multiple species use statistical learning also to process speech, but the neural underpinnings of speech segmentation in non-humans remain largely unknown. Here, we investigated computational and neural markers of speech segmentation in dogs, a phylogenetically distant mammal that efficiently navigates humans' social and linguistic environment. Using electroencephalography (EEG), we compared event-related responses (ERPs) for artificial words previously presented in a continuous speech stream with different distributional statistics. Results revealed an early effect (220-470 ms) of transitional probability and a late component (590-790 ms) modulated by both word frequency and transitional probability. Using fMRI, we searched for brain regions sensitive to statistical regularities in speech. Structured speech elicited lower activity in the basal ganglia, a region involved in sequence learning, and repetition enhancement in the auditory cortex. Speech segmentation in dogs, similar to that of humans, involves complex computations, engaging both domain-general and modality-specific brain areas. VIDEO ABSTRACT.

Language statistical learning responds to reinforcement learning principles rooted in the striatum

Statistical learning (SL) is the ability to extract regularities from the environment. In the domain of language, this ability is fundamental in the learning of words and structural rules. In lack of reliable online measures, statistical word and rule learning have been primarily investigated using offline (post-familiarization) tests, which gives limited insights into the dynamics of SL and its neural basis. Here, we capitalize on a novel task that tracks the online SL of simple syntactic structures combined with computational modeling to show that online SL responds to reinforcement learning principles rooted in striatal function. Specifically, we demonstrate—on 2 different cohorts—that a temporal difference model, which relies on prediction errors, accounts for participants’ online learning behavior. We then show that the trial-by-trial development of predictions through learning strongly correlates with activity in both ventral and dorsal striatum. Our results thus provide a detailed mechanistic account of language-related SL and an explanation for the oft-cited implication of the striatum in SL tasks. This work, therefore, bridges the long-standing gap between language learning and reinforcement learning phenomena.

Statistical learning is the ability to extract regularities from the environment; in the domain of language, this ability is fundamental in the learning of words and structural rules. This study uses a combination of computational modelling and functional MRI to reveal a fundamental link between online language statistical learning and reinforcement learning at the algorithmic and implementational levels.

Auditory Cortical Changes Precede Brainstem Changes During Rapid Implicit Learning: Evidence From Human EEG

The auditory system is sensitive to stimulus regularities such as frequently occurring sounds and sound combinations. Evidence of regularity detection can be seen in how neurons across the auditory network, from brainstem to cortex, respond to the statistical properties of the soundscape, and in the rapid learning of recurring patterns in their environment by children and adults. Although rapid auditory learning is presumed to involve functional changes to the auditory network, the chronology and directionality of changes are not well understood. To study the mechanisms by which this learning occurs, auditory brainstem and cortical activity was simultaneously recorded via electroencephalogram (EEG) while young adults listened to novel sound streams containing recurring patterns. Neurophysiological responses were compared between easier and harder learning conditions. Collectively, the behavioral and neurophysiological findings suggest that cortical and subcortical structures each provide distinct contributions to auditory pattern learning, but that cortical sensitivity to stimulus patterns likely precedes subcortical sensitivity.

Sustained Pupil Responses Are Modulated by Predictability of Auditory Sequences

The brain is highly sensitive to auditory regularities and exploits the predictable order of sounds in many situations, from parsing complex auditory scenes, to the acquisition of language. To understand the impact of stimulus predictability on perception, it is important to determine how the detection of predictable structure influences processing and attention. Here, we use pupillometry to gain insight into the effect of sensory regularity on arousal. Pupillometry is a commonly used measure of salience and processing effort, with more perceptually salient or perceptually demanding stimuli consistently associated with larger pupil diameters. In two experiments we tracked human listeners' pupil dynamics while they listened to sequences of 50-ms tone pips of different frequencies. The order of the tone pips was either random, contained deterministic (fully predictable) regularities (experiment 1, n = 18, 11 female) or had a probabilistic regularity structure (experiment 2, n = 20, 17 female). The sequences were rapid, preventing conscious tracking of sequence structure thus allowing us to focus on the automatic extraction of different types of regularities. We hypothesized that if regularity facilitates processing by reducing processing demands, a smaller pupil diameter would be seen in response to regular relative to random patterns. Conversely, if regularity is associated with heightened arousal and attention (i.e., engages processing resources) the opposite pattern would be expected. In both experiments we observed a smaller sustained (tonic) pupil diameter for regular compared with random sequences, consistent with the former hypothesis and confirming that predictability facilitates sequence processing.SIGNIFICANCE STATEMENT The brain is highly sensitive to auditory regularities. To appreciate the impact that the presence of predictability has on perception, we need to better understand how a predictable structure influences processing and attention. We recorded listeners' pupil responses to sequences of tones that followed either a predictable or unpredictable pattern, as the pupil can be used to implicitly tap into these different cognitive processes. We found that the pupil showed a smaller sustained diameter to predictable sequences, indicating that predictability eased processing rather than boosted attention. The findings suggest that the pupil response can be used to study the automatic extraction of regularities, and that the effects are most consistent with predictability helping the listener to efficiently process upcoming sounds.

Learning non-adjacent rules and non-adjacent dependencies from human actions in 9-month-old infants

Seven month old infants can learn simple repetition patterns, such as we-fo-we, and generalize the rules to sequences of new syllables, such as ga-ti-ga. However, repetition rule learning in visual sequences seems more challenging, leading some researchers to claim that this type of rule learning applies preferentially to communicative stimuli. Here we demonstrate that 9-month-old infants can learn repetition rules in sequences of non-communicative dynamic human actions. We also show that when primed with these non-adjacent repetition patterns, infants can learn non-adjacent dependencies that involve memorizing the dependencies between specific human actions-patterns that prior research has shown to be difficult for infants in the visual domain and in speech. We discuss several possible mechanisms that account for the apparent advantage stimuli involving human action sequences has over other kinds of stimuli in supporting non-adjacent dependency learning. We also discuss possible implications for theories of language acquisition.

Evidence of ordinal position encoding of sequences extracted from continuous speech

Infants' capacity to extract statistical regularities from sequential information is impressive and well documented. However, statistical learning's underlying mechanism remains mostly unknown, and its role in language acquisition is still under debate. To shed light on these issues, here we address the question of which information human subjects extract and encode after familiarisation with a continuous sequence of stimuli and its dependence on the type of segmentation cues and on the stimuli modality. Specifically, we investigate whether adults and 5-month-old infants learn the syllables' co-occurrence in the stream or generate a representation of the Words that include syllables' ordinal position. We test if subtle pauses signalling word boundaries change the encoding and, in adults, if it varies across modalities. In six behavioural experiments, we show that: (i) Adults and infants learn the streams' statistical structure. (ii) Ordinal encoding emerges in the auditory modality, and pauses enhanced it. However, (iii) ordinal encoding seems to depend on the learning stage and not on pauses marking Words' edges. Interestingly, (iv) for visual presentation of orthographic syllables, we do not find evidence of ordinal encoding in adults. Our results support the emergence, in the auditory modality, of a Word representation where its constituents are associated with an ordinal position at play already early in life, bringing new insights into speech processing and language acquisition. Additionally, we successfully use for the first time pupillometry in an infant segmentation task.

When forgetting fosters learning: A neural network model for statistical learning

Learning often requires splitting continuous signals into recurring units, such as the discrete words constituting fluent speech; these units then need to be encoded in memory. A prominent candidate mechanism involves statistical learning of co-occurrence statistics like transitional probabilities (TPs), reflecting the idea that items from the same unit (e.g., syllables within a word) predict each other better than items from different units. TP computations are surprisingly flexible and sophisticated. Humans are sensitive to forward and backward TPs, compute TPs between adjacent items and longer-distance items, and even recognize TPs in novel units. We explain these hallmarks of statistical learning with a simple model with tunable, Hebbian excitatory connections and inhibitory interactions controlling the overall activation. With weak forgetting, activations are long-lasting, yielding associations among all items; with strong forgetting, no associations ensue as activations do not outlast stimuli; with intermediate forgetting, the network reproduces the hallmarks above. Forgetting thus is a key determinant of these sophisticated learning abilities. Further, in line with earlier dissociations between statistical learning and memory encoding, our model reproduces the hallmarks of statistical learning in the absence of a memory store in which items could be placed.

Statistical learning of unbalanced exclusive-or temporal sequences in humans

A pervasive issue in statistical learning has been to determine the parameters of regularity extraction. Our hypothesis was that the extraction of transitional probabilities can prevail over frequency if the task involves prediction. Participants were exposed to four repeated sequences of three stimuli (XYZ) with each stimulus corresponding to the position of a red dot on a touch screen that participants were required to touch sequentially. The temporal and spatial structure of the positions corresponded to a serial version of the exclusive-or (XOR) that allowed testing of the respective effect of frequency and first- and second-order transitional probabilities. The XOR allowed the first-order transitional probability to vary while being not completely related to frequency and to vary while the second-order transitional probability was fixed (p(Z|X, Y) = 1). The findings show that first-order transitional probability prevails over frequency to predict the second stimulus from the first and that it also influences the prediction of the third item despite the presence of second-order transitional probability that could have offered a certain prediction of the third item. These results are particularly informative in light of statistical learning models.

Arc-shaped pitch contours facilitate item recognition in non-human animals

Acoustic changes linked to natural prosody are a key source of information about the organization of language. Both human infants and adults readily take advantage of such changes to discover and memorize linguistic patterns. Do they so because our brain is efficiently wired to specifically process linguistic stimuli? Or are we co-opting for language acquisition purposes more general principles that might be inherited from our animal ancestors? Here, we address this question by exploring if other species profit from prosody to better process acoustic sequences. More specifically, we test whether arc-shaped pitch contours defining natural prosody might facilitate item recognition and memorization in rats. In two experiments, we presented to the rats nonsense words with flat, natural, inverted and random prosodic contours. We observed that the animals correctly recognized the familiarization words only when arc-shaped pitch contours were implemented over them. Our results suggest that other species might also benefit from prosody for the memorization of items in a sequence. Such capacity seems to be rooted in general principles of how biological sounds are produced and processed.

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.

Different mechanisms underlie implicit visual statistical learning in honey bees and humans

The ability of developing complex internal representations of the environment is considered a crucial antecedent to the emergence of humans' higher cognitive functions. Yet it is an open question whether there is any fundamental difference in how humans and other good visual learner species naturally encode aspects of novel visual scenes. Using the same modified visual statistical learning paradigm and multielement stimuli, we investigated how human adults and honey bees (Apis mellifera) encode spontaneously, without dedicated training, various statistical properties of novel visual scenes. We found that, similarly to humans, honey bees automatically develop a complex internal representation of their visual environment that evolves with accumulation of new evidence even without a targeted reinforcement. In particular, with more experience, they shift from being sensitive to statistics of only elemental features of the scenes to relying on co-occurrence frequencies of elements while losing their sensitivity to elemental frequencies, but they never encode automatically the predictivity of elements. In contrast, humans involuntarily develop an internal representation that includes single-element and co-occurrence statistics, as well as information about the predictivity between elements. Importantly, capturing human visual learning results requires a probabilistic chunk-learning model, whereas a simple fragment-based memory-trace model that counts occurrence summary statistics is sufficient to replicate honey bees' learning behavior. Thus, humans' sophisticated encoding of sensory stimuli that provides intrinsic sensitivity to predictive information might be one of the fundamental prerequisites of developing higher cognitive abilities.

Statistical learning and memory

Learners often need to identify and remember recurring units in continuous sequences, but the underlying mechanisms are debated. A particularly prominent candidate mechanism relies on distributional statistics such as Transitional Probabilities (TPs). However, it is unclear what the outputs of statistical segmentation mechanisms are, and if learners store these outputs as discrete chunks in memory. We critically review the evidence for the possibility that statistically coherent items are stored in memory and outline difficulties in interpreting past research. We use Slone and Johnson's (2018) experiments as a case study to show that it is difficult to delineate the different mechanisms learners might use to solve a learning problem. Slone and Johnson (2018) reported that 8-month-old infants learned coherent chunks of shapes in visual sequences. Here, we describe an alternate interpretation of their findings based on a multiple-cue integration perspective. First, when multiple cues to statistical structure were available, infants' looking behavior seemed to track with the strength of the strongest one - backward TPs, suggesting that infants process multiple cues simultaneously and select the strongest one. Second, like adults, infants are exquisitely sensitive to chunks, but may require multiple cues to extract them. In Slone and Johnson's (2018) experiments, these cues were provided by immediate chunk repetitions during familiarization. Accordingly, infants showed strongest evidence of chunking following familiarization sequences in which immediate repetitions were more frequent. These interpretations provide a strong argument for infants' processing of multiple cues and the potential importance of multiple cues for chunk recognition in infancy.

Non-adjacent Dependency Learning in Humans and Other Animals

Learning and processing natural language requires the ability to track syntactic relationships between words and phrases in a sentence, which are often separated by intervening material. These nonadjacent dependencies can be studied using artificial grammar learning paradigms and structured sequence processing tasks. These approaches have been used to demonstrate that human adults, infants and some nonhuman animals are able to detect and learn dependencies between nonadjacent elements within a sequence. However, learning nonadjacent dependencies appears to be more cognitively demanding than detecting dependencies between adjacent elements, and only occurs in certain circumstances. In this review, we discuss different types of nonadjacent dependencies in language and in artificial grammar learning experiments, and how these differences might impact learning. We summarize different types of perceptual cues that facilitate learning, by highlighting the relationship between dependent elements bringing them closer together either physically, attentionally, or perceptually. Finally, we review artificial grammar learning experiments in human adults, infants, and nonhuman animals, and discuss how similarities and differences observed across these groups can provide insights into how language is learned across development and how these language-related abilities might have evolved.

Structured Sequence Learning: Animal Abilities, Cognitive Operations, and Language Evolution

Human language is a salient example of a neurocognitive system that is specialized to process complex dependencies between sensory events distributed in time, yet how this system evolved and specialized remains unclear. Artificial Grammar Learning (AGL) studies have generated a wealth of insights into how human adults and infants process different types of sequencing dependencies of varying complexity. The AGL paradigm has also been adopted to examine the sequence processing abilities of nonhuman animals. We critically evaluate this growing literature in species ranging from mammals (primates and rats) to birds (pigeons, songbirds, and parrots) considering also cross-species comparisons. The findings are contrasted with seminal studies in human infants that motivated the work in nonhuman animals. This synopsis identifies advances in knowledge and where uncertainty remains regarding the various strategies that nonhuman animals can adopt for processing sequencing dependencies. The paucity of evidence in the few species studied to date and the need for follow-up experiments indicate that we do not yet understand the limits of animal sequence processing capacities and thereby the evolutionary pattern. This vibrant, yet still budding, field of research carries substantial promise for advancing knowledge on animal abilities, cognitive substrates, and language evolution.

Statistical learning for vocal sequence acquisition in a songbird

Birdsong is a learned communicative behavior that consists of discrete acoustic elements (“syllables”) that are sequenced in a controlled manner. While the learning of the acoustic structure of syllables has been extensively studied, relatively little is known about sequence learning in songbirds. Statistical learning could contribute to the acquisition of vocal sequences, and we investigated the nature and extent of sequence learning at various levels of song organization in the Bengalese finch, Lonchura striata var. domestica. We found that, under semi-natural conditions, pupils (sons) significantly reproduced the sequence statistics of their tutor’s (father’s) songs at multiple levels of organization (e.g., syllable repertoire, prevalence, and transitions). For example, the probability of syllable transitions at “branch points” (relatively complex sequences that are followed by multiple types of transitions) were significantly correlated between the songs of tutors and pupils. We confirmed the contribution of learning to sequence similarities between fathers and sons by experimentally tutoring juvenile Bengalese finches with the songs of unrelated tutors. We also discovered that the extent and fidelity of sequence similarities between tutors and pupils were significantly predicted by the prevalence of sequences in the tutor’s song and that distinct types of sequence modifications (e.g., syllable additions or deletions) followed distinct patterns. Taken together, these data provide compelling support for the role of statistical learning in vocal production learning and identify factors that could modulate the extent of vocal sequence learning.

A common probabilistic framework for perceptual and statistical learning

System-level learning of sensory information is traditionally divided into two domains: perceptual learning that focuses on acquiring knowledge suitable for fine discrimination between similar sensory inputs, and statistical learning that explores the mechanisms that develop complex representations of unfamiliar sensory experiences. The two domains have been typically treated in complete separation both in terms of the underlying computational mechanisms and the brain areas and processes implementing those computations. However, a number of recent findings in both domains call in question this strict separation. We interpret classical and more recent results in the general framework of probabilistic computation, provide a unifying view of how various aspects of the two domains are interlinked, and suggest how the probabilistic approach can also alleviate the problem of dealing with widely different types of neural correlates of learning. Finally, we outline several directions along which our proposed approach fosters new types of experiments that can promote investigations of natural learning in humans and other species.

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.

Probability Learning in an Uncertain World: How Children Adjust to Changing Contingencies

We regularly make predictions about future events, even in a world where events occur probabilistically rather than deterministically. Our environment may even be non-stationary such that the probability of an event may change suddenly or from one context to another. 4-6 year olds and adults viewed 3 boxes and guessed the location of a hidden toy. After 80 trials with one set of probabilities assigned to the 3 boxes, the spatial distribution of these probabilities was altered. Adults easily responded to this change, with participants who maximized in the first half (by choosing the most common location at a higher rate than it was presented) being the fastest at making this shift. Only the older children successfully switched to the new location, with younger children either partially switching, perseverating on their original strategy, or failing to learn the first distribution, suggesting a fundamental development in children's response to changing probabilities.

Newborns are sensitive to multiple cues for word segmentation in continuous speech

Before infants can learn words, they must identify those words in continuous speech. Yet, the speech signal lacks obvious boundary markers, which poses a potential problem for language acquisition (Swingley, Philos Trans R Soc Lond. Series B, Biol Sci 364(1536), 3617-3632, 2009). By the middle of the first year, infants seem to have solved this problem (Bergelson & Swingley, Proc Natl Acad Sci 109(9), 3253-3258, 2012; Jusczyk & Aslin, Cogn Psychol 29, 1-23, 1995), but it is unknown if segmentation abilities are present from birth, or if they only emerge after sufficient language exposure and/or brain maturation. Here, in two independent experiments, we looked at two cues known to be crucial for the segmentation of human speech: the computation of statistical co-occurrences between syllables and the use of the language's prosody. After a brief familiarization of about 3 min with continuous speech, using functional near-infrared spectroscopy, neonates showed differential brain responses on a recognition test to words that violated either the statistical (Experiment 1) or prosodic (Experiment 2) boundaries of the familiarization, compared to words that conformed to those boundaries. Importantly, word recognition in Experiment 2 occurred even in the absence of prosodic information at test, meaning that newborns encoded the phonological content independently of its prosody. These data indicate that humans are born with operational language processing and memory capacities and can use at least two types of cues to segment otherwise continuous speech, a key first step in language acquisition.

Superior learning in synesthetes: Consistent grapheme-color associations facilitate statistical learning

In synesthesia activation in one sensory domain, such as smell or sound, triggers an involuntary and unusual secondary sensory or cognitive experience. In the present study, we ask whether the added sensory experience of synesthesia can aid statistical learning-the ability to track environmental regularities in order to segment continuous information. To investigate this, we measured statistical learning outcomes, using an aurally presented artificial language, in two groups of synesthetes alongside controls and simulated the multimodal experience of synesthesia in non-synesthetes. One group of synesthetes exclusively had grapheme-color (GC) synesthesia, in which the experience of color is automatically triggered by exposure to written or spoken graphemes. The other group had both grapheme-color and sound-color (SC+) synesthesia, in which the experience of color is also triggered by the waveform properties of a voice, such as pitch, timbre, and/or musical chords. Unlike GC-only synesthetes, the experience of color in the SC+ group is not perfectly consistent with the statistics that signal word boundaries. We showed that GC-only synesthetes outperformed both non-synesthetes and SC+ synesthetes, likely because the visual concurrents for GC-only synesthetes are highly consistent with the artificial language. We further observed that our simulations of GC synesthesia, but not SC+ synesthesia produced superior statistical learning, showing that synesthesia likely boosts learning outcomes by providing a consistent secondary cue. Findings are discussed with regard to how multimodal experience can improve learning, with the present data indicating that this boost is more likely to occur through explicit, as opposed to implicit, learning systems.

Capacities and neural mechanisms for auditory statistical learning across species

Statistical learning has been proposed as a possible mechanism by which individuals can become sensitive to the structures of language fundamental for speech perception. Since its description in human infants, statistical learning has been described in human adults and several non-human species as a general process by which animals learn about stimulus-relevant statistics. The neurobiology of statistical learning is beginning to be understood, but many questions remain about the underlying mechanisms. Why is the developing brain particularly sensitive to stimulus and environmental statistics, and what neural processes are engaged in the adult brain to enable learning from statistical regularities in the absence of external reward or instruction? This review will survey the statistical learning abilities of humans and non-human animals with a particular focus on communicative vocalizations. We discuss the neurobiological basis of statistical learning, and specifically what can be learned by exploring this process in both humans and laboratory animals. Finally, we describe advantages of studying vocal communication in rodents as a means to further our understanding of the cortical plasticity mechanisms engaged during statistical learning. We examine the use of rodents in the context of pup retrieval, which is an auditory-based and experience-dependent form of maternal behavior.

Prefrontal and Hippocampal Structure Predict Statistical Learning Ability in Early Childhood

Statistical learning can be used to gain sensitivity to many important regularities in our environment, including structure that is foundational to language and visual perception. As yet, little is known about how statistical learning takes place in the human brain, especially in children's developing brains and with regard to the broader neurobiology of learning and memory. We therefore explored the relationship between statistical learning and the thickness and volume of structures that are traditionally implicated in declarative and procedural memory, focusing specifically on the left inferior PFC, the hippocampus, and the caudate during early childhood (ages 5-8.5 years). We found that the thickness of the left inferior frontal cortex and volume of the right hippocampus predicted statistical learning ability in young children. Importantly, these regions did not change in thickness or volume with age, but the relationship between learning and the right hippocampus interacted with age such that older children's hippocampal structure more strongly predicted performance. Overall, the data show that children's statistical learning is supported by multiple neural structures that are more broadly implicated in learning and memory, especially declarative memory (hippocampus) and attention/top-down control (the PFC).

Auditory and Visual Sequence Learning in Humans and Monkeys using an Artificial Grammar Learning Paradigm

Language flexibly supports the human ability to communicate using different sensory modalities, such as writing and reading in the visual modality and speaking and listening in the auditory domain. Although it has been argued that nonhuman primate communication abilities are inherently multisensory, direct behavioural comparisons between human and nonhuman primates are scant. Artificial grammar learning (AGL) tasks and statistical learning experiments can be used to emulate ordering relationships between words in a sentence. However, previous comparative work using such paradigms has primarily investigated sequence learning within a single sensory modality. We used an AGL paradigm to evaluate how humans and macaque monkeys learn and respond to identically structured sequences of either auditory or visual stimuli. In the auditory and visual experiments, we found that both species were sensitive to the ordering relationships between elements in the sequences. Moreover, the humans and monkeys produced largely similar response patterns to the visual and auditory sequences, indicating that the sequences are processed in comparable ways across the sensory modalities. These results provide evidence that human sequence processing abilities stem from an evolutionarily conserved capacity that appears to operate comparably across the sensory modalities in both human and nonhuman primates. The findings set the stage for future neurobiological studies to investigate the multisensory nature of these sequencing operations in nonhuman primates and how they compare to related processes in humans.

Visual statistical learning deficits in memory-impaired individuals

Visual statistical learning (VSL) refers to the learning of environmental regularities. Classically considered an implicit process, one patient with isolated hippocampal damage is severely impaired at VSL tasks, suggesting involvement of explicit memory. Here, we asked whether memory impairment (MI) alone, absent of clear hippocampal pathology, predicted deficits across different VSL tasks. A classic VSL task revealed no learning in MI participants (Exp. 1), while imposing attentional demands (Exp. 2: flicker detection, Exp. 3: gender/location categorization) during familiarization revealed modest residual VSL. MI with nonspecific neural correlates predicted impaired VSL overall, but attentional processes may be harnessed for rehabilitation.

Statistical learning for speech segmentation: Age-related changes and underlying mechanisms

Statistical learning (SL) is a powerful learning mechanism that supports word segmentation and language acquisition in infants and young adults. However, little is known about how this ability changes over the life span and interacts with age-related cognitive decline. The aims of this study were to: (a) examine the effect of aging on speech segmentation by SL, and (b) explore core mechanisms underlying SL. Across four testing sessions, young, middle-aged, and older adults were exposed to continuous speech streams at two different speech rates, both with and without cognitive load. Learning was assessed using a two-alterative forced-choice task in which words from the stream were pitted against either part-words, which occurred across word boundaries in the stream, or nonwords, which never appeared in the stream. Participants also completed a battery of cognitive tests assessing working memory and executive functions. The results showed that speech segmentation by SL was remarkably resilient to aging, although age effects were visible in the more challenging conditions, namely, when words had to be discriminated from part-words, which required the formation of detailed phonological representations, and when SL was performed under cognitive load. Moreover, an analysis of the cognitive test data indicated that performance against part-words was predicted mostly by memory updating, whereas performance against nonwords was predicted mostly by working memory storage capacity. Taken together, the data show that SL relies on a combination of implicit and explicit skills, and that age effects on SL are likely to be linked to an age-related selective decline in memory updating. (PsycINFO Database Record (c) 2018 APA, all rights reserved).

Spontaneous Learning of Visual Structures in Domestic Chicks

Simple Summary

Our aim is to investigate the recognition of the structure of multi-element configurations; one mechanism that supports communicative functions in different species. Cognitive mechanisms involved in this ability might not have evolved specifically for communicative use, but derive from other functions. Thus, it is crucial to study these abilities in species that are not vocal learners and with stimuli from other modalities. We know already that domestic chicks can learn the temporal statistical structure of sequences of visual shapes, however their abilities to encode the spatial structure of visual patterns (configurations composed of multiple visual elements presented simultaneously side-by-side) is much less known. Using filial imprinting learning, we showed that chicks spontaneously recognize the structure of their imprinting stimulus, preferring it to one composed of the same elements in different configurations. Moreover, we found that in their affiliative responses chicks give priority to information located at the stimulus edges, a phenomenon that was so far observed only with temporal sequences. This first evidence of a spontaneous edge bias with spatial stimuli further stresses the importance of studying similarities and differences between the processing of linguistic and nonlinguistic stimuli and of stimuli presented in various sensory modalities.

Abstract

Effective communication crucially depends on the ability to produce and recognize structured signals, as apparent in language and birdsong. Although it is not clear to what extent similar syntactic-like abilities can be identified in other animals, recently we reported that domestic chicks can learn abstract visual patterns and the statistical structure defined by a temporal sequence of visual shapes. However, little is known about chicks’ ability to process spatial/positional information from visual configurations. Here, we used filial imprinting as an unsupervised learning mechanism to study spontaneous encoding of the structure of a configuration of different shapes. After being exposed to a triplet of shapes (ABC or CAB), chicks could discriminate those triplets from a permutation of the same shapes in different order (CAB or ABC), revealing a sensitivity to the spatial arrangement of the elements. When tested with a fragment taken from the imprinting triplet that followed the familiar adjacency-relationships (AB or BC) vs. one in which the shapes maintained their position with respect to the stimulus edges (AC), chicks revealed a preference for the configuration with familiar edge elements, showing an edge bias previously found only with temporal sequences.

Neurophysiological Markers of Statistical Learning in Music and Language: Hierarchy, Entropy, and Uncertainty

Statistical learning (SL) is a method of learning based on the transitional probabilities embedded in sequential phenomena such as music and language. It has been considered an implicit and domain-general mechanism that is innate in the human brain and that functions independently of intention to learn and awareness of what has been learned. SL is an interdisciplinary notion that incorporates information technology, artificial intelligence, musicology, and linguistics, as well as psychology and neuroscience. A body of recent study has suggested that SL can be reflected in neurophysiological responses based on the framework of information theory. This paper reviews a range of work on SL in adults and children that suggests overlapping and independent neural correlations in music and language, and that indicates disability of SL. Furthermore, this article discusses the relationships between the order of transitional probabilities (TPs) (i.e., hierarchy of local statistics) and entropy (i.e., global statistics) regarding SL strategies in human's brains; claims importance of information-theoretical approaches to understand domain-general, higher-order, and global SL covering both real-world music and language; and proposes promising approaches for the application of therapy and pedagogy from various perspectives of psychology, neuroscience, computational studies, musicology, and linguistics.

Non-adjacent Dependencies Processing in Human and Non-human Primates

Human and non-human primates share the ability to extract adjacent dependencies and, under certain conditions, non-adjacent dependencies (i.e., predictive relationships between elements that are separated by one or several intervening elements in a sequence). In this study, we explore the online extraction dynamics of non-adjacent dependencies in humans and baboons using a serial reaction time task. Participants had to produce three-target sequences containing deterministic relationships between the first and last target locations. In Experiment 1, participants from the two species could extract these non-adjacent dependencies, but humans required less exposure than baboons. In Experiment 2, the data show for the first time in a non-human primate species the successful generalization of sequential non-adjacent dependencies over novel intervening items. These findings provide new evidence to further constrain current theories about the nature and the evolutionary origins of the learning mechanisms allowing the extraction of non-adjacent dependencies.

Behavioral evaluation of auditory stream segregation in rats

Perceptual organization of sound sequences into separate sound sources or streams is called auditory stream segregation. Neural substrates for this process in both the spectral and temporal domains remain to be elucidated. Despite abundant knowledge about their auditory physiology, behavioral evidence for auditory streaming in rodents is still limited. We provided behavioral evidence for auditory streaming in the go/no-go discrimination task, but not in the two-alternative choice task. In the go/no-go discrimination phase, rats were able to discriminate different rhythms corresponding to segregated or integrated tone sequences in both short inter-tone interval (ITI) and long ITI conditions. Nevertheless, performance was poorer in the long ITI group. In probe testing, which assessed the ability to discriminate one of the segregated tone sequences from ABA- tone sequences, the detection rate increased with the difference in frequency (ΔF) for short (100 ms), but not long (200 ms) ITIs. Our results indicate that auditory streaming in rats on both the spectral and temporal features in the ABA- tone paradigm is qualitatively analogous to that observed in human psychophysics studies. This suggests that rodents are a valuable model for investigating the neural substrates of auditory streaming.

Constraints on Statistical Learning Across Species

Both human and nonhuman organisms are sensitive to statistical regularities in sensory inputs that support functions including communication, visual processing, and sequence learning. One of the issues faced by comparative research in this field is the lack of a comprehensive theory to explain the relevance of statistical learning across distinct ecological niches. In the current review we interpret cross-species research on statistical learning based on the perceptual and cognitive mechanisms that characterize the human and nonhuman models under investigation. Considering statistical learning as an essential part of the cognitive architecture of an animal will help to uncover the potential ecological functions of this powerful learning process.

Statistical learning in songbirds: from self-tutoring to song culture

At the onset of vocal development, both songbirds and humans produce variable vocal babbling with broadly distributed acoustic features. Over development, these vocalizations differentiate into the well-defined, categorical signals that characterize adult vocal behaviour. A broadly distributed signal is ideal for vocal exploration, that is, for matching vocal production to the statistics of the sensory input. The developmental transition to categorical signals is a gradual process during which the vocal output becomes differentiated and stable. But does it require categorical input? We trained juvenile zebra finches with playbacks of their own developing song, produced just a few moments earlier, updated continuously over development. Although the vocalizations of these self-tutored (ST) birds were initially broadly distributed, birds quickly developed categorical signals, as fast as birds that were trained with a categorical, adult song template. By contrast, siblings of those birds that received no training (isolates) developed phonological categories much more slowly and never reached the same level of category differentiation as their ST brothers. Therefore, instead of simply mirroring the statistical properties of their sensory input, songbirds actively transform it into distinct categories. We suggest that the early self-generation of phonological categories facilitates the establishment of vocal culture by making the song easier to transmit at the micro level, while promoting stability of shared vocabulary at the group level over generations.

This article is part of the themed issue ‘New frontiers for statistical learning in the cognitive sciences’.