Hebbian learning can explain rhythmic neural entrainment to statistical regularities

In many domains, learners extract recurring units from continuous sequences. For example, in unknown languages, fluent speech is perceived as a continuous signal. Learners need to extract the underlying words from this continuous signal and then memorize them. One prominent candidate mechanism is statistical learning, whereby learners track how predictive syllables (or other items) are of one another. Syllables within the same word predict each other better than syllables straddling word boundaries. But does statistical learning lead to memories of the underlying words-or just to pairwise associations among syllables? Electrophysiological results provide the strongest evidence for the memory view. Electrophysiological responses can be time-locked to statistical word boundaries (e.g., N400s) and show rhythmic activity with a periodicity of word durations. Here, I reproduce such results with a simple Hebbian network. When exposed to statistically structured syllable sequences (and when the underlying words are not excessively long), the network activation is rhythmic with the periodicity of a word duration and activation maxima on word-final syllables. This is because word-final syllables receive more excitation from earlier syllables with which they are associated than less predictable syllables that occur earlier in words. The network is also sensitive to information whose electrophysiological correlates were used to support the encoding of ordinal positions within words. Hebbian learning can thus explain rhythmic neural activity in statistical learning tasks without any memory representations of words. Learners might thus need to rely on cues beyond statistical associations to learn the words of their native language. RESEARCH HIGHLIGHTS: Statistical learning may be utilized to identify recurring units in continuous sequences (e.g., words in fluent speech) but may not generate explicit memory for words. Exposure to statistically structured sequences leads to rhythmic activity with a period of the duration of the underlying units (e.g., words). I show that a memory-less Hebbian network model can reproduce this rhythmic neural activity as well as putative encodings of ordinal positions observed in earlier research. Direct tests are needed to establish whether statistical learning leads to declarative memories for words.

Hebbian, correlational learning provides a memory-less mechanism for Statistical Learning irrespective of implementational choices: Reply to Tovar and Westermann (2022)

Statistical learning relies on detecting the frequency of co-occurrences of items and has been proposed to be crucial for a variety of learning problems, notably to learn and memorize words from fluent speech. Endress and Johnson (2021) (hereafter EJ) recently showed that such results can be explained based on simple memory-less correlational learning mechanisms such as Hebbian Learning. Tovar and Westermann (2022) (hereafter TW) reproduced these results with a different Hebbian model. We show that the main differences between the models are whether temporal decay acts on both the connection weights and the activations (in TW) or only on the activations (in EJ), and whether interference affects weights (in TW) or activations (in EJ). Given that weights and activations are linked through the Hebbian learning rule, the networks behave similarly. However, in contrast to TW, we do not believe that neurophysiological data are relevant to adjudicate between abstract psychological models with little biological detail. Taken together, both models show that different memory-less correlational learning mechanisms provide a parsimonious account of Statistical Learning results. They are consistent with evidence that Statistical Learning might not allow learners to learn and retain words, and Statistical Learning might support predictive processing instead.

Socio-Cultural Values Are Risk Factors for COVID-19-Related Mortality

To assess whether socio-cultural values are population-level risk factors for health, I sought to predict COVID-19-related mortality between 2 weeks and 6 months after the first COVID-19-related death in a country based on values extracted from the World Values Survey for different country sets, after controlling for various confounding variables. COVID-19-related mortality was increased in countries endorsing political participation but decreased in countries with greater trust in institutions and materialistic orientations. The values were specific to COVID-19-related mortality, did not predict general health outcomes, and values predicting increased COVID-19-related mortality predicted decreased mortality from other outcomes (e.g., environmental-related mortality).

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 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.

Sequential Presentation Protects Working Memory From Catastrophic Interference

Neural network models of memory are notorious for catastrophic interference: Old items are forgotten as new items are memorized (French, 1999; McCloskey & Cohen, 1989). While working memory (WM) in human adults shows severe capacity limitations, these capacity limitations do not reflect neural network style catastrophic interference. However, our ability to quickly apprehend the numerosity of small sets of objects (i.e., subitizing) does show catastrophic capacity limitations, and this subitizing capacity and WM might reflect a common capacity. Accordingly, computational investigations (Knops, Piazza, Sengupta, Eger & Melcher, 2014; Sengupta, Surampudi & Melcher, 2014) suggest that mutual inhibition among neurons can explain both kinds of capacity limitations as well as why our ability to estimate the numerosity of larger sets is limited according to a Weber ratio signature. Based on simulations with a saliency map-like network and mathematical proofs, we provide three results. First, mutual inhibition among neurons leads to catastrophic interference when items are presented simultaneously. The network can remember a limited number of items, but when more items are presented, the network forgets all of them. Second, if memory items are presented sequentially rather than simultaneously, the network remembers the most recent items rather than forgetting all of them. Hence, the tendency in WM tasks to sequentially attend even to simultaneously presented items might not only reflect attentional limitations, but also an adaptive strategy to avoid catastrophic interference. Third, the mean activation level in the network can be used to estimate the number of items in small sets, but it does not accurately reflect the number of items in larger sets. Rather, we suggest that the Weber ratio signature of large number discrimination emerges naturally from the interaction between the limited precision of a numeric estimation system and a multiplicative gain control mechanism.

A Simple, Biologically Plausible Feature Detector for Language Acquisition

Language has a complex grammatical system we still have to understand computationally and biologically. However, some evolutionarily ancient mechanisms have been repurposed for grammar so that we can use insight from other taxa into possible circuit-level mechanisms of grammar. Drawing upon recent evidence for the importance of disinhibitory circuits across taxa and brain regions, I suggest a simple circuit that explains the acquisition of core grammatical rules used in 85% of the world's languages: grammatical rules based on sameness/difference relations. This circuit acts as a sameness detector. “Different” items are suppressed through inhibition, but presenting two “identical” items leads to inhibition of inhibition. The items are thus propagated for further processing. This sameness detector thus acts as a feature detector for a grammatical rule. I suggest that having a set of feature detectors for elementary grammatical rules might make language acquisition feasible based on relatively simple computational mechanisms.

Interference and memory capacity limitations

Working memory (WM) is thought to have a fixed and limited capacity. However, the origins of these capacity limitations are debated, and generally attributed to active, attentional processes. Here, we show that the existence of interference among items in memory mathematically guarantees fixed and limited capacity limits under very general conditions, irrespective of any processing assumptions. Assuming that interference (a) increases with the number of interfering items and (b) brings memory performance to chance levels for large numbers of interfering items, capacity limits are a simple function of the relative influence of memorization and interference. In contrast, we show that time-based memory limitations do not lead to fixed memory capacity limitations that are independent of the timing properties of an experiment. We show that interference can mimic both slot-like and continuous resource-like memory limitations, suggesting that these types of memory performance might not be as different as commonly believed. We speculate that slot-like WM limitations might arise from crowding-like phenomena in memory when participants have to retrieve items. Further, based on earlier research on parallel attention and enumeration, we suggest that crowding-like phenomena might be a common reason for the 3 major cognitive capacity limitations. As suggested by Miller (1956) and Cowan (2001), these capacity limitations might arise because of a common reason, even though they likely rely on distinct processes. (PsycINFO Database Record

Transitional probabilities count more than frequency, but might not be used for memorization

Learners often need to extract recurring items from continuous sequences, in both vision and audition. The best-known example is probably found in word-learning, where listeners have to determine where words start and end in fluent speech. This could be achieved through universal and experience-independent statistical mechanisms, for example by relying on Transitional Probabilities (TPs). Further, these mechanisms might allow learners to store items in memory. However, previous investigations have yielded conflicting evidence as to whether a sensitivity to TPs is diagnostic of the memorization of recurring items. Here, we address this issue in the visual modality. Participants were familiarized with a continuous sequence of visual items (i.e., arbitrary or everyday symbols), and then had to choose between (i) high-TP items that appeared in the sequence, (ii) high-TP items that did not appear in the sequence, and (iii) low-TP items that appeared in the sequence. Items matched in TPs but differing in (chunk) frequency were much harder to discriminate than items differing in TPs (with no significant sensitivity to chunk frequency), and learners preferred unattested high-TP items over attested low-TP items. Contrary to previous claims, these results cannot be explained on the basis of the similarity of the test items. Learners thus weigh within-item TPs higher than the frequency of the chunks, even when the TP differences are relatively subtle. We argue that these results are problematic for distributional clustering mechanisms that analyze continuous sequences, and provide supporting computational results. We suggest that the role of TPs might not be to memorize items per se, but rather to prepare learners to memorize recurring items once they are presented in subsequent learning situations with richer cues.

The cost of proactive interference is constant across presentation conditions

Proactive interference (PI) severely constrains how many items people can remember. For example, Endress and Potter (2014a) presented participants with sequences of everyday objects at 250ms/picture, followed by a yes/no recognition test. They manipulated PI by either using new images on every trial in the unique condition (thus minimizing PI among items), or by re-using images from a limited pool for all trials in the repeated condition (thus maximizing PI among items). In the low-PI unique condition, the probability of remembering an item was essentially independent of the number of memory items, showing no clear memory limitations; more traditional working memory-like memory limitations appeared only in the high-PI repeated condition. Here, we ask whether the effects of PI are modulated by the availability of long-term memory (LTM) and verbal resources. Participants viewed sequences of 21 images, followed by a yes/no recognition test. Items were presented either quickly (250ms/image) or sufficiently slowly (1500ms/image) to produce LTM representations, either with or without verbal suppression. Across conditions, participants performed better in the unique than in the repeated condition, and better for slow than for fast presentations. In contrast, verbal suppression impaired performance only with slow presentations. The relative cost of PI was remarkably constant across conditions: relative to the unique condition, performance in the repeated condition was about 15% lower in all conditions. The cost of PI thus seems to be a function of the relative strength or recency of target items and interfering items, but relatively insensitive to other experimental manipulations.

Words, rules, and mechanisms of language acquisition

We review recent artificial language learning studies, especially those following Endress and Bonatti (Endress AD, Bonatti LL. Rapid learning of syllable classes from a perceptually continuous speech stream. Cognition 2007, 105:247-299), suggesting that humans can deploy a variety of learning mechanisms to acquire artificial languages. Several experiments provide evidence for multiple learning mechanisms that can be deployed in fluent speech: one mechanism encodes the positions of syllables within words and can be used to extract generalization, while the other registers co-occurrence statistics of syllables and can be used to break a continuum into its components. We review dissociations between these mechanisms and their potential role in language acquisition. We then turn to recent criticisms of the multiple mechanisms hypothesis and show that they are inconsistent with the available data. Our results suggest that artificial and natural language learning is best understood by dissecting the underlying specialized learning abilities, and that these data provide a rare opportunity to link important language phenomena to basic psychological mechanisms. For further resources related to this article, please visit the WIREs website.

Large capacity temporary visual memory

Visual working memory (WM) capacity is thought to be limited to 3 or 4 items. However, many cognitive activities seem to require larger temporary memory stores. Here, we provide evidence for a temporary memory store with much larger capacity than past WM capacity estimates. Further, based on previous WM research, we show that a single factor--proactive interference--is sufficient to bring capacity estimates down to the range of previous WM capacity estimates. Participants saw a rapid serial visual presentation of 5-21 pictures of familiar objects or words presented at rates of 4/s or 8/s, respectively, and thus too fast for strategies such as rehearsal. Recognition memory was tested with a single probe item. When new items were used on all trials, no fixed memory capacities were observed, with estimates of up to 9.1 retained pictures for 21-item lists, and up to 30.0 retained pictures for 100-item lists, and no clear upper bound to how many items could be retained. Further, memory items were not stored in a temporally stable form of memory but decayed almost completely after a few minutes. In contrast, when, as in most WM experiments, a small set of items was reused across all trials, thus creating proactive interference among items, capacity remained in the range reported in previous WM experiments. These results show that humans have a large-capacity temporary memory store in the absence of proactive interference, and raise the question of whether temporary memory in everyday cognitive processing is severely limited, as in WM experiments, or has the much larger capacity found in the present experiments.

Bayesian learning and the psychology of rule induction

In recent years, Bayesian learning models have been applied to an increasing variety of domains. While such models have been criticized on theoretical grounds, the underlying assumptions and predictions are rarely made concrete and tested experimentally. Here, I use Frank and Tenenbaum's (2011) Bayesian model of rule-learning as a case study to spell out the underlying assumptions, and to confront them with the empirical results Frank and Tenenbaum (2011) propose to simulate, as well as with novel experiments. While rule-learning is arguably well suited to rational Bayesian approaches, I show that their models are neither psychologically plausible nor ideal observer models. Further, I show that their central assumption is unfounded: humans do not always preferentially learn more specific rules, but, at least in some situations, those rules that happen to be more salient. Even when granting the unsupported assumptions, I show that all of the experiments modeled by Frank and Tenenbaum (2011) either contradict their models, or have a large number of more plausible interpretations. I provide an alternative account of the experimental data based on simple psychological mechanisms, and show that this account both describes the data better, and is easier to falsify. I conclude that, despite the recent surge in Bayesian models of cognitive phenomena, psychological phenomena are best understood by developing and testing psychological theories rather than models that can be fit to virtually any data.

Early conceptual and linguistic processes operate in independent channels

Language and concepts are intimately linked, but how do they interact? In the study reported here, we probed the relation between conceptual and linguistic processing at the earliest processing stages. We presented observers with sequences of visual scenes lasting 200 or 250 ms per picture. Results showed that observers understood and remembered the scenes' abstract gist and, therefore, their conceptual meaning. However, observers remembered the scenes at least as well when they simultaneously performed a linguistic secondary task (i.e., reading and retaining sentences); in contrast, a nonlinguistic secondary task (equated for difficulty with the linguistic task) impaired scene recognition. Further, encoding scenes interfered with performance on the nonlinguistic task and vice versa, but scene processing and performing the linguistic task did not affect each other. At the earliest stages of conceptual processing, the extraction of meaning from visually presented linguistic stimuli and the extraction of conceptual information from the world take place in remarkably independent channels.

Learning melodies from non-adjacent tones

Language acquisition might heavily rely on statistical learning mechanisms. This has led many researchers to investigate the computational constraints that limit such learning. In particular, it has been argued that statistical relations among non-adjacent items cannot be tracked, as this might lead to a "computational explosion" making statistical learning intractable. In line with this view, previous research suggests that listeners cannot track relations among non-adjacent musical tones (Creel, Newport, & Aslin, 2004). Here I show that participants readily track non-adjacent tone relations when these are implemented in a musically meaningful way. Specifically, participants readily track non-adjacent tone relations in tonal melodies, but find it more difficult to track non-adjacent tone relations in random melodies, suggesting that non-adjacent relations are easier to track when listeners face "ecological", musically meaningful stimuli.

The apes' edge: positional learning in chimpanzees and humans

A wide variety of organisms produce actions and signals in particular temporal sequences, including the motor actions recruited during tool-mediated foraging, the arrangement of notes in the songs of birds, whales and gibbons, and the patterning of words in human speech. To accurately reproduce such events, the elements that comprise such sequences must be memorized. Both memory and artificial language learning studies have revealed at least two mechanisms for memorizing sequences, one tracking co-occurrence statistics among items in sequences (i.e., transitional probabilities) and the other one tracking the positions of items in sequences, in particular those of items in sequence-edges. The latter mechanism seems to dominate the encoding of sequences after limited exposure, and to be recruited by a wide array of grammatical phenomena. To assess whether humans differ from other species in their reliance on one mechanism over the other after limited exposure, we presented chimpanzees (Pan troglodytes) and human adults with brief exposure to six items, auditory sequences. Each sequence consisted of three distinct sound types (X, A, B), arranged according to two simple temporal rules: the A item always preceded the B item, and the sequence-edges were always occupied by the X item. In line with previous results with human adults, both species primarily encoded positional information from the sequences; that is, they kept track of the items that occurred in the sequence-edges. In contrast, the sensitivity to co-occurrence statistics was much weaker. Our results suggest that a mechanism to spontaneously encode positional information from sequences is present in both chimpanzees and humans and may represent the default in the absence of training and with brief exposure. As many grammatical regularities exhibit properties of this mechanism, it may be recruited by language and constrain the form that certain grammatical regularities take.

Primitive computations in speech processing

Previous research suggests that artificial-language learners exposed to quasi-continuous speech can learn that the first and the last syllables of words have to belong to distinct classes (e.g., Endress & Bonatti, 2007; Peña, Bonatti, Nespor, & Mehler, 2002). The mechanisms of these generalizations, however, are debated. Here we show that participants learn such generalizations only when the crucial syllables are in edge positions (i.e., the first and the last), but not when they are in medial positions (i.e., the second and the fourth in pentasyllabic items). In contrast to the generalizations, participants readily perform statistical analyses also in word middles. In analogy to sequential memory, we suggest that participants extract the generalizations using a simple but specific mechanism that encodes the positions of syllables that occur in edges. Simultaneously, they use another mechanism to track the syllable distribution in the speech streams. In contrast to previous accounts, this model explains why the generalizations are faster than the statistical computations, require additional cues, and break down under different conditions, and why they can be performed at all. We also show that that similar edge-based mechanisms may explain many results in artificial-grammar learning and also various linguistic observations.

Evidence of an evolutionary precursor to human language affixation in a non-human primate

Human language, and grammatical competence in particular, relies on a set of computational operations that, in its entirety, is not observed in other animals. Such uniqueness leaves open the possibility that components of our linguistic competence are shared with other animals, having evolved for non-linguistic functions. Here, we explore this problem from a comparative perspective, asking whether cotton-top tamarin monkeys (Saguinus oedipus) can spontaneously (no training) acquire an affixation rule that shares important properties with our inflectional morphology (e.g. the rule that adds -ed to create the past tense, as in the transformation of walk into walk-ed). Using playback experiments, we show that tamarins discriminate between bisyllabic items that start with a specific 'prefix' syllable and those that end with the same syllable as a 'suffix'. These results suggest that some of the computational mechanisms subserving affixation in a diversity of languages are shared with other animals, relying on basic perceptual or memory primitives that evolved for non-linguistic functions.

Rapid learning of syllable classes from a perceptually continuous speech stream

To learn a language, speakers must learn its words and rules from fluent speech; in particular, they must learn dependencies among linguistic classes. We show that when familiarized with a short artificial, subliminally bracketed stream, participants can learn relations about the structure of its words, which specify the classes of syllables occurring in first and last word positions. By studying the effect of familiarization length, we compared the general predictions of associative theories of learning and those of models postulating separate mechanisms for quickly extracting the word structure and for tracking the syllable distribution in the stream. As predicted by the dual-mechanism model, the preference for structurally correct items was negatively correlated with the familiarization length. This result is difficult to explain by purely associative schemes; an extensive set of neural network simulations confirmed this difficulty. Still, we show that powerful statistical computations operating on the stream are available to our participants, as they are sensitive to co-occurrence statistics among non-adjacent syllables. We suggest that different learning mechanisms analyze speech on-line: A rapid mechanism extracting structural information about the stream, and a slower mechanism detecting statistical regularities among the items occurring in it.

Perceptual constraints and the learnability of simple grammars

Cognitive processes are often attributed to statistical or symbolic general-purpose mechanisms. Here we show that some spontaneous generalizations are driven by specialized, highly constrained symbolic operations. We explore how two types of artificial grammars are acquired, one based on repetitions and the other on characteristic relations between tones ("ordinal" grammars). Whereas participants readily acquire repetition-based grammars, displaying early electrophysiological responses to grammar violations, they perform poorly with ordinal grammars, displaying no such electrophysiological responses. This outcome is problematic for both general symbolic and statistical models, which predict that both types of grammars should be processed equally easily. This suggests that some simple grammars are acquired using perceptual primitives rather than general-purpose mechanisms; such primitives may be elements of a "toolbox" of specialized computational heuristics, which may ultimately allow constructing a psychological theory of symbol manipulation.

The role of salience in the extraction of algebraic rules

Recent research suggests that humans and other animals have sophisticated abilities to extract both statistical dependencies and rule-based regularities from sequences. Most of this research stresses the flexibility and generality of such processes. Here the authors take up an equally important project, namely, to explore the limits of such processes. As a case study for rule-based generalizations, the authors demonstrate that only repetition-based structures with repetitions at the edges of sequences (e.g., ABCDEFF but not ABCDDEF) can be reliably generalized, although token repetitions can easily be discriminated at both sequence edges and middles. This finding suggests limits on rule-based sequence learning and new interpretations of earlier work alleging rule learning in infants. Rather than implementing a computerlike, formal process that operates over all patterns equally well, rule-based learning may be a highly constrained and piecemeal process driven by perceptual primitives--specialized type operations that are highly sensitive to perceptual factors.