MDLChunker: a MDL-based cognitive model of inductive learning

Grouping in working memory guides chunk formation in long-term memory: Evidence from the Hebb effect

The Hebb effect refers to the improvement in immediate memory performance on a repeated list compared to unrepeated lists. That is, participants create a long-term memory representation over repetitions, on which they can draw in working memory tests. These long-term memory representations are likely formed by chunk acquisition: The whole list becomes integrated into a single unified representation. Previous research suggests that the formation of such chunks is rather inflexible and only occurs when at least the beginning of the list repeats across trials. However, recent work has shown that repetition learning strongly depends on participants recognizing the repeated information. Hence, successful chunk formation may depend on the recognizability of the repeated part of a list, and not on its position in the list. Across six experiments, we compared these two alternatives. We tested immediate serial recall of eight-letter lists, some of which partially repeated across trials. We used different partial-repetition structures, such as repeating only the first half of a list, or only every second item. We manipulated the salience of the repeating structure by spatially grouping and coloring the lists according to the repetition structure. We found that chunk formation is more flexible than previously assumed: Participants learned contiguous repeated sequences regardless of their position within the list, as long as they were able to recognize the repeated structure. Even when the repeated sequence occurred at varying positions over repetitions, learning was preserved when the repeated sequence was made salient by the spatial grouping. These findings suggest that chunk formation requires recognition of which items constitute a repeating group, and demonstrate a close link between grouping of information in working memory, and chunk formation in long-term memory.

The dynamics of multiword sequence extraction

Being able to process multiword sequences is central for both language comprehension and production. Numerous studies support this claim, but less is known about the way multiword sequences are acquired, and more specifically how associations between their constituents are established over time. Here we adapted the Hebb naming task into a Hebb lexical decision task to study the dynamics of multiword sequence extraction. Participants had to read letter strings presented on a computer screen and were required to classify them as words or pseudowords. Unknown to the participants, a triplet of words or pseudowords systematically appeared in the same order and random words or pseudowords were inserted between two repetitions of the triplet. We found that response times (RTs) for the unpredictable first position in the triplet decreased over repetitions (i.e., indicating the presence of a repetition effect) but more slowly and with a different dynamic compared with items appearing at the predictable second and third positions in the repeated triplet (i.e., showing a slightly different predictability effect). Implicit and explicit learning also varied as a function of the nature of the triplet (i.e., unrelated words, pseudowords, semantically related words, or idioms). Overall, these results provide new empirical evidence about the dynamics of multiword sequence extraction, and more generally about the role of statistical learning in language acquisition.

Chunking and data compression in verbal short-term memory

Short-term verbal memory is improved when words can be chunked into larger units. Miller (1956) suggested that the capacity of verbal short-term memory is determined by the number of chunks that can be stored in memory, rather than by the number of items or the amount of information. But how does the improvement due to chunking come about, and is memory really determined by the number of chunks? One possibility is that chunking is a form of data compression. It allows more information to be stored in the available capacity. An alternative is that chunking operates primarily by redintegration. Chunks exist only in long-term memory, and enable the corresponding items in short-term memory to be reconstructed more reliably from a degraded trace. We review the data favoring each of these views and discuss the implications of treating chunking as data compression. Contrary to Miller, we suggest that memory capacity is primarily determined both by the amount of information that can be stored but also by the underlying representational vocabulary of the memory system. Given the limitations on the representations that can be stored in verbal short-term memory, chunking can sometimes allow the information capacity of short-term memory to be exploited more efficiently. (202 words).

Measuring children's auditory statistical learning via serial recall

Statistical learning (SL) has been a prominent focus of research in developmental and adult populations, guided by the assumption that it is a fundamental component of learning underlying higher-order cognition. In developmental populations, however, there have been recent concerns regarding the degree to which many current tasks reliably measure SL, particularly in younger children. In the current article, we present the results of two studies that measured auditory statistical learning (ASL) of linguistic stimuli in children aged 5-8 years. Children listened to 6 min of continuous syllables comprising four trisyllabic pseudowords. Following the familiarization phase, children completed (a) a two-alternative forced-choice task and (b) a serial recall task in which they repeated either target sequences embedded during familiarization or foils, manipulated for sequence length. Results showed that, although both measures consistently revealed learning at the group level, the recall task better captured learning across the full range of abilities and was more reliable at the individual level. We conclude that, as has also been demonstrated in adults, the method holds promise for future studies of individual differences in ASL of linguistic stimuli.

What Mechanisms Underlie Implicit Statistical Learning? Transitional Probabilities Versus Chunks in Language Learning

In a prior review, Perrruchet and Pacton (2006) noted that the literature on implicit learning and the more recent studies on statistical learning focused on the same phenomena, namely the domain-general learning mechanisms acting in incidental, unsupervised learning situations. However, they also noted that implicit learning and statistical learning research favored different interpretations, focusing on the selection of chunks and the computation of transitional probabilities aimed at discovering chunk boundaries, respectively. This paper examines the state of the debate 12 years later. The link between contrasting theories and their historical roots has disappeared, but a number of studies were aimed at contrasting the predictions of these two approaches. Overall, these studies strongly question the still prevalent account based on the statistical computation of pairwise associations. Various chunk-based models provide much better predictions in a number of experimental situations. However, these models rely on very different conceptual frameworks, as illustrated by a comparison between Bayesian models of word segmentation, PARSER, and a connectionist model (TRACX).

When learning goes beyond statistics: Infants represent visual sequences in terms of chunks

Much research has documented infants' sensitivity to statistical regularities in auditory and visual inputs, however the manner in which infants process and represent statistically defined information remains unclear. Two types of models have been proposed to account for this sensitivity: statistical models, which posit that learners represent statistical relations between elements in the input; and chunking models, which posit that learners represent statistically-coherent units of information from the input. Here, we evaluated the fit of these two types of models to behavioral data that we obtained from 8-month-old infants across four visual sequence-learning experiments. Experiments examined infants' representations of two types of structures about which statistical and chunking models make contrasting predictions: illusory sequences (Experiment 1) and embedded sequences (Experiments 2-4). In all four experiments, infants discriminated between high probability sequences and low probability part-sequences, providing strong evidence of learning. Critically, infants also discriminated between high probability sequences and statistically-matched sequences (illusory sequences in Experiment 1, embedded sequences in Experiments 2-3), suggesting that infants learned coherent chunks of elements. Experiment 4 examined the temporal nature of chunking, and demonstrated that the fate of embedded chunks depends on amount of exposure. These studies contribute important new data on infants' visual statistical learning ability, and suggest that the representations that result from infants' visual statistical learning are best captured by chunking models.

Compression in Working Memory and Its Relationship With Fluid Intelligence

Working memory has been shown to be strongly related to fluid intelligence; however, our goal is to shed further light on the process of information compression in working memory as a determining factor of fluid intelligence. Our main hypothesis was that compression in working memory is an excellent indicator for studying the relationship between working-memory capacity and fluid intelligence because both depend on the optimization of storage capacity. Compressibility of memoranda was estimated using an algorithmic complexity metric. The results showed that compressibility can be used to predict working-memory performance and that fluid intelligence is well predicted by the ability to compress information. We conclude that the ability to compress information in working memory is the reason why both manipulation and retention of information are linked to intelligence. This result offers a new concept of intelligence based on the idea that compression and intelligence are equivalent problems.

Predictive Movements and Human Reinforcement Learning of Sequential Action

Sequential action makes up the bulk of human daily activity, and yet much remains unknown about how people learn such actions. In one motor learning paradigm, the serial reaction time (SRT) task, people are taught a consistent sequence of button presses by cueing them with the next target response. However, the SRT task only records keypress response times to a cued target, and thus it cannot reveal the full time‐course of motion, including predictive movements. This paper describes a mouse movement trajectory SRT task in which the cursor must be moved to a cued location. We replicated keypress SRT results, but also found that predictive movement—before the next cue appears—increased during the experiment. Moreover, trajectory analyses revealed that people developed a centering strategy under uncertainty. In a second experiment, we made prediction explicit, no longer cueing targets. Thus, participants had to explore the response alternatives and learn via reinforcement, receiving rewards and penalties for correct and incorrect actions, respectively. Participants were not told whether the sequence of stimuli was deterministic, nor if it would repeat, nor how long it was. Given the difficulty of the task, it is unsurprising that some learners performed poorly. However, many learners performed remarkably well, and some acquired the full 10‐item sequence within 10 repetitions. Comparing the high‐ and low‐performers’ detailed results in this reinforcement learning (RL) task with the first experiment's cued trajectory SRT task, we found similarities between the two tasks, suggesting that the effects in Experiment 1 are due to predictive, rather than reactive processes. Finally, we found that two standard model‐free reinforcement learning models fit the high‐performing participants, while the four low‐performing participants provide better fit with a simple negative recency bias model.

Developmental Abilities to Form Chunks in Immediate Memory and Its Non-Relationship to Span Development

Both adults and children –by the time they are 2–3 years old– have a general ability to recode information to increase memory efficiency. This paper aims to evaluate the ability of untrained children aged 6–10 years old to deploy such a recoding process in immediate memory. A large sample of 374 children were given a task of immediate serial report based on SIMON®, a classic memory game made of four colored buttons (red, green, yellow, blue) requiring players to reproduce a sequence of colors within which repetitions eventually occur. It was hypothesized that a primitive ability across all ages (since theoretically already available in toddlers) to detect redundancies allows the span to increase whenever information can be recoded on the fly. The chunkable condition prompted the formation of chunks based on the perceived structure of color repetition within to-be-recalled sequences of colors. Our result shows a similar linear improvement of memory span with age for both chunkable and non-chunkable conditions. The amount of information retained in immediate memory systematically increased for the groupable sequences across all age groups, independently of the average age-group span that was measured on sequences that contained fewer repetitions. This result shows that chunking gives young children an equal benefit as older children. We discuss the role of recoding in the expansion of capacity in immediate memory and the potential role of data compression in the formation of chunks in long-term memory.

Optimal behavioral hierarchy

Human behavior has long been recognized to display hierarchical structure: actions fit together into subtasks, which cohere into extended goal-directed activities. Arranging actions hierarchically has well established benefits, allowing behaviors to be represented efficiently by the brain, and allowing solutions to new tasks to be discovered easily. However, these payoffs depend on the particular way in which actions are organized into a hierarchy, the specific way in which tasks are carved up into subtasks. We provide a mathematical account for what makes some hierarchies better than others, an account that allows an optimal hierarchy to be identified for any set of tasks. We then present results from four behavioral experiments, suggesting that human learners spontaneously discover optimal action hierarchies.

In order to accomplish everyday tasks, we often divide them up into subtasks: to make spaghetti, we (1) get out a pot, (2) fill it with water, (3) bring the water to a boil, and so forth. But how do we learn to subdivide our goals in this way? Work from computer science suggests that the way a task is subdivided or decomposed can have a dramatic impact on how easy the task is to accomplish: certain decompositions speed learning and planning compared to others. Moreover, some decompositions allow behaviors to be represented more simply. Despite this general insight, little work has been done to formalize these ideas. We outline a mathematical framework to address this question, based on methods for comparing between statistical models. We then present four behavioral experiments, showing that human learners spontaneously discover optimal task decompositions.

New evidence for chunk-based models in word segmentation

There is large evidence that infants are able to exploit statistical cues to discover the words of their language. However, how they proceed to do so is the object of enduring debates. The prevalent position is that words are extracted from the prior computation of statistics, in particular the transitional probabilities between syllables. As an alternative, chunk-based models posit that the sensitivity to statistics results from other processes, whereby many potential chunks are considered as candidate words, then selected as a function of their relevance. These two classes of models have proven to be difficult to dissociate. We propose here a procedure, which leads to contrasted predictions regarding the influence of a first language, L1, on the segmentation of a second language, L2. Simulations run with PARSER (Perruchet & Vinter, 1998), a chunk-based model, predict that when the words of L1 become word-external transitions of L2, learning of L2 should be depleted until reaching below chance level, at least before extensive exposure to L2 reverses the effect. In the same condition, a transitional-probability based model predicts above-chance performance whatever the duration of exposure to L2. PARSER's predictions were confirmed by experimental data: Performance on a two-alternative forced choice test between words and part-words from L2 was significantly below chance even though part-words were less cohesive in terms of transitional probabilities than words.

The use of limericks to engage student interest and promote active learning in an undergraduate course in functional anatomy

The study of anatomy is a content-dense discipline with a challenging vocabulary. A mnemonic is a series of letters, a word, a phrase, or a rhyme that students can use when studying to facilitate recall. This project was designed to promote active learning in undergraduate students studying anatomy and physiology by asking them to create limericks based on course content and then to evaluate the limericks written by their peers for learning value, accuracy, style, and adherence to limerick characteristics. Students (278 and 288, respectively, in the 2009 and 2010 sections of ANP1107) worked in groups of three to create a total of 242 limericks. Peer evaluation was accomplished in two stages using a 20-point marking rubric. In Stage 1, students were randomly divided into 10 groups (n = 23 ± 2 students) with each group member evaluating the same 12 ± 1 limericks. In Stage 2, the top 19% of limericks were reevaluated by all students so that the best three could be chosen. In each of the two years, 60% of students completed all parts of the assignment. Higher percentages (75-80%) participated in limerick writing and one of the two assessment stages. A positive association was noted between level of student participation in the limerick assignment and final course marks. Limerick creation and evaluation can be used to promote active learning by encouraging students to review functional-anatomy-based content to create limericks with good learning value and to provide valid assessments of limericks written by their peers.

What's magic about magic numbers? Chunking and data compression in short-term memory

Short term memory is famously limited in capacity to Miller's (1956) magic number 7±2-or, in many more recent studies, about 4±1 "chunks" of information. But the definition of "chunk" in this context has never been clear, referring only to a set of items that are treated collectively as a single unit. We propose a new more quantitatively precise conception of chunk derived from the notion of Kolmogorov complexity and compressibility: a chunk is a unit in a maximally compressed code. We present a series of experiments in which we manipulated the compressibility of stimulus sequences by introducing sequential patterns of variable length. Our subjects' measured digit span (raw short term memory capacity) consistently depended on the length of the pattern after compression, that is, the number of distinct sequences it contained. The true limit appears to be about 3 or 4 distinct chunks, consistent with many modern studies, but also equivalent to about 7 uncompressed items of typical compressibility, consistent with Miller's famous magical number.