Statistical Learning at the UC Davis Language Learning Lab
Embedded Files
What is Statistical Learning?
A neighborhood dog parks every time it sees a bird. You have two foods that you always eat together. There is a mesmerizing poster you see, which has a series of repeating shapes. During a speech, you recognize where one word ends and the next begins.
While these may not seem like remarkable events, they highlight a mechanism known as statistical learning. Every day, we are exposed to countless patterns of co-occurring objects, images, sounds, and events. Humans are incredibly good at detecting these patterns, starting in infancy. Statistical learning is a domain-general process, meaning that it applies to multiple components of our lives. We are able to detect and interpret patterns in all kinds of environmental structures. For example, infants are not explicitly taught the boundaries of words, but learn the statistical regularities of how their native language sounds, giving them experience with language before they can speak.
In previous research, statistical learning has been associated with language acquisition, memory development, and education (Apfelbaum et al., 2013). In fact, infants who are better at detecting patterns in spoken language have larger vocabularies. Picking up on patterns across modalities is highly beneficial during early learning. For its wide range of applications, it is important to understand the underlying components that drive statistical learning, and how children within different contexts learn to recognize patterns in their lives. Dr. Graf Estes’s Language Learning Lab has been conducting meaningful research investigating statistical learning since 2006.
Insert awesome video
How do different hearing experiences affect pattern learning?
This study aims to understand how children with different hearing and language experiences learn patterns across multiple modalities. We want to learn about the role that signed versus spoken language experience plays in pattern learning. Based on previous research, we expect that any visual language experience (regardless of hearing status) will be associated with better visual statistical tracking (Hall et al., 2017).
Appointments involve multiple tasks to better understand pattern recognition and cognitive control. First, 5–7-year-old children participate in a visual statistical learning task. Using an eye tracker, their gaze is recorded as they view patterns of cartoon aliens in spaceships. They are then shown clusters of aliens and asked to recall if they saw the same pattern (the correct aliens in the correct spaceships in order) during the learning phase. By analyzing eyetracking data and pattern learning performance, we hope to learn more about the role that hearing experience plays in statistical learning. In addition to the visual pattern learning portion, the child listens to an artificial language—the same language the aliens from earlier spoke in—for eight minutes (if they can hear), and then answers questions about what they heard. Children also play a cognitive control game, where they have to respond to different probes. Lastly, they complete a receptive vocabulary test (in either English or ASL).
We are collecting data in the lab, in local schools, and working towards adapting the study for in-home opportunities. By collecting data on hearing and non-hearing children, we want to analyze the role that hearing experience plays in visual statistical learning and cognition. These findings can help to inform healthcare and educational policies.
Word patterns can be tracked visually (syllables in writing) and auditorily (boundaries between sounds).
One of the aliens from our study!
How do infants detect patterns across different settings?
This research project seeks to understand how infants learn to connect two different names to one object across settings. While previous research has found that infants are able to pick up patterns across different contexts, less is known about this same process for infants acquiring two-to-one mappings (e.g., a dog is a puppy and a perro) more frequently, such as in bilingual environments. We hope to identify a link between background differences and word-object association performance (Bialystok et al., 2010).
This research project seeks to understand how infants learn to connect two different names to one object across settings. While previous research has found that infants are able to pick up patterns across different contexts, less is known about this same process for infants acquiring two-to-one mappings (e.g., a dog is a puppy and a perro) more frequently, such as in bilingual environments. We hope to identify a link between background differences and word-object association performance (Bialystok et al., 2010).
Many researchers disagree about a possible cognitive benefit of bilingualism. By analyzing infants’ performance in mapping two names to a single object, we can better understand the mechanisms and experiences necessary to support cross-situational word-mapping. These findings can help inform future research and policy related to bilingualism and early language acquisition.
A parent and child in the lab!
How do physical interactions affect pattern recognition?
The goal of this research project is to better understand how an infant's physical interaction with their surrounding environment might change their perception and association with co-occurring objects. In the past, similar experiments would virtually display three separate objects clustered together, and in each trial one of these objects would be switched out with a unique object. Infants tended to look longer at these new objects, and these longer gazes were interpreted as surprise that their previously learned pattern (the original three objects) had changed (Wu et al., 2011). In addition to showing these objects on a 2D screen, this specific experiment has a section where the objects are 3D printed toys that are magnetized to form clusters with each other. This first part of the experiment will observe infants as they physically manipulate the 3 separate objects in their individual clusters. This added aspect aims to further capture how infants might come across and learn about multiple objects in their daily environment through physical interactions in their 3D environment as opposed to simply viewing the objects. The second section of this experiment will have infants view the previous objects on a computer screen and the infants’ location and length of their gaze will be recorded.
Currently, the prediction of the study is that the infants who spent more time playing around with the 3D printed objects will be able to notice differences in the cluster patterns. Consistent with other visual statistical learning experiments, infants will also stare longer at clusters with unfamiliar objects.
1
1
11111
Examples of our unique, 3d printed toys! In these clusters, the top left toy is changing over the course of the trials.
Examples of our unique, 3d printed toys! In these clusters, the top left toy is changing over the course of the trials.
sfsse
se
se
sf
Dr. Graf Estes' Older Studies
Dr. Graf Estes, the principal investigator of this lab, has also done multiple statistical learning studies in the past. Here is a summary of the main findings:
Similar to the methods used in this lab’s statistical learning studies, Dr. Graf Estes has also used artificial language to understand how infants pick up patterns in languages. In one of her studies, she starts off by exposing some of the infants out of the study’s participant pool to an artificial language, while the other infants were not exposed. After the initial exposure, the duration and location of all the infant’s gazes were observed as they stared at the experiment’s words/part-words (words consist of made up syllables that will always co-occur with each other, while part-words consist of made up syllables that only sometimes co-occur with each other). The results found that infants who were pre-exposed to the language would stare longer at the part-words. The longer gazes suggest some sort of surprise from the infants, indicating that the infants must have learned a language pattern (normal words) from the first phase of the experience that was now being challenged (new co-occurances with syllables aka part-words). Infants who were not pre-exposed did not have any difference in their looking times and this is because they did not learn any patterns. Additionally, children were able to discriminate between words and part-words even with a difference in the speakers’ voices, implying a child’s concept of language is flexible and not static (Graf Estes et al., 2012).
Other studies have also added onto these findings. In a study from 2016, it was found that infants with larger vocabularies were able to pick up phonotactic patterns better. These patterns are the rules by which sounds in languages abide by, essentially acting as building blocks (Graf Estes et al., 2016). An example of this in English is how consonants rarely follow each other in a word – you don’t often see “bm” or “mn” next to each other. Simultaneously, infants who have a better phonotactic understanding of their first language would struggle with learning new patterns (Antovich et al., 2020). These sentiments are echoed again in another study from 2016, where it was found that infants who can learn statistical patterns well will be able to understand speech faster (Lany et al., 2016). Children with a bigger vocabulary are also able to associate meanings to words faster because they do not have to actively learn the label. For example, a child who already knows the word “dinosaur” doesn’t have to channel their energy into learning the word but instead can focus on learning what a dinosaur actually is and the multiple types of dinosaurs that fall into the label’s category (Erickson et al., 2014).
In more recent years, Dr. Graf Estes has shifted her research focus towards bilingual infants. Newer studies have found that when infants are exposed to two different streams of artificial languages, monolingual infants would struggle with learning patterns for the second language. Bilingual children, on the other hand, were able to recognize words from both languages, suggesting that the younger children are exposed to multiple languages, the better infants can apply statistical learning to multiple languages (Antovich et al., 2020). Additionally, bilingual infants retain sensitivity to pitch contours & tonal variations longer than monolingual infants, whose flexibility will begin to become static after the first year of their life. This means struggling to process tonal languages such as Thai & Mandarin – both being languages where the pitch of the voice changes the meaning of the word (Graf Estes et al., 2015).
Page updated
Report abuse