The idea that babies have capabilities beyond our recognition is one that has been expressed in many different forms across the years, such as in children’s television shows (e.g., Nickelodeon’s Rugrats and a few ill-fated movies (Baby’s Day Out, Baby Geniuses, Superbabies: Baby Geniuses 2). A common theme in these shows and movies is that adults severely underestimate the intelligence of their children.
While babies are not able to take over the world, these fictional themes are backed by a bit of truth in the form of surprising experimental findings: Young infants are able to learn and recognize patterns in language, even before they are able to articulate the patterns themselves. These learning processes in infants can often exceed our own (adult) expectations.
One particularly influential study that has demonstrated this type of learning in infants is by Marcus and colleagues in which they demonstrate simple grammar-like rule learning in 7-month-old babies. The researchers constructed fake “words” with regularities in syllable repetition, such as “li ti li” and presented the spoken stimuli to the infants for familiarization. After the infants were familiarized with the artificial words, researchers presented a new set of spoken “words”, half of which were from the artificial grammar from the familiarization task (e.g., “wo fe wo”; ABA) and half of which were from a separate and unfamiliar artificial grammar (e.g., “wo fe fe”; ABB). Each set of artificial words was played from randomly-chosen sound speakers in the room, and the researchers recorded the amount of time that infants looked at each speaker (the Head Turn Preference Procedure). Marcus and colleagues found significant differences in the amount of time that the infants would gaze at speakers playing words from each grammar, indicating a familiarity with the artificial grammar. Thus, it seems the 7-month-old infants were able to pick up on enough regularities in the speech patterns to discriminate between two artificial grammars.
This study, and others like it, has spurred an important question in developmental research: What is the underlying mechanism by which young infants are able to learn basic rules? This has led to a number of different theoretical and computational accounts of basic rule-learning in infants, such as domain-general statistical learning accounts, simple recurrent networks, auto-associator models, Bayesian models, and many others.
With the increase of research in basic rule learning comes a separate, but important issue: Despite the emergence of many different types of computational models, there is often little overlap between the theoretical and practical elements of these models, and there is no consensus amongst researchers on how to evaluate the efficacy of these different models against each other. The lack of commonalities amongst these different accounts of rule learning in infants has slowed progress toward a consensus on the exact mechanisms underlying the effect.
In a recent paper in the Psychonomic Society’s Psychonomic Bulletin & Review Raquel G. Alhama and Willem Zuidema (pictured below) review the state of computational accounts of rule learning in infants. They go on to propose 9 guidelines that outline the theoretical issues that future models should address. While I will not list their desiderata here (the authors do it best!), what follows is an overview of their recommendations.
Consider Marr’s Levels of Analysis
Alhama and Zuidema argue that many overall differences can be characterized using David Marr’s topology for computational models of cognition. At one level (the computational level), theories are constructed to simply define the problems a system faces and attempts to overcome. On a different level (the algorithmic/representational level), theories are constructed to define a mechanistic account of how the system processes information and representations. On the last level (the implementational level), theories are constructed to define the physical realization of a system.
Different models of basic rule learning represent different levels of Marr’s taxonomy. For example, Marcus and colleagues’ recommendations for such a model can be characterized as being a computational-level explanation, where the model defines the problem space but does not offer many details that describe the mechanisms needed to overcome these problems. Neural network accounts, on the other hand, represent an entirely different level of analysis. Because most neural network models incorporate sub-symbolic representations, they are closer to implementation-level models. Frank and Tenenbaum’s Bayesian model incorporates features of both algorithmic and computational accounts of rule learning, but like other models, it does not share features with models from different levels of analysis, making comparisons difficult.
With this in mind, Alhama and Zuidema recommend that researchers construct models that can be compared across Marr’s levels of analysis by integrating neural-symbolic components, and also by analyzing internal representations of implementational-level models (i.e., neural networks) and how these representations can incorporate the rules that infants seem to learn. Better comparisons between these levels will also require investigation into the features that should be represented in computational models so that researchers can identify the amount of overlap in the input data that is required to replicate empirical findings, as well as the features that are over- or under-weighted during learning. Garnering consensus on how language is represented in computational models will also allow researchers to evaluate their models using similar criteria.
The Role of Prior Experience
A difficult task in modeling human performance is accounting for differences that experiments cannot control, such as prior experience. This is equally (if not more) true in infants, where the influence of prior exposure to spoken language on simple learning tasks is unknown. Some models of simple rule learning in infants account for contextual variables (e.g., prior experience), the vast majority do not. Models such as neural networks often start with unconstrained parameters that are trained over thousands of trials.
Alhama and Zuidema’s first recommendation on this matter is for researchers to incorporate relevant prior experience in the initial state of the model in order to simulate how infants learn over the course of the task. The incorporation of prior experience can also help simulate how infants learn simple rules over a relatively short number of trials, minimizing the need for model training approaches that are not developmentally plausible. Finally, the authors propose the use of a continuous representation of time, arguing that representations of “time” in prior models, where time is represented as discrete time-steps, are not plausible or realistic.
Conclusions
Babies may not be able to take over the world, and they may not be “geniuses” in the traditional meaning of the word, but their remarkable ability to learn basic rules and generalize them remains an exciting avenue for developmental research and computational modeling. Alhama and Zuidema’s review effectively synthesizes conclusions from different modeling traditions and provides guidelines for future research.
Psychonomics article featured in this post:
Alhama, R. G., & Zuidema, W. (2019). A review of computational models of basic rule learning: The neural-symbolic debate and beyond. Psychonomic Bulletin & Review, 26(4), 1174-1194. doi: 10.3758/s13423-019-01602-z.