Research Interests

My research interests center around language learning and the interaction between language and everything else. To study these questions I use both empirical methods with children and adults and computational modeling using connectionist networks. I believe that the strength of this approach is the interplay between experimental results and predictions from the model - an empirical finding may suggest a computational model and in turn this computational model makes predictions that suggest more experiments...

Language basics			Nouns		Relations
Word Segmentation	Rule Learning	Symbol Use	Cross-linguistic studies	Simulations	Relational Correlations	Spatial Relations

Language Basics - Before children can talk, they show abilities that can be thought of as necessary for the task of learning language. These are abilities such as finding words and abstract regularities in an auditory stream and mapping abstract symbols to concrete objects.

Word Segmentation - how do children segment the continuous auditory stream that is a linguistic utterance into meaningful segments, such as words? Recent evidence has shown that infants of only 8 months of age are able to perform this task by using the transitional probabilities between syllables. We show a simulation of this task within the Playpen framework.

Gasser, M. & Colunga, E. (1999). How babies learn to find words. Proceedings of the International Conference on Cognitive Science, 2, 277--281. Postscript version
Rule Learning - Even more impressive than 8-month-olds ability to find word boundaries is their apparent ability to learn abstract "grammar" rules embodied in an auditory stream that follows them. These are rules such as ABB (the last two "words" are the same) or ABA (the first and the last "words" are the same). Infants' ability to learn these sorts of abstract regularities seems to indicate their access to things such as variables and rules, we present a variable-less, rule-less model of this phenomenon, within the Playpen framework. We test predictions of this model in a visual version of the infant's task.
Emergence of Symbol Use - The ability to use abstract symbols, such as words, has been called by some the "hallmark of human cognition". However, there is evidence that suggests that this ability is not present in humans from the start, but rather that it develops. We concentrate on words as symbols. We are interested in the following questions: how do words become symbols? how do words come to refer? why are words used as names for things? and what are the characteristics of a learner that can come to use words as symbols? We use empirical tests on infants and connectionist network simulations to explore these questions. We suggest that words, as symbols, get their special status by virtue of the way they are used.

Nouns - Once they start learning language, children seem to be expert noun learners - they can guess what a noun refers to, even if they have only heard it once. To explain this proficiency people have proposed several mechanisms in the form of constraints or biases, but where do these biases come from? Are they innate or learned? What is the mechanism that gets them into children's heads?

Cross-linguistic differences and similarities - does the language children are learning affect the way they learn language? Current evidence suggests both universals and specific differences in the word-learning biases shown by speakers of different languages. To try to puzzle this out, we have been conducting studies with Japanese speakers, Spanish speakers, and English speakers. We suggest that similarities and differences in the structures of the languages they speak can explain the similarities and the differences in children's word-learning biases.

Colunga, E., Gasser, M. & Smith, L.B. (submitted) The Attention to Different Cues in Naming Tasks Depends on the Structure of the Language Being Learned – submitted to the 23^rd Annual Conference of the Cognitive Science Society PDF version

Colunga, E., & Smith, L.B. (2000). Learning to Learn Words: A Cross-Linguistic Study of the Shape and Material Biases. Proceedings of the 24th Annual Boston University Conference on Language Development. Pages Postscript version
Simulations - do children learn to learn words? research has shown that there are regularities in the language that roughly map the children's expectations about the meanings of nouns. Do children learn what to expect from these regularities? What kind of a learner could do such a thing? We start with a simple statistical learner - a neural network- and show that given a structured vocabulary this learner will learn not only the words in the vocabulary, but also the regularities of the underlying structure. This model also accounts for both similarities and differences found between languages and makes verifiable (and verified!) predictions about children's noun learning.

Relations - Learning names for relations seems, at least logically, secondary to learning names for objects -- understanding a relation implies understanding the objects that participate in it. However, even young infants understand some relational concepts, such as those involved in situations of occlusion, containment or support. How do these concepts develop? What is the role of action? What is the role of language?

Relational correlations -- Where do relations come from? What are they made out of? Playpen is a connectionist network with units (Relation Units) that are about two things. This allows us to represent "micro-relations". We propose that relational concepts are formed by correlations of these micro-relations. We use this framework to model diverse things such as the learning of rules, relational words and spatial relations.

Gasser, M., & Colunga, E. (1998). Where do relations come from? Indiana University Cognitive Science Program, Technical Report #221. Bloomington, IN. Postscript version

Colunga, E. & Gasser, M. (1998). Linguistic relativity and word acquisition: a computational approach. Proceedings of the Annual Conference of the Cognitive Science Society, 20, 244–249. Postscript version
Spatial relations -- Studying spatial relations seems like a good place to start studying the development of relations because they are concrete. In addition, research has shown that even very young infants seem to understand some of the "rules" of certain spatial relations. In this project we look at what infants know about putting things into containers and about relative size. We look at the effects on babies' and neural network's performance of experience putting things in and of knowing words like "in", "little", "big" and "bigger".