Computational Modeling

A neural network An important aspect of the types of systems we study is that they are dynamic and complex. Because of this, one of the major difficulties in studying them is that we can often only observe the end result of the cognitive or perceptual process we are interested in. Ideally, we would like to know not only the properties of the behavior we observe, but also what types of information processing are involved and how neural processes give rise to that behavior. Brain imaging and neurophysiological methods provide us with important information about the neural basis of these processes at both gross and cellular levels, but they do not provide us with a functional description of how neural activity translates into the behaviors we observe. Computational modeling fills the void between these two levels of analysis by allowing us to construct models that describe the functional relationship between brain and behavior. Cognitive and developmental scientists use modeling as a tool for testing hypotheses about how complex dynamic systems are organized to give rise to observable behavior.

Our lab uses several different types of computational models to study a variety of research topics. These range from neural networks to parametric statistical models to non-mechanistic models. Modeling plays a large role in many of our research projects and gives us a method of answering questions and making predictions about processes ranging from word learning to recognizing speech sounds.


Computational models we use:

For more information, see:

McMurray, B. (2000). Connectionism for ...er... linguists. In Crosswhite, K., and McDonough, J., eds., The University of Rochester Working Papers in the Language Sciences, 2000(1), 72-96. PDF

TRACE

The TRACE Model (McClelland & Elman, 1986) is a classic model of word recognition. TRACE envisions word recognition as a process of gradually building activation (or evidence) for a number of candidate words over time. As activation accumulates, these words compete and the winner is the ultimate word that was recognized. For example, while hearing "beatle", TRACE will initially build activation for both "beatle" and "beaker". However, over time, "beatle" inhibits "beaker" until it alone is active.

TRACE is a dynamical system. Activation flows between the input and phonemic units, between phoneme units and words, and back down from words to phonemes over time, as the network gradually settles on a winner. For this reason, it has been very successful in modeling the temporal dynamics of word recognition that we can measure using the Visual World Paradigm. Perhaps the most controversial aspect of TRACE has been the fact that it posits feedback from words to sublexical units. The MACLab has been testing this hypothesis in the Compensation for Coarticulation project and has so far found support for this hypothesis. In this same light, we are applying the same theoretical framework (though not this exact model) to music with some startling results. We've also been using TRACE to understand how the system is sensitive to and retains fine-grained detail over time. More recently, we have begun using this model to understand the possible computational underpinnings of specific language impairment. By changing the parameters that control TRACE's operations, (and matching them to VWP data collected from SLI children), we can get a clearer understanding of what processes may be different in these children.

Recently, thanks to the efforts of Jim Magnuson, there is now a java implementation of TRACE (jTRACE) that makes it simple for anyone to run their own simulations and apply TRACE to their own data. The language sciences owe Jim a huge debt of gratitute for making this freely available.

Students
Vicki Samelson 		Collaborators
J. Bruce Tomblin Michael Tanenhaus (University of Rochester)
Richard Aslin (University of Rochester)
References
McClelland, J. and Elman, J. (1986) The TRACE model of speech perception. Cognitive Psychology, 18(1), 1-86.

Back to top

Mixture of Gaussians (MOG)

Mixture of Gaussians (MOG) model

A MOG is a standard statistical approach for solving problems that involve determining the distribution of a set of data. The model consists of a set of Gaussian distributions (the mixture) defined by their means and variances. The model is trained by giving it input on the dataset to be learned. During training, the parameters of the Gaussians are updated using an algorithm for maximum likelihood estimation in order to match the distribution of the data.

We are using the MOG approach to model the development of speech categories. The acoutsic cues that define phonetically meaningful categories are distinguished by multimodal Gaussian distributions. This provides an excellent dataset for training an MOG. We have found that, for a variety of different phonetic contrasts across different languages, the MOG is able to learn the distributions corresponding to the phonetic contrast it is trained on. The figure on the right shows the process of training the model on a distribution similar to the one found for VOT in English. This tells us how a data-driven learning mechanism can account for the formation of speech sound categories and it allows us to exmaine the dynamics of this learning process over time.

The MOG model is used in the projects on statistical learning of speech categories and the projects on cue integration in perception and over development.


Students
Joe Toscano 		Collaborators
Richard Aslin

Back to top



Site maintained by Bob McMurray, Cheyenne Munson, and Joe Toscano.