Research

I am interested in the characterization of the structure of semantic networks and how this interacts with processes operating on these structures. The statistical structure of large-scale semantic networks, such as word association, WordNet, and Roget's thesaurus are characterized by a small-world structure with short average path lengths, strong local clustering and scale-free patterns of connectivity with most nodes having relatively few connections joined together through a small number of hubs with many connections. This pattern of connectivity is difficult to explain on the basis of Euclidian spaces.

.

Steyvers, M., & Tenenbaum, J. (2005). The Large Scale Structure of Semantic Networks: Statistical Analyses and a Model of Semantic Growth. Cognitive Science, 29(1), 41-78. 

I am also interested in developing probabilistic topic models to explain these structured semantic representations. Probabilistic topic models were introduced in the domain of machine learning and information retrieval under more technical names such as probabilistic Latent Semantic Indexing (pLSI) and Latent Dirichlet Allocation (LDA). These models are based on the idea that documents are mixtures of topics, where a topic is a probability distribution over words. A topic model is a generative model for documents: it specifies a simple probabilistic procedure by which documents can be generated. To make a new document, one chooses a distribution over topics. Then, for each word in that document, one chooses a topic at random according to this distribution, and draws a word from that topic. An efficient Markov Chain Monte Carlo (MCMC) technique can be used to infer the set of topics that were responsible for generating a collection of documents.

Steyvers, M. & Griffiths, T. (in press). Probabilistic topic models. In T. Landauer, D McNamara, S. Dennis, and W. Kintsch (eds), Latent Semantic Analysis: A Road to Meaning. Laurence Erlbaum

Griffiths, T., & Steyvers, M. (2004). Finding Scientific Topics. Proceedings of the National Academy of Sciences, 101 (suppl. 1), 5228-5235. 

Our work with these probabilistic topic models has shown that the large-scale structure of the model's representation has statistical properties that correspond well with those of semantic networks produced by humans:

Griffiths, T.L., & Steyvers, M. (2002). Prediction and semantic association. In: Advances in Neural Information Processing Systems, 15.  

Griffiths, T.L., & Steyvers, M. (2002). A probabilistic approach to semantic representation. In: Proceedings of the Twenty-Fourth Annual Conference of Cognitive Science Society. George Mason University, Fairfax, VA. 

We are currently extending this model to make precise predictions for episodic memory tasks such as recognition and recall where semantic properties play an important role (e.g. "false memory").

For a demo of the Gibbs sampler as applied to word-sense disambiguation, check out this demo:

demo of word sense disambiguation (--> click on "next iteration" to see through the iterations of the Gibbs sampler). Colors and numbers indicate the assignment of words to topics. Note the ambiguous word "PLAY". This word, over the course of learning, is assigned to different topics that highlight the different senses.

Another research direction is to formulate increasingly more structured representations that will be useful in both cognitive science as well as machine learning/information retrieval. For example, standard topic models do not make any assumptions about the order of words as they appear in documents. This is known as the bag-of-words assumption, and is common to many statistical models of language.

Of course, word-order information might contain important cues to the content of a document and this information is not utilized by the model. Griffiths, Steyvers, Blei, and Tenenbaum (2005) present an extension of the topic model that is sensitive to word-order and automatically learns which words characterize the content of a document and which words are mere function words that are needed to form a sentence. This research is both useful in the domain of information retrieval to automatically identify relevant content words as well cognitive science, to characterize the online processing of syntactic and semantic information. 

Griffiths, T.L., & Steyvers, M.,  Blei, D.M., & Tenenbaum, J.B. (2005). Integrating Topics and Syntax. In: Advances in Neural Information Processing Systems, 17.  

Recently, with Padhraic Smyth (ICS, UC Irvine),  we developed the author-topic model, an extension of the topic model that integrates authorship information with content (e.g., Steyvers, Smyth, Rosen-Zvi, and Griffiths, 2004; Rosen-Zvi, Griffiths, Steyvers, and Smyth, 2004). Instead of associating each document with a distribution over topics, the author-topic model associates each author with a distribution over topics and assumes each multi-authored document expresses a mixture of the authors’ topic mixtures. Using a large corpus of 500,000 Enron emails recently released by the Justice department, we applied the model to learn what topics Enron employees wrote about. We also applied the model to a large collection of CiteSeer abstracts to learn about the main researchers and topic trends in different areas of computer science research.

Steyvers, M., Smyth, P., Rosen-Zvi, M., & Griffiths, T. (2004). Probabilistic Author-Topic Models for Information Discovery. The Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Seattle, Washington.  

Rosen-Zvi, M., Griffiths T., Steyvers, M., & Smyth, P. (2004). The Author-Topic Model for Authors and Documents. In 20th Conference on Uncertainty in Artificial Intelligence. Banff, Canada

Based on the derived representations, statistical inference can be used to pose the following queries: 1) what topics does a given author write about? 2) given a document, what author is most likely to have written about the topics expressed in the document? 3) how broad is the research of an author as expressed by the topics distribution? 4) how unusual is a paper for a given author? and 5), what author is similar to a given author? These queries are not only relevant when exploring a scientific domain or developing an author profile, but also in practical situations when finding targets for funding or assigning reviewers to a paper or grant proposal.

Understanding how people make decisions, from very simple perceptual to complex cognitive decisions, is an important area of research in psychology. In this collaborative research with Scott Brown, we examine decision-making behavior in dynamically changing decision contexts. Real world decision contexts are continually varying, and good decision makers must continually adjust their behavior to track environmental changes.

Consider the case of a military observer making decisions about the identity (friend vs. enemy) of noisy stimuli from reconnaissance pictures. The difficulty of these decisions will change throughout the task, as more or less clear pictures are used, or more or less uniform terrain is observed. An ideal observer must dynamically adjust their decision making process to reflect changes in the environment. For example, if it becomes easier to identify friendly stimuli in new terrain, observers should relax their criterion for identifying enemy stimuli.

Previous research has most often assumed static models for decision making that ignore sequential dependencies between environments and the effect of history on current decision making. Some more recent research has focused on dynamic models of decision making, in which certain parameters of the decision process are allowed to vary from decision to decision. However, this research typically makes another static assumption: namely that the environment is stationary.

In our research, we develop new experimental paradigms in which dynamic decision making environments force participants to change their decision making processes in order to remain (approximately) ideal. This paradigm allows us to observe decision makers tracking changes in the environment. Currently, one demonstration program is near completion in which an aircraft is flying through a canyon environment (see screenshots below). During the flight, the aircraft is attacked by incoming missiles. There are two types of incoming missiles and the participant in the experiment has to make a quick decision about the correct type of missile in order to choose the appropriate counter-measures. Building on this demo, the goal is to measure decision speed in natural environments and also to measure how well participants adapt to changes in the decision making environment (e.g., by making the two types of missiles more or less similar during the course of the experiment).

We are also developing two models for the decision process in dynamic environments. One model is an ideal observer system in which statistical evidence for a changed environment is weighed in optimal fashion against evidence for a stable environment. The ideal observer analysis results in estimates for the (optimal) number of trials it takes to detect and adjust to new decision environments. Our other model is a dynamic SDT model that estimates how long it actually takes for individual decision makers to adapt to novel decision environments. By comparing predictions from the ideal observer model to the parameter estimates from the decision model (from individual decision makers), we can quantify the degree of mismatch between ideal and actual observer.
 

Brown, S.D., & Steyvers, M. (2005). The Dynamics of Experimentally Induced Criterion Shifts. Journal of Experimental Psychology: Learning, Memory & Cognition, 31(4), 587-599.

Steyvers, M., & Brown, S. (in press). Prediction and Change Detection. In: Advances in Neural Information Processing Systems, 19.

In research on episodic, lexical and semantic memory, I am trying to understand both the processes that underlie tasks such as recognition, recall, and lexical decision, as well as the representations that support these processes. The basis for this research is the theory of REM (Retrieving Effectively from Memory) developed by Richard M. Shiffrin at Indiana University and myself (Shiffrin & Steyvers, 1997). The model was able to handle some basic recognition memory phenomena that have been difficult to handle with extant models. For example, in order to model mirror effects – the finding that many experimental manipulations (e.g., list length, strength and word frequency) simultaneously raise hit rates and lower false alarm rates – many models adjust response thresholds without explaining how and why these thresholds vary as they do. In contrast, in the REM theory, memory decisions are based on a Bayesian inference process that contrasts the evidence for one decision (e.g. “old”) with the evidence for another decision (e.g. “new”), which naturally leads to mirror effects.

Shiffrin, R.M. & Steyvers, M. (1997). A model for recognition memory: REM: Retrieving Effectively from Memory. Psychonomic Bulletin & Review, 4 (2), 145-166. 

Shiffrin, R. M., & Steyvers, M. (1998). The effectiveness of retrieval from memory. In M. Oaksford & N. Chater (Eds.). Rational models of cognition. (pp. 73-95), Oxford, England: Oxford University Press. 

Steyvers, M. (2000). Modeling semantic and orthographic similarity effects on memory for individual words. Dissertation, Psychology Department, Indiana University. Formatted for 55 pages 

Wagenmakers, Steyvers, Raaijmakers, Shiffrin, van Rijn, & Zeelenberg (2004) have developed a model for lexical decision based on the same principles as the REM model. The evidence for one decision (“WORD”) is contrasted with the evidence for the other decision (“NONWORD”) on the basis of a Bayesian inference process. The long-term goal in the REM framework is to develop a unified account of episodic, lexical and semantic memory.

Wagenmakers, E.J.M., Steyvers, M., Raaijmakers, J.G.W., Shiffrin, R.M., van Rijn, H., & Zeelenberg, R. (2004). A Model for Evidence Accumulation in the Lexical Decision Task. Cognitive Psychology, 48, 332-367. 

Steyvers, M., Wagenmakers, E.J.M., Shiffrin, R.M., Zeelenberg, R., & Raaijmakers, J.G.W. (2001). A Bayesian model for the time-course of lexical processing. In: Proceedings of the Fourth International Conference on Cognitive Modeling. George Mason University, Fairfax, VA. 

Finally, I am interested in explaining word frequency effects in recognition memory in terms of aspects of words other than word frequency perse. For example, what is the effect of the number of contexts a word has appeared in, irrespective of the number of total times a word has appeared? Are words that appear in few contexts more memorable? Also, what is the effect of rare or common letter features within words? Are words with rare features more memorable?

Steyvers, M., & Malmberg, K. (2003). The effect of normative context variability on recognition memory. Journal of Experimental Psychology: Learning, Memory, & Cognition, 29(5), 760-766. 

Malmberg, K. J., Steyvers, M., Stephens, J. D., & Shiffrin, R.M. (2002).  Feature-frequency effects in recognition memory.  Memory & Cognition, 30(4), 607-613. 
 

One theme in my research is finding appropriate mental representations for stimuli often used in cognitive tasks (e.g., words, faces, visual scenes). Traditional multidimensional scaling techniques place stimuli as points in a multidimensional space with similarity inversely related to the distances between points. Typically, pairwise similarity judgments are used to infer these multidimensional representations but I have shown how to extend this framework to learn semantic spaces for words based on word association (Steyvers, Shiffrin, & Nelson, 2004) and perceptual representations for faces based on physical features as well as similarity ratings (Steyvers & Busey, 2000).

Steyvers, M. (2002). Multidimensional Scaling. In: Encyclopedia of Cognitive Science. Nature Publishing Group, London, UK.

Steyvers, M., & Busey, T. (2000). Predicting Similarity Ratings to Faces using Physical Descriptions. In M. Wenger, & J. Townsend (Eds.), Computational, geometric, and process perspectives on facial cognition: Contexts and challenges. Lawrence Erlbaum Associates.

Steyvers, M., Shiffrin, R.M., & Nelson, D.L. (2004). Word Association Spaces for Predicting Semantic Similarity Effects in Episodic Memory. In A. Healy (Ed.), Experimental Cognitive Psychology and its Applications. 

In research on causal reasoning, I study peoples ability to infer causal structure from both observation and intervention, and to choose informative interventions on the basis of purely observational data. I develop computational models of how people infer causal structure from data and how they plan intervention experiments, based on the representational framework of causal Bayesian networks and the inferential principles of optimal Bayesian decision-making and maximizing expected information gain.

Steyvers, M., Tenenbaum, J., Wagenmakers, E.J., Blum, B. (2003). Inferring Causal Networks from Observations and Interventions. Cognitive Science, 27, 453-489. 

Some of my work has involved testing subjects on the web (see website here) and comparing their performance to subjects tested in the lab. Part of the success in luring web subjects to my site may have something to do with the "Hall of Fame" where the subjects performance score is directly posted and compared to other subjects. So far, I have published two papers where I show detailed comparisons between web and lab subjects.

Steyvers, M., & Malmberg, K. (2003). The effect of normative context variability on recognition memory. Journal of Experimental Psychology: Learning, Memory, & Cognition, 29(5), 760-766. 

Steyvers, M., Tenenbaum, J., Wagenmakers, E.J., Blum, B. (2003). Inferring Causal Networks from Observations and Interventions. Cognitive Science, 27, 453-489.