Co-evolution of event memory and knowledge
Visual Inference and Visual Perception
Optimal and Non-optimal Decision Making
Explanations for Quantitative Data
Diffusion Modeling of Perception and Memory
General and Specific Context in Event Memory
Model Selection and Data Priors
Rationality and Game Theory
Identifying Visual Features
Visual Masking and Decision Making
Assessing Mental Categories with MCMC
Model Selection via Individual or Group Data Analysis
Stereotype Threat Inhibits Learning
Knowledge grows from experienced events, and events are encoded through knowledge. We are exploring and modeling this process. Initial research by Dr. Angela Nelson (USC, former graduate student) trained initially unknown Chinese characters for two weeks to form knowledge, and then tested pseudo-lexical decision, episodic recognition memory, and two-alternative forced-choice perception (article in preparation). The roles of frequency of training were elucidated: Two major effects of frequency are seen: 1) The contexts in which items are experienced has important effects. Items occurring in temporal and spatial proximity lend features to each other’s knowledge traces, causing the knowledge representations to grow more similar. 2) When context is controlled (in our studies characters are trained in isolation), frequency of event occurrence nevertheless has potent effects. Our model of the co-evolution process and the expression of the knowledge representation in later tests incorporates both factors. Current research headed by graduate student Greg Cox (with other collaborators Professor Mike Jones and graduate student Brendan Johns) trains pairs of characters that are studied in ‘isolation’ from other pairs. Un-noticed by participants (in most cases) some pairs overlap in ways that link together to produce more complex knowledge structures: linear orders and trees. We then carry out transfer tests to assess the expression of this structural knowledge in priming of knowledge retrieval, in episodic memory, in perceptual identification, and in memory for co-occurrence during study. This first study asks how implicit structure in knowledge affects retrieval of knowledge, event memory, and perception. Later studies will make more explicit during study the need to store structure among items. (Funded by AFOSR)
Paper (PDF) - Nelson, A.B. Examining the co-evolution of knowledge and event memory. Dissertation.
Research started with Professor Dave Huber (UCSD, former graduate student), continued with Dr. Christoph Weidemann, (U Penn, former graduate student), and pursued presently with postdoctoral researcher Dr. Stephen Denton explores and models the inference processes by which we form our percept of the visual world. In particular, we model the process by which evidence for the presence of a particular feature is discounted when the presence of that feature can be attributed to another source. The project investigates the extraction, temporal course, masking , neural mechanisms, and use of visual features in perception tasks. Bayesian and neural net models are introduced and tested. Earlier research showed that features of a prime stimulus are confused with features of an immediately following target stimulus. Evidence due to those prime features is discounted slightly for short and unattended primes, producing positive priming (a tendency to see and choose something resembling the prime), and is discounted strongly for long and attended primes, producing negative priming (a tendency to see and choose something that does not resemble the prime). The present research studies the effect of moving the eyes between the prime and the target presentation: To what degree are features of the prime and target confused? To what degree are confused features discounted? (Funded by NSF, in a joint grant with Dave Huber).
Various papers on this topic may be found on the above links. The Psychological Review paper that initiated this project may be found below:
Paper (PDF) - Huber et al. (2003) Visual Inference and Visual Perception.
There have been many successes in modeling decision making in perception, memory and concept formation tasks with adaptive, rational, optimal approaches, most often employing Bayesian models. There have been many failures to employ ‘rational’ models for more complex tasks, often employing verbal scenarios (e.g. the conjunction fallacy as seen in the ‘Linda’ problem introduced by Kahneman and Tversky). We attempt to study the mechanisms causing the different results in these two subfields of decision making by employing a common perceptual task: Os are asked to combine evidence from two halves of a face, when the two face halves are presented in normal, upside-down, and split arrangements. Somewhat surprisingly, quite non-optimal decision making is found in all cases. In addition, the degree of non-optimality is a bit higher when faces are split (presumably because the usual perceptual face perception processes are interrupted). This research project is under the present direction of graduate student Jared Hotaling and is a joint project with Professors Jerry Busemeyer (IU) and Professor Andrew Cohen (UMass, former postdoctoral researcher).
Paper (PDF) - Hotaling (2009). Optimal and Non-optimal Decision Making: Draft of Article Submission.
Scientists almost invariably begin by studying two-dimensional plots of their quantitative data, even when these plots represent marginals extracted from high dimensional data. These data plots are often quite noisy, but nonetheless play an important role in the scientist’s formulation of explanations, models, and theories. We therefore explore the way scientists balance fit, complexity, and other factors, in coming up with good (possibly causal) explanations for the plotted data. We do so by asking Os to estimate best functional explanations (by in effect drawing best functions on the data). Both the raw data and the estimated functions are analyzed with Bayesian model selection techniques (a combination of selection among a plausible set of simple basis functions that act as priors and a tendency to ‘track the data smoothly’). The analysis is implemented both in standard Bayesian formulations and in (equivalent) Gaussian Process formulations. This research in directed by postdoctoral researcher Dr. Dan Little.
Paper (PDF) - Little (2009). Explanations for Quantitative Data: Draft of Research in Progress.
When information about correlated events is encountered, it most often occurs in richly structured situations where there is ambiguity about what is correlated with what. For example, in early development, a spoken word and its object referent may be ambiguous because the situation contains many other spoken words and many possibly referents. Such ambiguity of course tends to be resolved across multiple situations containing the association, because the associated pair tends to co-occur statistically often. Professor Chen Yu (IU), former graduate student Dr. Krystal Klein, and present graduate student George Kachergis are exploring and modeling the processes and mechanisms of statistical learning, studying such factors as rate of learning, degree of ambiguity, intentional vs. incidental learning, the effects of prior knowledge, and the statistical structure of the environment. Relevant papers can be found on the links in this paragraph.
A project in its very earliest stages explores the role of noise in diffusion and race models of decision making in memory and perception tasks. This is joint research with Professor Roger Ratcliff (Ohio State) and Chris Donkin (U. Newcastle, joining our laboratory as a postdoctoral researcher in January 2010).
Recent years have seen many demonstrations of the importance of local context in the formation and retrieval of episodic memories. One factor is general list context, not usually manipulated experimentally, but often studied indirectly. Dr. Krystal Klein (former graduate student) explored this factor with a judgment of relative recency. A second factor is specific context that is more easily varied experimentally. For example, memory for a list of words is strongly affected by the words that are nearby during list study, the relation to words that are nearby at test, and the semantic and other similarity of such words at both study and test. This is the subject of research presently directed by graduate student George Kachergis.
What makes a good model, and how should one choose among competing models? This is a difficult question that has answers with both qualitative and quantitative components. The last ten to fifteen years especially have seen a great deal of progress on statistical comparison of models, particularly through use the techniques of Minimum Description Length (in the form of Normalized Maximum Likelihood, or NML) and Bayesian Model Selection (or BMS). Former postdoctoral researcher Dr. Woojae Kim co-directed a continuing project that seeks to: 1) clarify the conceptual underpinnings of model selection techniques, particularly elucidating the close underlying relation between NML and BMS, and 2) introduce methods to incorporate ‘data priors’ (our prior beliefs about expected patterns of data outcomes) into BMS and NML. I also consider how to incorporate inference about the validity of data sets, and how to incorporate ‘higher level’ considerations into the model evaluation picture. I give here two powerpoints, one a draft of an address given to the 2010 Judgment and Decision Making (JDM) meetings, and a second with some elaboration and additional technical material.
Paper (PDF) - Cohen, A. L., Sanborn, A. N., & Shiffrin, R. M. (2008). Model Evaluation Using Grouped or Individual Data. Psychological Bulletin & Review, 15, 692-712
Rationality. I use analysis of famous paradoxes to justify the claim that rationality is not a normative concept, but rather a sufficiently large social consensus of people judged to be sufficiently expert.
Game Theory. In the Prisoner's Dilemma each player gains more by defecting whatever
the other player chooses: If both therefore defect (the Nash Equilibrium) they get a jointly
bad result. I argue that rational players should cooperate and obtain a jointly good result
(not because they have the goal of maximizing joint return, but because such a strategy gives
a 'selfish' player a better return). Similarly consider a two trial 'centipede game': The
first player can STOP and both get a return of zero, or can GO. Then the second player can
STOP, getting 10 and giving the first player -1, or GO, giving both players 9. Usual game
theory stipulates the following: Knowing the second player will STOP, the first player will
STOP giving both zero. I argue that two rational players should both GO, giving both 9.
The present talk starts with such a premise and searches for algorithms that will find such 'rational' solutions for jointly rational players in multiplayer sequential decision games of arbitrary complexity and length. Such a solution plays a role similar to 'ideal observer theory', and provides a baseline to which human behavior can be compared. The basic assumptions: 1) All players are rational and know all players are rational. 2) All players are selfish and try to maximize personal return. 3) Players do not know other players utilities so only ordinal returns matter: More is better but how much more is irrelevant.
For two player games a unique rational solution is always available, and a relatively simple algorithm finds it. When there are more than two players, the complexities expand super-exponentially: Rational solutions do not always exist (as shown by examples). An extension of the two player algorithm will find an important class of solutions, but not all. I discuss attempts to find ways to determine all cases that do have rational solutions, and attempts to find algorithms that extend the class of cases for which solutions do exist. This is joint research with Professor Michael Lee (UCI) and UCI graduate student Shunan Zhang.
Professor Jason Gold (IU) and Andrew Cohen (UMass, former postdoctoral researcher) co-direct(ed) a project to use classification image and machine learning techniques to extract features visual features that observers use to identify and perceive objects in the environment. Subjects are asked to judge whether a very noisy visual image does or does not contain a particular visual object. Machine learning techniques are used to extract visual patterns that are the most highly correlated with the classification responses, and these arte candidate visual features that the observer may be using to carry out the task.
Papers (PDF) -
Gold, J. M., Cohen, A. L., and Shiffrin, R. M. (2006). Visual noise reveals category representations. Psychological Bulletin & Review, 13, 649-655.
Cohen, A. L., Shiffrin, R. M., Gold, J. M., Ross, D. A., & Ross, M. G. (2007). Inducing features from visual noise. Journal of Vision, 7(8):15, 1-14.
Dr. Adam Sanborn (Gatsby Institute, former graduate student) and Ken Malmberg (former Postdoctoral Researcher, now Associate Professor, University of South Florida) studied identification decisions when brief visual targets (words) are followed by pattern masks. Such masks introduce sufficient noise into the decision process that observers abandon the use of low level physical features such as form and color, and instead use higher level features that presumably are sometimes extracted before the mask interrupts processing.
Paper (PDF) - Sanborn, A., Malmberg, K, and Shiffrin, R. M. (2004). High-level effects of masking on perceptual identification. Vision Research, 44, 1427-1436
Dr. Adam Sanborn and Professor Tom Griffiths (UC Berkeley) have observers judge which of two candidates is a better member of a mental category. Candidates are generated with a process found in Markov Chain Monte Carlo, so that the chain of responses and presented members converges on the modal category representation.
Paper (PDF) - Sanborn, A., Griffiths, T, & Shiffrin, R.M.. Uncovering mental representations with Markov Chain Monte Carlo. Cognitive Psychology.
Dr. Adam Sanborn and Professor Andrew Cohen studied conditions that make analysis by individuals or analysis by grouped data produce superior selection of the correct generating model. For many complex models, small amounts of data make analysis by group superior
Paper (PDF) - Cohen, A. L., Sanborn, A. N., & Shiffrin, R. M. (2008). Model evaluation using grouped or individual data. Psychological Bulletin & Review, 15, 692-712.
Research headed by IU Assistant Professor B.J. Rydell explores the effects of stereotype threat on the learning of automatic attention (“popout’) to fixed targets in visual search. Observers are women in two groups: those made aware that women are a group that underperforms in visual tasks, and those not made aware. All perform one session of searching for one of several fixed targets (Chinese characters) in a subsequent display of different Chinese character foils. The display has two or four characters and half the time contains the target. The control group showed learning: The visual search slope dropped regularly and steadily across the blocks within session. The stereotype threat group showed no change in slope. The next study is testing the possibility that learning had taken place for the observers under threat, but that expression of this learning had been suppressed (perhaps by a strategy to continue effortful serial search).