| John R. Anderson: |
| ACT-R 5.0 |
| ACT-R is a cognitive theory and simulation system for developing cognitive models. It assumes cognition emerges through the interaction of a procedural memory of productions with a declarative memory of chunks. The ACT-R 4.0 version of the theory was detailed in the book "The Atomic Components of Thought" by John R. Anderson and Christian Lebiere, published in 1998 by Lawrence Erlbaum Associates. Since its release in 1997, ACT-R 4.0 has supported the development of over 50 cognitive models, published in the literature by as many different researchers. These models cover topics as diverse as driving behavior, implicit memory, learning backgammon, metaphor processing, and emotion. We have recently developed a new version, ACT-R 5.0 that extends the ACT-R 4.0 to be more interruptible, to achieve greater across-task parameter consistency, to have better mechanisms of production learning, and to be more in correspondence with our knowledge of brain function. While the new system extends the capabilities of ACT-R 4.0 it involves relatively few changes and is actually simpler. This short tutorial will provide an overview of ACT-R, as it is specified in the 5.0 version. It will not assume a prior background in ACT-R 4.0. |
| Randy O'Reilly: |
| No abstract available yet |
| Roger Ratcliff: |
| Diffusion Models |
| Roger Ratcliff is planning to discuss the diffusion model for performance in two-choice RT tasks. Recent developments in the diffusion model, such as incorporating variability and estimation of contaminants, allow the model to fit response time distributions for both correct and error responses. The mechanisms of the diffusion model and the fitting process will be dealt with in some detail. Comparisons between other stochastic models will be examined briefly Application to data from four experimental paradigms to examine the effect of aging on processing will be presented. |
| Erik Altmann: |
| Functional decay in working memory: Subtle effects of a pervasive process |
| Decay of distracting information is essential to a functioning memory system - without it, interference would quickly become catastrophic. Evidence from three domains will be reviewed to support the implication that decay is as pervasive as memory itself. In task switching, decay allows the system to distinguish the current task from previous tasks. It also predicts the novel effect of within-run slowing, a short-term reversal of practice effects. In language comprehension, decay allows the system to distinguish current syntactic structures, and explains the time course of Stroop effects. In the probe digit paradigm, decay predicts that presentation rate should interact with serial position - as it does in Exp. 1 of Waugh and Norman (1965), an empirical pillar of interference theory. Model fits for each domain will be presented. In each case, the model incorporates a decay process in which activation rises and falls with frequency of instance retrieval. |
| Conor Dolan: |
| A comprehensive approach to ML estimation of parameters of (mixtures of) common reaction time distributions given optional truncation or censoring |
| We present a comprehensive approach to distributional reaction time (RT)
analysis, based on maximum likelihood estimation. Given certain information
concerning distribution functions, one can estimate the parameters of this
distribution and of finite mixtures including this distribution. This
information relates to the integrals and derivatives of the distribution
function. In addition left and/or right censoring or truncation may be imposed. Censoring and truncation are useful methods of accommodating outlying observations, which are a pervasive problem in RT research. We employ quasi-Newton optimization to obtain maximum likelihood estimates. Multi-case analyses can be carried out enabling one to conduct detailed comparisons of observed RT distributions. Parameters may be freely estimated, estimated subject to boundary constraints, constrained to be equal, or fixed. |
| Arthur Jacobs: |
| Modeling ROC curves with a localist connectionist model (MROM) |
| This study provides further experimental and computational evidence for the hypothesis that performance in one of the most widely used cognitive tests, the lexical decision task, is based on the output of a continuous strength, fast-guess mechanism. According to the Multiple Read-Out Model of word recognition (MROM; Grainger & Jacobs, 1996), and the recent, revised dual route cascaded (DRC) model of reading (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001) this mechanism monitors overall activity in the mental lexicon and computes a familiarity (or lexicality) index during the early phases of stimulus processing that influences both positive and negative responses in the lexical decision task. We extend the MROM to allow predictions of receiver operating characteristics (ROCs) in a data-limited variant of the lexical decision task and show that the model provides accurate descriptions of the participants' behavior. This implies that a simple signal detection model can parsimoniously account for performance in data-limited lexical decision tasks. Our results challenge all models of word recognition that do not include a familarity assessment mechanism and provide indirect support for models (other than the MROM) that do. In addition, the results suggest that the deadline mechanism for generating "no" responses in the lexical decision task, as implemented by both the MROM and the revised DRC, may not always be necessary. |
| Han van der Maas: |
| Phase Transitions in the Trade-Off Between Speed and Accuracy |
| Most variants of the random walk model (RWM) of choice reaction time predict
a continuous trade-off between speed and accuracy. That is, the change from
guessing to accurate responding takes place smoothly, and each intermediate
level of accuracy and speed is stable. In this talk a phase-transition model
(PTM) of the trade-off is proposed, which is a dynamic extension of the
older fast guess models. In our elementary phase transition model (i.e., a
so-called cusp model) sudden switches in reaction time and accuracy occur as
function of smooth and small changes in payoffs for speed and accuracy.
According to this model, intermediate states are unstable. To compare the
PTM model to the RWM, two experiments were conducted, one on bimodality and
one hysteresis. The bimodality experiment forces subject to behave in-between guessing and accurate responding, which requires, according to the RWM, adjustment of the bounds of the decision process, and, according to the PTM, sampling from two different modes. In the first case RTs should be unimodal, in the second case they should be bimodal. Our data indicate bimodality. In the hysteresis experiment, the payoffs for speed and accuracy are slowly varied between values favoring guessing and values favoring accurate responding. Jumps between the states should differ in position depending on the direction of change in the payoffs. Initial evidence for this effect, called hysteresis, will be presented. |
| Michael E. J. Masson: |
| When Fluency Begets Bias: A Multinomial Model of Priming in Masked Word Identification |
| In the course of attempting to identify a briefly presented target word, unexpected processing fluency can create a sense of familiarity or a perceptual experience that gives rise to response bias. Experiments will be described in which long- and short-term priming were used to generate fluent processing of one member of a pair of probes in a two-alternative forced-choice version of the masked word identification task. These experiments reveal parallel patterns of bias effects across long- and short-term priming paradigms, whereby subjects tend to select the primed member of a pair of probes. A simple multinomial model that assumes a common set of processing principles for both paradigms will be described and fits of that model to results generated by these paradigms will be presented. The model emphasizes the influence of priming on the fluency with which probes or internally generated candidates are processed. |
| M. Meeter, J.M.J. Murre & L. Talamini: |
| Connectionist models of the hippocampus: structural adequacy and new sources of data at three different levels of organization |
| Not many neural network models can be used for quantitative modeling of
behavioral data. Neural networks can make up for this limitation by, in
the terminology of Barbara Webb (BBS, in press) a higher structural
adequacy. Moreover, they may attempt to explain data of a different
kind, such as lesion data or electrophysiological data. These two forms
of additional support will be explored in the context of three different
models of hippocampal functioning, that target at three different levels
of organization. It will also be discussed how, by forming a hierarchy
from basic to very abstract, these models can inform and constrain
each-other. The first of the three discussed models is TraceLink, an abstract model of consolidation and amnesia. Its main assumption is that there is a consolidation process that helps transfer memories from a temporary, medial temporal lobe repository to permanent neocortical storage sites. TraceLink can be used to explain a wide variety of neuropsychological memory disorders. The second model explores the role of the parahippocampal region as an interface between the hippocampus and neocortical processing areas. With this model we have simulated list learning, and attempt to explain the profile of memory deficits seen in schizophrenia. The third model attempts to capture the details of hippocampal functioning. It has as goal predicting firing patterns across the hippocampus when novel stimuli are presented. |
| Dennis Norris: |
| Applying random-walk decision process to models of spoken word recognition |
| This paper will describe preliminary work on integrating a random-walk decision process with the Merge (Shortlist) model of spoken word recognition. Merge, like most connectionist models, produces activation functions which can only be related to RTs and error rates qualitatively. When the output of the model is used to drive a a random-walk decision process it can give a precise quantitative account of both speed and error rate in lexical decision and phonemic classification experiments. Some emergent properties of the new model generate a novel explanation of existing data. |
| Jaap Murre, Antonio Chessa, Martijn Meeter: |
| A mathematical model of learning and forgetting applied to disorders of long-term memory |
| We assume that memory processes can be decomposed into a number of stores that contain memory representations and that a newly learned memory passes through one or more of the stores. We will show how for long-term memory, the stores can be identified as hippocampus and cortex. Model fits to animal data of retrograde and anterograde amnesia will be shown, as well as to recent data with mice that lack cortical learning. The model will also be fitted to human data. A problem is that in human tests of retrograde amnesia, the difficulty of questions is often varied arbitrarily. We will demonstrate that a suitable transformation of the test results can still make them amenable to quantitative fitting. |
| Randy O'Reilly: |
| Models of Hippocampal and Neocortical Contributions to Memory |
| Hippocampal amnesia reveals that the hippocampus and the spared neocortex make different contributions to memory. Ken Norman and I have been applying neural network models of the neocortex and hippocampus to a variety of recognition memory phenomena to more precisely and mechanistically characterize these different contributions. Specifically, the models show how the neocortical and hippocampal contributions differ in their sensitivity to manipulations of item similarity and interference, making predictions that have been confirmed in neuropsychological and behavioral tests. |
| Jeroen Raaijmakers: |
| Modeling Implicit Memory |
| Research on implicit memory or 'priming' phenomena has been an important topic in the past 10-15 years. Despite this great popularity, such phenomena have, until recently, not been treated by computational models. However, in the past few years a number of promising approaches have been developed. I will discuss a number of these global approaches giving special attention to the recently developed REM model that holds out the promise of providing an integrated theory for both implicit and explicit memory. Despite the successes of these approaches, there do remain some problems, most notably the question regarding the exact locus of priming effects. |
| Roger Ratcliff: |
| No abstract available yet |
| Hedderik van Rijn: |
| An ACT-R model for Lexical Decision |
| We present an ACT-R model of lexical decision, based on the new ACT-R 5.0 mechanism of asynchronical retrieval of memory chunks. The model uses the complete CELEX four-letter word-form lexicon initialized using the corresponding empirical frequencies as its mental lexicon. The lexical decision task is performed by activating those entries in the lexicon that are similar to the probe and selecting the entry that is most active. This model is used to fit data from the signal-to-respond paradigm (see Wagenmakers, this workshop) by assuming that if a reaction has to be given before a correct entry is retrieved, the model simply guesses. We will demonstrate that this approach provides a similar fit as the REM-LD model's approach which simultaneously considers the diagnosticity of the evidence for the 'WORD' response and the 'NONWORD' response. |
| Richard Shiffrin: |
| Bayesian-based Memory Modeling |
| In recent years, my colleagues and I have developed memory models rooted
in the assumption that memory retrieval and decision making has adapted to
approach optimal performance, conditional on a set of basic capacity
limitations (usually characterized in terms of sources of noise). The
approach has been termed REM, for Retrieving Effectively from Memory.
Bayesian analysis is used to develop models for given tasks. The approach
has been used for knowledge retrieval, episodic retrieval, and the ways
these interact (i.e. implicit memory--see the morning talk by Jeroen
Raaijmakers). In this talk, I review some applications to episodic retrieval, and describe new extensions we are starting to explore to link memory tasks to categorization and other tasks. |
| Niels Taatgen: |
| Explicit Memory: Psychological Reality or Psychologist's Reality? |
| The distinction between implicit and explicit memory and implicit and explicit learning has proven to be useful in psychology. Although usually treated as seperate processes and systems, the interaction between the two has recently become a focus of attention. An interesting question to ask in this respect is whether a theory is possible that can explain the phenomena, but doesn't have to rely one seperate memory systems or seperate processes. I will outline such an alternative based on the ACT-R architecture, in which all learning and all memory can be considered as implicit, and in which explicit learning and memory can be explained by implicit processing. |
| Ingmar Visser: |
| The simple recurrent network and hidden Markov model for implicit sequence learning |
| We have studied implicit sequence learning in two experiments and compared reaction times and prediction and generation ability of subjects. The simple recurrent network model for sequence learning predicts an exact inverse relationship between prediction ability and reaction times. In our first experiment we established this relationship between online predictions and reaction times. In the second experiment we included online generation trials, i.e. subjects had to freely generate series of trials in between sequences of RT trials. We use hidden Markov models to characterize these generated sequences. The fitted models show that during learning subjects build more complex models of the presented sequences. |
| Eric-Jan Wagenmakers: |
| A Bayesian Model for Lexical Decision |
| We present a new model for lexical decision, REM-LD, that is based on REM theory (e.g., Shiffrin & Steyvers, 1997). REM-LD uses a principled (i.e., Bayesian) decision process that simultaneously considers the diagnosticity of the evidence for the 'WORD' response and the 'NONWORD' response. The model calculates the odds that the presented stimulus is a word by accumulating likelihood ratios for each lexical entry in a small neighborhood of similar words. Using a signal-to-respond paradigm, we will demonstrate how REM-LD can handle the effects of word frequency, nonword lexicality, and repetition priming. We will also discuss some ideas on how REM-LD can be extended to incorporate neighborhood effects and predict response times as well as percentage correct. |
| Trish Van Zandt: |
| Response Reversals in Recognition Memory |
| Confidence judgments are commonly elicited in recognition memory and psychophysical tasks, under the assumption that the reported level of confidence is a measure of the perceived level of familiarity or signal strength. Recent work (Van Zandt, 2000) has shown effects on the confidence-based ROC curve that are inconsistent with this assumption and are more consistent with a dynamic ``balance of evidence'' model. In this model, information toward alternative responses is accumulated on parallel channels and confidence is based on the difference between the channels at the time a decision is made. This model makes predictions about the frequency with which rememberers should change their minds when making recognition decisions under bias. These predictions are tested in several experiments, in which speeded old/new recognition decisions are collected, and then rememberers are asked to make confidence judgments about those decisions on a bipolar scale. The rate of response reversals, where the confidence rating does not support the primary old/new decision, can be explained as an outcome of the dynamics of the accumulation process. |
| René Zeelenberg: |
| Repetition Priming in Implicit Memory Tasks: Effects of Bias and Enhanced Discriminability |
| Ratcliff and McKoon (1996, 1997) showed that prior study results in bias. Using a two-alternative forced-choice procedure, they showed that prior study of the target resulted in a benefit. However, prior study of the foil resulted in a cost. An important question is whether prior study results in only bias or whether there is also evidence for an enhanced discriminability effect. In three experiments (picture identification, auditory word identification and visual word identification) we found that performance was better when both alternatives were studied compared to when neither of the alternatives was studied. Several models of bias and enhanced discriminability will be discussed. |