Computational Modeling of Cognition and Behavior

Visit the link below to download the full version of this book:

https://medipdf.com/product/computational-modeling-of-cognition-and-behavior/

Click Download Now

University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre, New Delhi – 110025, India
79 Anson Road, #06–04/06, Singapore 079906

Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781107109995
DOI: 10.1017/9781316272503
© Simon Farrell and Stephan Lewandowsky 2018
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2018
Printed in the United Kingdom by Clays, St Ives plc
A catalogue record for this publication is available from the British Library.
Library of Congress Cataloging-in-Publication Data
Names: Farrell, Simon, 1976– author. | Lewandowsky, Stephan, 1958– author.
Title: Computational modeling of cognition and behavior / Simon Farrell,
University of Western Australia, Perth, Stephan Lewandowsky, University of Bristol.
Description: New York, NY : Cambridge University Press, 2018.
Identifiers: LCCN 2017025806 | ISBN 9781107109995 (Hardback) | ISBN 9781107525610 (paperback)
Subjects: LCSH: Cognition–Mathematical models. | Psychology–Mathematical models.
Classification: LCC BF311 .F358 2018 | DDC 153.01/5118–dc23
LC record available at https://lccn.loc.gov/2017025806
ISBN 978-1-107-10999-5 Hardback
ISBN 978-1-107-52561-0 Paperback
Cambridge University Press has no responsibility for the persistence or accuracy
of URLs for external or third-party Internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
To Jodi, Alec, and Sylvie, with love (S.F.)
To Annie and the tribe (Ben, Rachel, Thomas, Jess, and Zachary) with love (S.L.)
 Contents

List of Illustrations page xiii

List of Tables xviii
List of Contributors xix
Preface xxi

Part I Introduction to Modeling 1

1 Introduction 3
1.1 Models and Theories in Science 3
1.2 Quantitative Modeling in Cognition 6
1.2.1 Models and Data 6
1.2.2 Data Description 9
1.2.3 Cognitive Process Models 13
1.3 Potential Problems: Scope and Falsifiability 17
1.4 Modeling as a “Cognitive Aid” for the Scientist 20
1.5 In Vivo 22

2 From Words to Models 24

2.1 Response Times in Speeded-Choice Tasks 24
2.2 Building a Simulation 26
2.2.1 Getting Started: R and RStudio 26
2.2.2 The Random-Walk Model 27
2.2.3 Intuition vs. Computation: Exploring the Predictions of a
Random Walk 31
2.2.4 Trial-to-Trial Variability in the Random-Walk Model 33
2.2.5 A Family of Possible Sequential-Sampling Models 37
2.3 The Basic Toolkit 38
2.3.1 Parameters 38
2.3.2 Connecting Model and Data 40
2.4 In Vivo 40

vii
viii Contents

Part II Parameter Estimation 45

3 Basic Parameter Estimation Techniques 47

3.1 Discrepancy Function 47
3.1.1 Root Mean Squared Deviation (RMSD) 48
3.1.2 Chi-Squared (χ 2 ) 49
3.2 Fitting Models to Data: Parameter Estimation Techniques 50
3.3 Least-Squares Estimation in a Familiar Context 50
3.3.1 Visualizing Modeling 51
3.3.2 Estimating Regression Parameters 53
3.4 Inside the Box: Parameter Estimation Techniques 57
3.4.1 Simplex 57
3.4.2 Simulated Annealing 61
3.4.3 Relative Merits of Parameter Estimation Techniques 64
3.5 Variability in Parameter Estimates 65
3.5.1 Bootstrapping 65
3.6 In Vivo 70

4 Maximum Likelihood Parameter Estimation 72

4.1 Basics of Probabilities 72
4.1.1 Defining Probability 72
4.1.2 Properties of Probabilities 73
4.1.3 Probability Functions 75
4.2 What Is a Likelihood? 80
4.3 Defining a Probability Distribution 85
4.3.1 Probability Functions Specified by the Psychological Model 86
4.3.2 Probability Functions via Data Models 86
4.3.3 Two Types of Probability Functions 91
4.3.4 Extending the Data Model 92
4.3.5 Extension to Multiple Data Points and Multiple Parameters 93
4.4 Finding the Maximum Likelihood 95
4.5 Properties of Maximum Likelihood Estimators 101
4.6 In Vivo 103

5 Combining Information from Multiple Participants 105

5.1 It Matters How You Combine Data from Multiple Units 105
5.2 Implications of Averaging 106
5.3 Fitting Aggregate Data 109
5.4 Fitting Individual Participants 111
5.5 Fitting Subgroups of Data and Individual Differences 113
5.5.1 Mixture Modeling 113
5.5.2 K-Means Clustering 118
5.5.3 Modeling Individual Differences 121
5.6 In Vivo 123
 Contents ix

6 Bayesian Parameter Estimation 126

6.1 What Is Bayesian Inference? 126
6.1.1 From Conditional Probabilities to Bayes Theorem 126
6.1.2 Marginalizing Probabilities 129
6.2 Analytic Methods for Obtaining Posteriors 130
6.2.1 The Likelihood Function 130
6.2.2 The Prior Distribution 131
6.2.3 The Evidence or Marginal Likelihood 134
6.2.4 The Posterior Distribution 135
6.2.5 Estimating the Bias of a Coin 136
6.2.6 Summary 139
6.3 Determining the Prior Distributions of Parameters 139
6.3.1 Non-Informative Priors 139
6.3.2 Reference Priors 142
6.4 In Vivo 143

7 Bayesian Parameter Estimation 146

7.1 Markov Chain Monte Carlo Methods 146
7.1.1 The Metropolis-Hastings Algorithm for MCMC 147
7.1.2 Estimating Multiple Parameters 153
7.2 Problems Associated with MCMC Sampling 160
7.2.1 Convergence of MCMC Chains 161
7.2.2 Autocorrelation in MCMC Chains 162
7.2.3 Outlook 162
7.3 Approximate Bayesian Computation: A Likelihood-Free Method 163
7.3.1 Likelihoods That Cannot be Computed 163
7.3.2 From Simulations to Estimates of the Posterior 164
7.3.3 An Example: ABC in Action 166
7.4 In Vivo 170

8 Bayesian Parameter Estimation 172

8.1 Gibbs Sampling 172
8.1.1 A Bivariate Example of Gibbs Sampling 173
8.1.2 Gibbs vs. Metropolis-Hastings Sampling 176
8.1.3 Gibbs Sampling of Multivariate Spaces 176
8.2 JAGS: An Introduction 177
8.2.1 Installing JAGS 177
8.2.2 Scripting for JAGS 177
8.3 JAGS: Revisiting Some Known Models and Pushing Their Boundaries 182
8.3.1 Bayesian Modeling of Signal-Detection Theory 182
8.3.2 A Bayesian Approach to Multinomial Tree Models:
The High-Threshold Model 186
8.3.3 A Bayesian Approach to Multinomial Tree Models 190
8.3.4 Summary 198
8.4 In Vivo 198
x Contents

9 Multilevel or Hierarchical Modeling 203

9.1 Conceptualizing Hierarchical Modeling 203
9.2 Bayesian Hierarchical Modeling 204
9.2.1 Graphical Models 204
9.2.2 Hierarchical Modeling of Signal-Detection Performance 207
9.2.3 Hierarchical Modeling of Forgetting 211
9.2.4 Hierarchical Modeling of Inter-Temporal Preferences 218
9.2.5 Summary 226
9.3 Hierarchical Maximum Likelihood Modeling 228
9.3.1 Hierarchical Maximum Likelihood Model of Signal Detection 228
9.4 Recommendations 233
9.5 In Vivo 234

Part III Model Comparison 239

10 Model Comparison 241

10.1 Psychological Data and the Very Bad Good Fit 241
10.1.1 Model Complexity and Over-Fitting 243
10.2 Model Comparison 248
10.3 The Likelihood Ratio Test 249
10.4 Akaike’s Information Criterion 256
10.5 Other Methods for Calculating Complexity and Comparing Models 261
10.5.1 Cross-Validation 262
10.5.2 Minimum Description Length 262
10.5.3 Normalized Maximum Likelihood 263
10.6 Parameter Identifiability and Model Testability 264
10.6.1 Identifiability 264
10.6.2 Testability 269
10.7 Conclusions 270
10.8 In Vivo 271

11 Bayesian Model Comparison Using Bayes Factors 273

11.1 Marginal Likelihoods and Bayes Factors 273
11.2 Methods for Obtaining the Marginal Likelihood 277
11.2.1 Numerical Integration 278
11.2.2 Simple Monte Carlo Integration and Importance Sampling 280
11.2.3 The Savage-Dickey Ratio 284
11.2.4 Transdimensional Markov Chain Monte Carlo 287
11.2.5 Laplace Approximation 294
11.2.6 Bayesian Information Criterion 297
11.3 Bayes Factors for Hierarchical Models 301
11.4 The Importance of Priors 303
11.5 Conclusions 306
11.6 In Vivo 306
 Contents xi

Part IV Models in Psychology 309

12 Using Models in Psychology 311

12.1 Broad Overview of the Steps in Modeling 311
12.2 Drawing Conclusions from Models 312
12.2.1 Model Exploration 312
12.2.2 Analyzing the Model 314
12.2.3 Learning from Parameter Estimates 315
12.2.4 Sufficiency of a Model 316
12.2.5 Model Necessity 318
12.2.6 Verisimilitude vs. Truth 323
12.3 Models as Tools for Communication and Shared Understanding 324
12.4 Good Practices to Enhance Understanding and Reproducibility 326
12.4.1 Use Plain Text Wherever Possible 326
12.4.2 Use Sensible Variable and Function Names 327
12.4.3 Use the Debugger 327
12.4.4 Commenting 328
12.4.5 Version Control 328
12.4.6 Sharing Code and Reproducibility 329
12.4.7 Notebooks and Other Tools 330
12.4.8 Enhancing Reproducibility and Runnability 331
12.5 Summary 332
12.6 In Vivo 332

13 Neural Network Models 334

13.1 Hebbian Models 334
13.1.1 The Hebbian Associator 334
13.1.2 Hebbian Models as Matrix Algebra 339
13.1.3 Describing Networks Using Matrix Algebra 348
13.1.4 The Auto-Associator 349
13.1.5 Limitations of Hebbian Models 356
13.2 Backpropagation 356
13.2.1 Learning and the Backpropagation of Error 360
13.2.2 Applications and Criticisms of Backpropagation in Psychology 364
13.3 Final Comments on Neural Networks 365
13.4 In Vivo 366

14 Models of Choice Response Time 369

14.1 Ratcliff’s Diffusion Model 369
14.1.1 Fitting the Diffusion Model 371
14.1.2 Interpreting the Diffusion Model 383
14.1.3 Falsifiability of the Diffusion Model 385
14.2 Ballistic Accumulator Models 386
14.2.1 Linear Ballistic Accumulator 386
14.2.2 Fitting the LBA 388
xii Contents

14.3 Summary 391

14.4 Current Issues and Outlook 392
14.5 In Vivo 393

15 Models in Neuroscience 395

15.1 Methods for Relating Neural and Behavioral Data 397
15.2 Reinforcement Learning Models 398
15.2.1 Theories of Reinforcement Learning 398
15.2.2 Neuroscience of Reinforcement Learning 404
15.3 Neural Correlates of Decision-Making 410
15.3.1 Rise-to-Threshold Models of Saccadic Decision-Making 410
15.3.2 Relating Model Parameters to the BOLD Response 411
15.3.3 Accounting for Response Time Variability 413
15.3.4 Using Spike Trains as Model Input 415
15.3.5 Jointly Fitting Behavioral and Neural Data 417
15.4 Conclusions 420
15.5 In Vivo 421

Appendix A Greek Symbols 424

Appendix B Mathematical Terminology 425

References 427
Index 455
 Illustrations

1.1 An example of data that defy easy description and explanation without a
quantitative model. 4
1.2 The geocentric model of the solar system developed by Ptolemy. 5
1.3 Observed recognition scores as a function of observed classification confidence
for the same stimuli (each number identifies a unique stimulus). 7
1.4 Observed and predicted classification (left panel) and recognition (right panel). 8
1.5 Sample power law learning function (solid line) and alternative exponential
function (dashed line) fitted to the same data. 11
1.6 The representational assumptions underlying GCM. 14
1.7 The effects of distance on activation in the GCM. 15
1.8 Stimuli used in a classification experiment by Nosofsky (1991). 16
1.9 Four possible hypothetical relationships between theory and data involving
two measures of behavior (A and B). 19
2.1 Graphical illustration of a simple random-walk model. 25
2.2 Predicted decision-time distributions from the simple random-walk model
when the stimulus is non-informative. 31
2.3 Predicted decision-time distributions from the simple random-walk model
with a positive drift rate (set to 0.03 for this example). 32
2.4 Predicted decision-time distributions from the modified random-walk model
with a positive drift rate (set to 0.035 for this example) and trial-to-trial
variability in the starting point (set to 0.8). 35
2.5 Predicted decision-time distributions from the modified random-walk model
with a positive drift rate (set to 0.03 for this example) and trial-to-trial
variability in the drift rate (set to 0.025). 36
2.6 Overview of the family of sequential-sampling models. 38
2.7 The basic idea: We seek to connect model predictions to the data from our
experiment(s). 40
3.1 Data (plotting symbols) from Experiment 1 of Carpenter et al. (2008)
(test/study condition) with the best-fitting predictions (solid line) of a power
function. 48
3.2 An “error surface” for a linear regression model given by y = X b + e. 51
3.3 Two snapshots during parameter estimation of a simple regression line. 56
3.4 Two-dimensional projection of the error surface in Figure 3.2. 58

xiii
xiv List of Illustrations

3.5 Probability with which a worse fit is accepted during simulated annealing as a
function of the increase in discrepancy ( f ) and the temperature parameter (T). 63
3.6 The process of obtaining parameter estimates for bootstrap samples. 66
3.7 Histograms of parameter estimates obtained by the bootstrap procedure, where
data are generated from the model and the model is fit to the generated
bootstrap samples. 68
4.1 An example probability mass function: the probability of responding A to
exactly NA out of N=10 items in a categorization task, where the probability of
an A response to any particular item is PA = 0.7. 76
4.2 An example cumulative distribution function (CDF). 77
4.3 An example probability density function (PDF). 78
4.4 Reading off the probability of discrete data (top panel) or the probability
density for continuous data (bottom panel). 81
4.5 Distinguishing between probabilities and likelihoods. 83
4.6 The probability of a data point under the binomial model, as a function of the
model parameter PA and the data point NA , the number of A responses in a
categorization task. 84
4.7 Different ways of generating a predicted probability function, depending on
the nature of the model and the dependent variable. 91
4.8 The joint likelihood function for the Wald parameters m and a given the data set
t = [0.6 0.7 0.9]. 94
4.9 A likelihood function (left panel), and the corresponding log-likelihood
function (middle) and deviance function (−2 log likelihood; right panel). 97
4.10 A scatterplot between the individual data points (observed proportion A
responses for the 34 faces) and the predicted probabilities from GCM under
the maximum likelihood parameter estimates. 101
5.1 Simulated consequences of averaging of learning curves. 107
5.2 A simulated saccadic response time distribution from the gap task. 114
5.3 Left panel: Accuracy serial position function for immediate free recall of a
list of 12 words presented as four groups of three items. Right panel: Serial
position functions for three clusters of individuals identified using K-means
analysis. 118
5.4 The gap statistic for different values of k. 120
5.5 A structural equation model for choice RT. 122
6.1 Two illustrative Beta distributions obtained by the R code in Listing 6.1. 133
6.2 Bayesian prior and posterior distributions obtained by a slight modification of
the R code in Listing 6.1. 137
6.3 Jeffreys prior, Beta(0.5,0.5), for a Bernoulli process. 140
7.1 MCMC output obtained by Listing 7.1 for different parameter values. 150
7.2 MCMC output obtained by Listing 7.2 for different parameter values. 153
7.3 Experimental procedure for a visual working memory task in which
participants have to remember the color of a varying number of squares. 154
7.4 Data (circles) from a single subject in the color estimation experiment of
Zhang and Luck (2008) and fits of the mixture model (solid lines). 155
 List of Illustrations xv

7.5 Posterior distributions of parameter estimates for g and σvM obtained when
fitting the mixture model to the data in Figure 7.4. 160
7.6 Overview of a simple Approximate Bayesian Computation (ABC) rejection
algorithm. 165
7.7 a. Data from an hypothetical recognition memory experiment in which people
respond “old” or “new” to test items that are old or new. b. Signal-detection
model of the data in panel a. 167
8.1 Illustration of a Gibbs sampler for a bivariate normal distribution. 174
8.2 Overview of how JAGS is being used from within R. 178
8.3 Output obtained from R using the plot command with an MCMC object
returned by the function coda.samples. 181
8.4 a. Data from an hypothetical recognition memory experiment in which people
respond “old” or “new” to test items that are old or new. b. Signal-detection
model of the data in panel a. 182
8.5 Output from JAGS for the signal detection model illustrated in Figure 8.4. 185
8.6 Convergence diagnostics for the JAGS signal detection model reported in
Figure 8.5. 186
8.7 The high-threshold (1HT) model of recognition memory expressed as a
multinomial processing tree model. 187
8.8 Output from JAGS for the high-threshold (1HT) model illustrated in Figure 8.7. 190
8.9 a. Autocorrelation pattern for the output shown in Figure 8.8. b. The same
autocorrelations after thinning. Only every fourth sample is considered during
each MCMC chain. 191
8.10 The no-conflict MPT model proposed by Wagenaar and Boer (1987) to account
for performance in the inconsistent-information condition in their experiment. 193
8.11 Output from a run of the no-conflict model for the data of Wagenaar and Boer
(1987) using Listings 8.8 and 8.9. 197
8.12 Example of a 95% highest density interval (HDI). 199
8.13 Diagram of the normal model, in the style of the book, Doing Bayesian Data
Analysis (Kruschke, 2015). 201
8.14 Diagram of the normal model, in the style of conventional graphical models. 202
9.1 Graphical model for the signal-detection example from Section 8.3.1. 205
9.2 Graphical model for a signal-detection model that is applied to a number of
different conditions or participants. 206
9.3 Graphical model for a signal-detection model that is applied to a number of
different conditions or participants. 207
9.4 Hierarchical estimates of individual hit rates (left panel) and false alarm
rates (right) shown as a function of the corresponding individual frequentist
estimates for the data in Table 9.2. 210
9.5 Graphical model for a hierarchical model of memory retention. 213
9.6 Results of a run of the hierarchical exponential forgetting model defined in
Listings 9.3 and 9.4. 216
9.7 Posterior densities of the parameters a, b, and α of the hierarchical exponential
forgetting model defined in Listings 9.3 and 9.4. 216
xvi List of Illustrations

9.8 Results of a run of the hierarchical power forgetting model defined in Listings
9.5 and 9.6. 218
9.9 Graphical model for a hierarchical model of intertemporal choice. 221
9.10 Data from 15 participants of an intertemporal choice experiment reported by
Vincent (2016). 223
9.11 Snippet of the data file from the experiment by Vincent (2016) that is used by
the R script in Listing 9.8. 224
9.12 Predictions of the hierarchical intertemporal choice model for the experimental
conditions explored by Vincent (2016). 226
9.13 Posterior densities for the parameters of the hierarchical intertemporal choice
model when it is applied to the experimental conditions explored by Vincent
(2016). 227
10.1 Fits of the polynomial law of sensation to noisy data generated from a
logarithmic function. 242
10.2 Predictions from a polynomial function of order 2 (left panel) and order 10
(right panel), with randomly sampled parameter values. 243
10.3 An illustration of the bias-variance trade-off. 246
10.4 The bias-variance trade-off. As model complexity (the order of the fitted
polynomial) increases, bias decreases and variance increases. 247
10.5 Out-of-set prediction error. 248
10.6 The two functions underlying prospect theory. 251
10.7 K-L distance is a function of models and their parameters. 257
10.8 Prior probability (solid horizontal line) and posterior probabilities (lines
labeled β and ) for two parameters in a multinomial tree model that are
“posterior-probabilistically-identified.” 267
11.1 Illustration of how the marginal likelihood can implement the principle of
parsimony. 275
11.2 Illustration of the Savage-Dickey density ratio for the signal detection model,
examining whether b = 0. 286
11.3 Autocorrelations in samples of the model indicator pM2 using noninformative
pseudo-priors (left panel) and pseudo-priors approximating the posterior (right
panel). 294
11.4 Predicted hit and false alarm rates in a change-detection task derived using
non-informative (left-hand quadrants) and informative (right-hand quadrants)
prior distributions for two models of visual working memory. 305
12.1 A flowchart of modeling. 312
12.2 The effect of the response suppression parameter η in Lewandowsky’s (1999)
connectionist model of serial recall. 314
12.3 A schematic depiction of sufficiency and necessity. 319
13.1 Architecture of a Hebbian model of associative memory. 335
13.2 Different ways of representing information in a connectionist model. 336
13.3 Schematic depiction of the calculation of an outer product W between two
vectors o and c. 342
13.4 Generalization in the Hebbian model. 346