NEUROSCIENCE OF ERROR DETECTION
By
Alena Grigorievna Patsenka

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
DOCTOR OF PHILOSOPHY
Psychology
2010

ABSTRACT
NEUROSCIENCE OF ERROR DETECTION
By
Alena Grigorievna Patsenka
Several accounts of anterior cingulated cortex (ACC) function currently exist – a
conflict monitoring theory (e.g., Carter et al., 1998; Botvinick et al., 2001), a
reinforcement learning theory (e.g., Holroyd & Coles, 2002), and an error-likelihood
theory (Brown & Braver, 2005). The theories agree that the ACC is involved in error
detection; however, the precise role of the ACC is still a subject of debate. In the conflict
monitoring view, the ACC detects a conflict between intended and actual responses. In
reinforcement learning view, the ACC is activated when events are worse-than-expected.
In error-likelihood theory, the ACC stores information about error-likelihood. The results
of the present research challenge the idea that the ACC performs conflict monitoring, and
they are consistent with the idea that the ACC responds to worse-than-expected
outcomes. The fMRI finding that the same region in the left dACC responds both to
errors that people are aware of and to unexpected negative feedback is consistent with the
reinforcement learning account and it is problematic for the conflict monitoring theory.
The results of the present research also lead to reinterpretation of the findings taken to
support the error-likelihood theory.
In the present study, I develop a paradigm in which a memory task and a multiple
object tracking tasks are combined. Participants receive feedback on the accuracy of their
performance only after both tasks are completed. Negative feedback is ambiguous in
terms of what type of error occurred (memory or tracking). After receiving negative

feedback, participants are asked to interpret it. The behavioral results showed that most of
the time the interpretations are correct (e.g. when participants make a memory error they
indicate that they made a memory error); however, sometimes the interpretations are
incorrect (e.g. when participants make a memory error they indicate that they made a
tracking error). I assume here that participants interpret feedback correctly if they detect
an error and they interpret feedback incorrectly if they miss an error.
Contrasting all errors and correct trials, I identified a region within the left dACC
responding greater to errors. A region of interest analysis revealed an interaction between
error detection and phase of the trial: The region was more active to detected errors than
undetected errors during task performance, but during feedback the pattern was reversed,
such that the region was more active to undetected errors than to detected errors. The
pattern is problematic for the conflict monitoring theory, because there is no conflict
during feedback processing. The results are consistent with the reinforcement learning
theory, since both error detection and processing of unexpected negative feedback are
worse-than-expected outcomes.
The imaging results also showed that an error related signal within the left dACC
is content specific. Comparing the trials on which participants failed to retrieve correct
information with the trials on which participants lost track of dots, I found that tracking
errors lead to greater ACC activation. Several explanations are possible: One of them is
that the ACC contains error-specific neurons similar to feature detectors in visual cortex.
Alternatively, participants might have stronger expectations regarding the accuracy of
their performance on tracking task than memory task; therefore, violations of stronger
expectations result in greater ACC activation.

To my father, Grigori Stepanovich Patsenko, for teaching me to set high standards and
achieve goals regardless of how difficult and impossible they might seem.

 

iv

ACKNOWLEDGMENTS
Most importantly, I would like to thank my adviser Erik Altmann. I realize now
that no matter what I write here, it will not express the depth of my gratitude to him. He
not only made my dream possible, but insured that the road to it was enjoyable! Erik was
an amazing advisor: He set a great example of a true scientist, by being intelligent,
thinking critically about ideas, and ready to give up ideas in the face of conflicting
evidence. Being always open to new ideas and giving me a lot of freedom in my research,
he, at the same time, provided me with strong guidance. Importantly, Erik was very
generous in giving praise and encouraging me to keep working. It is hard to overestimate
the motivation he gave me when I was buried in doubts in my abilities. I was very happy
to have Erik as an advisor, he forever set the highest expectations for people of science
for me.
Also I would like to thank members of my Master’s and Doctorial committees:
Tom Carr, Fred Dyer, Zack Hambrick, Taosheng Lue, and Rose Zacks, for their critique
of my work, their comments improved my work greatly and made me prepared for every
possible question about it. I am also grateful to them for the creative work environment
that they support in the psychology department and cognitive science program at
Michigan State University. Special thanks also goes to Bruce Burns for initially inviting
me to Michigan State and giving me the opportunity to start a scientific research career.
In addition to faculty members, I am also thankful to graduate students who made
these years more enjoyable: John Dewey, Paul Engelhardt, Matt Husband, Graham
Pierce, and Emily Swensen (Darowski). Special thanks to Paul Engelhardt, who was not

 

v

only an extremely helpful colleague, advising me in all possible aspects of graduate life,
but also he became a close friend throughout these years. I am also thankful to my nonacademic friends, Nadia Karatkevich, Marina Kaspirovich, Lena Luchaninova, Oksana
Naumenko, Kathy and Sebastian Seifarth, Anna Senkina, Paul Tucci for diversifying my
life with parties, wine-tasting, dancing, and conversing with me on “human” subjects.
Finally, I want to thank my family. I am grateful to my parents, Galina
Nikolaevna and Grigori Stepanovich Patsenko, for teaching me how to think. They taught
me to set high standards and achieve impossible goals. They believed that one could
become whoever they want if the desire meets the effort. I am also grateful to my sister,
Natasha Patsenko (Sidorova) for setting an example in every aspect of a nonacademic
life. I give my special thanks to my aunt Angelina Stepanovna and uncle Isaac Aronovich
Garbar for making this “trip” to the U.S. possible. But no one deserves more thanks than
my boyfriend, Yuri Shalagin! I am so grateful to you for believing in me even when there
was no basis for it, or when you could not possibly know if there was. Thank you for
taking me on trips in the most inconvenient time, and thank you for going to parties alone
when I absolutely could not make it. You are always in my heart!

 

vi

TABLE OF CONTENTS
LIST OF TABLES..………………………………………………………………..…….ix
LIST OF FIGURES………………………………………………………………..……...x
1 INTRODUCTION……………………………………………………………………..1
1.1 The ACC Involvement in Error Detection…………………………………....2
1.1.1 Mismatch (or comparator) theory of ACC function………………2
1.1.2 Conflict monitoring theory of ACC function……………………..4
1.1.3 Reinforcement learning theory of ACC function………………....8
1.2 Overview of Current Research……………………………………………….12
2 ERROR DETECTION IN COMPLEX TASKS……………………………………15
2.1 Experiment 1: Behavioral study of error detection…………………………..15
2.1.1 Method……………………………………………………………..16
2.1.2 Results and Discussion…………………………………………….19
2.2 Experiment 2: Imaging study of error detection……………………………..21
2.2.1 Method……………………………………………………………..22
2.2.2 Behavioral Results…………………………………………………29
2.2.3 fMRI Results……………………………………………………….31
2.2.3.1 Whole brain analysis: all errors versus correct trials during
Task phase…………………………………………….….31
2.2.3.2 ROI analysis dACC during Task phase..…………...……33
2.2.3.3 ROI analysis dACC: Task versus Feedback……….…….38
2.2.3.4 Whole-brain analysis: Interaction between Consistency and
Time point…………………………………………….….41
3 GENERAL DISCUSSION………...............................................................................44
3.1 The dACC in response to worse-than-expected outcomes…………………..44
3.1.1 Evidence concerning the Conflict monitoring theory……………...46
3.1.2 Evidence concerning the Reinforcement learning theory………….49
3.1.3 Evidence concerning the Error-likelihood learning………………..50
3.1.4 Integrating the finding from ERP and neuroimaging studies……...53
3.2 Content-specific signal in the ACC………………………………………….56
3.3 ACC activation in tasks without response conflict…………………………..59
4 FUTURE DIRECTIONS………………………….…………………………………60
4.1 Does the ACC respond to better-than-expected outcomes?............................60
4.2 Violations of expectations in the context of anxiety………………………...61
4.3 Negative emotions and rate of learning……………………………………..62
4.4 Lingering effect of emotions………………………………………………...64
4.5 Lingering effect of emotions in a context of anxiety………………………..65
5 CONCLUSIONS……………………………………………………………………...67

 

vii

APPENDIX A: Analysis of regions other than the left dACC……..…………...……….68
APPENDIX B: Correct memory phase versus correct tracking phase………………......79
REFERENCES…………………………………………………………………………..80

 

viii

LIST OF TABLES
Table 1. Group analysis results for the contrast: “all errors > all correct”………………32
Table 2. Group analysis results for the contrast: “Consistently interpreted errors >
correct”…………………………………………………………………………………..37
Table 3. Group analysis results for the interaction between Consistency and Task
Phase……………………………………………………………………………………..42
Table A1. ROI analysis: Percentage BOLD signal change observed during feedback
processing for consistently and inconsistently interpreted tracking and memory errors in a
small region in the dACC (a) and in the left SFG (b). Repeated measures ANOVAs
showed neither the main effects (Consistency or Error Type) nor the interaction between
factors……………………………………………………………………………………71
Table A2. Group analysis results for the contrast: “Consistently interpreted tracking
errors > correct”.………………………………………………………………………...74
Table A3. Group analysis results for the contrast: “Inconsistently interpreted errors >
Consistently interpreted errors” during the Feedback phase…………………………….76

 

ix

LIST OF FIGURES
Figure 1. An example of a trial with a negative feedback in a modified multiple object
tracking task. ………………………………………………………….……....................19
Figure 2. Frequency of different types of interpretations as a function of error type in
Experiment 1 ……………………………………………………………..…….……......21
Figure 3. An example of a trial with a negative feedback in a modified task used for
imaging study…………………………………………………………………………….26
Figure 4. Frequency of different types of interpretations as a function of error type in
Experiment 2……………………………………………………………………………..30
Figure 5. Error-related activation. Regions identified in a whole brain analysis contrasting
all error trials and all correct trials (Terr_t + Terr_m + Merr_t +Merr_m) – 2 (Tcor +
Mcor). The arrow points to a region in left dorsal ACC, which is more active for error
trials………………………………………………………………………………………31
Figure 6. ROI analysis: the left dACC. The region shows Error Type effect: tracking
errors activate the region greater than memory errors, consistent with the idea that the
ACC represents information regarding types of errors people make. There is no effect of
Interpretation Type. The interaction of Error Type and Interpretation Type approaches
significance, suggesting an effect of Consistency of interpretation: the ACC responds
greater to consistently interpreted errors (Terr_t and Merr_m) compared to inconsistently
interpreted errors (Terr_m and Merr_t). The error bars show SEM…………………….34
Figure 7. Estimated Hemodynamic Response Functions (HDR) for different types of
error during the Task phase. The graph shows the effect of Error type – blue lines above
red lines, and the effect of consistency – solid lines above dashed lines………………..35
Figure 8. ROI analysis: the left dACC. The region shows Consistency effect: Errors
interpreted consistently lead to a greater dACC activation than errors interpreted
inconsistently. There is also an effect of Error Type. The interaction between Consistency
of Interpretation and Error Type is not significant. The error bars show SEM………….36
Figure 9. Consistently interpreted errors. Regions identified in a whole brain analysis
contrasting consistently interpreted errors and correct trials (Terr_t + Merr_m) – (Tcor +
Mcor). The arrows point to the regions more active for consistently interpreted errors.
They are two regions in the left dACC (big and small), the left MFG, and the left
SFG………………………………………………………………………………………37
Figure 10. ROI analysis: the left dACC. The region shows Consistency by Time Point
interaction: During task performance, errors interpreted consistently lead to a greater
dACC activation than errors interpreted inconsistently. However, during feedback
processing, errors interpreted inconsistently lead to greater dACC activation than errors
 

x

interpreted consistently. Neither of the main effects is significant. The error bars show
SEM……………………………………………………………………………………...39
Figure 11. ROI analysis: the left dACC. The region shows Consistency effect during
feedback processing only for tracking errors: Errors interpreted inconsistently lead to
greater dACC activation than errors interpreted inconsistently. The effect of consistency
is not significant for memory errors. The error bars show SEM………………………...40
Figure 12. Estimated HDR functions for all types of error during the Feedback phase...41
Figure 13. Consistency by Time point interaction. Regions identified in a whole-brain
analysis contrasting (Terr_t + FTer_m) and (Terr_m + FTer_t ). The arrows point to the
regions more active for (Terr_t + FTer_m). They are the left MFG and the left
Caudate…………………………………………………………………………………..42
Figure 14. Summary of the results from three studies: Holroyd & Coles, 2002 (A),
present study (B), and Brown & Braver, 2005 (C)………………………………………54
Figure A1. ROI analysis: a big region of the left dACC. During negative feedback
processing, the region shows Consistency effect for tracking errors, but not for memory
errors. The error bars show SEM………………………………………………………..70
Figure A2. Consistently interpreted tracking errors. Regions identified in a whole brain
analysis contrasting consistently interpreted tracking errors and correct trials (Terr_t –
Tcor). The arrows point to the regions more active for consistently interpreted tracking
errors. They are the dACC bilaterally, the left medial frontal gyrus (MFG), the left
superior frontal gyrus (SFG), and the inferior parietal gyri (IPG) bilaterally…………...72
Figure A3. ROI analysis: consistently interpreted tracking errors versus inconsistently
interpreted tracking errors during feedback processing in brain regions identified with the
contrast (Terr_t – Tcor). Neither of the comparisons reaches significant level. Thus the
bilateral IP, the left MFG, and the bilateral ACC do not show a consistency effect during
feedback processing. (The region in the left ACC described here is more anterior and
inferior compared to the regions in the left ACC identified in “all errors vs. correct trials”
contrast and in “consistent interpretations vs. correct trials” contrast.)………………….74
Figure A4. Feedback: Inconsistently interpreted errors. Contrasting inconsistently
interpreted errors and consistently interpreted errors during Feedback (FTer_m + FMer_t)
– (FTer_t + FMer_m), I find a single area in the left IPG responding greater to
inconsistently interpreted errors………………………………………………………….76
Figure A5. ROI analysis: the left dACC. The region shows an effect of the Trial Phase:
correct tracking phase is associated with a greater dACC activation compared to correct
memory retrieval phase………………………………………………………………......79

 

xi

“There are only two mistakes one can make along the road to truth:
not starting and not going all the way.”
Buddha

 

xii

1 INTRODUCTION
Human performance is far from being infallible. Errors pervade all our activities:
while typing one might press a wrong key; while preparing a cup of tea one might pour
water in a bowl instead of a cup; while solving a problem one might follow a wrong path.
Detection of an error, or deviation from accuracy, allows people to adjust their behavior
correspondingly. These adjustments could have a general nature, for example, people
could slow down after a typing error. However, errors often require task specific
adjustments such as narrowing attentional focus on a central letter in the Eriksen task or
processing ink color instead of word reading in the Stroop task (the tasks will be
described later). The question is how people know what adjustments are appropriate in a
current situation. One possibility is that error itself is informative in terms of how
behavior should be modified. The other is that an error signals only generically that
something has failed and people rely on various heuristics to determine what went wrong
and what adjustments need to be made.
In spite of the pervasiveness of errors in behavior and their role in learning and
performance, there is relatively little behavioral research on the topic (e.g. Norman, 1981;
Reason, 1990). Within the last decade, following the discovery of a neural signature for
errors (Falkenstein, Hohnsbein, Hoorman, & Blanke, 1991; Gehring, Goss, Coles, Meyer,
& Donchin, 1993), there has been an explosion in neuroscience research on error
detection mechanisms (e.g. Botvinick, Braver, Barch, Carter, & Cohen, 2001; Carter,
Braver, Barch, Botvinick, Noll, & Cohen, 1998; Holroyd & Coles, 2002; Miltner, Braun,
& Coles, 1997). An area in the medial frontal cortex, the anterior cingulate cortex (ACC),
has been found to be active during error commission (Carter et al., 1998). Understanding
 

1

the nature of the error signal that the ACC carries – whether or not the signal is contentspecific – could provide insights into the question of how the system determines what
went wrong and how to adjust behavior following an error.
1.1 The ACC Involvement in Error Detection
Research accumulated during the past decade agrees that the ACC is involved in
error detection; however, the precise role of the ACC is still a subject of debate. Several
theories of ACC function have been proposed; among them are a mismatch theory, a
conflict monitoring theory, and a reinforcement learning theory. None of the theories
addresses the question of whether the ACC carries content-specific information, the
information about the type of an error.
1.1.1 Mismatch (or comparator) theory of ACC function.
In the mismatch view, the medial frontal cortex detects a mismatch between
actual and intended responses. The theory was proposed by Gehring and colleagues
(1993) who found a negative going deflection in event-related potentials (ERPs)
originating in the medial frontal cortex following an error commission. The error related
negativity, known as ERN, correlated with a number of measures of compensatory
behavior.
In this study, participants performed a version of the Eriksen flanker task. On
each trial, participants viewed an array of letters (e.g., HHSHH) and their task was to
respond to the central letter by squeezing a dynamometer in the corresponding hand (e.g.,
left hand for S and right hand for H). On congruent trials, all letters in the array were
identical – HHHHH or SSSSS; on incongruent trials, the center letter was different from
 

2

the flankers – HHSHH or SSHSS. As in previous studies (Eriksen & Schultz, 1979),
incongruent trials resulted in lower accuracy and higher reaction time (RT) compared
with congruent trials. In this study, the authors found a sharp negative going deflection
100 ms after the onset of electromyographic (EMG) activity in the limb that was about to
make an error. Dipole modeling suggested a medial frontal generator of the negativity,
possibly the ACC (ERPs do not allow conclusions regarding precise spatial localization).
The relation of the ERN to error detection is also evident from the correlations of
the amplitude of the signal with three measures of compensatory behavior. First, the
authors found that large ERN was associated with greater probability of correcting the
error on the current trial. Second, the amplitude of ERN positively correlated with error
force (or amplitude of error squeeze measured in Kg). Finally, larger ERN was associated
with higher RT on the next correct trial. The association of the frontal negativity with
error trials and measures of compensatory behavior suggested that ERN was a neural
signature of an error detection mechanism.
The idea that ERN reflects a mismatch between actual and intended response was
largely driven by the nature of errors in the task used in the study. In the Eriksen flanker
task, the majority of errors are caused by impulsive behavior, when a response is issued
before the processing of the central stimulus is complete. The authors hypothesized that
an error is detected when an efferent copy of the actual response is compared with the
correct response that soon becomes available.

 

3

The effects of error detection, discussed within this work, have a general nature.
The system adjusts the overall speed of processing. The trials following error trials have
usually higher RTs and probably as a result of that higher accuracy.
1.1.2 Conflict monitoring theory of ACC function.
In the conflict monitoring view, the ACC detects a conflict or competition
between several simultaneously activated alternatives, and errors are just a special case of
conflict. The theory was proposed by Carter and colleagues (1998) who found that the
same area within the ACC is active on both types of trials: error trials and correct trials
that involve response conflict. The ACC is believed to detect a presence of conflict and
transmit this information to the areas responsible for cognitive control, for example,
prefrontal cortex (PFC) (Botvinick et al., 2001). Kerns and colleagues (2004) provided
evidence for that idea showing that the ACC activation on a current trial was positively
correlated with PFC activation on a subsequent trial.
The ERP study done by Gehring and colleagues (1993) suggested that the medial
frontal cortex, possibly the ACC, responds to errors. A methodology that has a better
spatial resolution than ERPs was required to confirm that the ACC was the locus of the
activation on error trials. Carter and colleagues (1998) used fMRI to address the question.
In their study, participants performed an AX-CPT task. In this task, the sequence of
events was as follows: First, participants saw a letter A or a letter B, which could be
followed by the letters X or Y. The participants’ task was to press a button when the
sequence AX was presented, otherwise the response should have been withheld. The
authors found that an area within the ACC was more active on error trials (e.g., when

 

4

participants responded to AY, BX, or BY sequences) compared with correct trials. This
finding confirmed the idea that the ACC responds to errors.
In addition to comparing error and correct trials, Carter and colleagues (1998)
also examined the ACC activation on correct trials alone. They separated all correct trials
into high conflict and low conflict trials. Conflict here is simultaneous activation of
responding behavior and withholding a response. AY or BX trials were classified as high
conflict trials, because, although the participants had to withhold response on these trials,
letters A and X could activate responding at least partially. BY trials were classified as
low conflict trials, because participants were to withhold their response on these trials and
neither of the letters was associated with pressing a button. The authors found that the
same area of the ACC that was active on error trials was also more active on high conflict
correct trials compared to low conflict correct trials. Carter and colleagues (1998)
concluded that the ACC performs a general conflict monitoring function and errors are
another instance of conflict.
Botvinick and colleagues (2001, 2004) elaborated the idea that the ACC serves a
performance monitoring function and proposed that conflict (or error) detection triggers
top-down control mechanisms. Thus, episodes of high conflict engage cognitive control,
whereas episodes of low conflict relax it. Botvinick and colleagues (2001, 2004)
modified computational models of the tasks that usually show ACC engagement (e.g.,
Eriksen flanker task, the Stroop task) by including a conflict-monitoring unit. In their
model, the conflict-monitoring unit computes level of conflict on a current trial and
transmits this information to a control unit, which is biasing processing on the following
trial.
 

5

Thus, in the Eriksen flanker task described earlier, the conflict arises on
incongruent trials – HHSHH or SSHSS– when flankers and the central letter
simultaneously activate responding with left and right hands. The conflict-monitoring
unit registers the level of conflict and transfers the information to the control unit, which
biases processing to the central stimuli on the next trial. In the Stroop task, participants
are presented with words written in different ink colors and their task is to name the ink
color and ignore the words. The conflict, in this task, arises when participants are
presented, for example, with the word RED written in blue color. Although the task is to
name the ink color – blue – word reading being automatic can partially activate an
incorrect response – red. A conflict-monitoring unit registers the presence of the conflict
and transmits this information to control units, which biases processing toward color
naming on the next trial.
In the model proposed by Botvinick and colleagues (2001, 2004), the level of
control is not fixed as in previous models, but is flexibly modified throughout the course
of a session by the presence of conflict. This addition allows the model to accommodate a
number of effects reported previously: for example, sequential adjustment effects in the
Eriksen task and the effect of trial-type frequency in the Stroop task. One of the
sequential adjustment effects in the Eriksen task is that an incongruent trial preceded by
another incongruent trial (iI trial) has lower RT than an incongruent trial preceded by a
congruent trial (cI trial). A current incongruent trial (I) seems to have a benefit of control
being engaged by the preceding incongruent trial (i). The Stroop model simulates the
effect of trial-type frequency on performance. The difference in performance (i.e. RTs
and accuracy) between incongruent and congruent trials is larger when incongruent trials
 

6

are infrequent (20%) compared with when incongruent and congruent trials are more
balanced (50%). Thus the level of control is not hard-coded by an experimenter in the
models, but is a result of the amount of conflict accumulated throughout an experiment.
In the simulations, the conflict-monitoring unit influences the strength of the
control and it does not provide information regarding the content of the adjustment that
need to be made. Its influence is quantitative rather than qualitative. The nature of the
adjustments is task specific and is determined a priori by the experimenter. The system is
never faced with the question of what adjustments are appropriate in a current situation.
In the Eriksen task, the control units bias processing toward the central letter, whereas in
the Stroop task, the control units bias processing toward color naming. The assumption of
the model is that the conflict monitoring units provide only non-specific information
regarding the occurrence of conflict. However, the authors noted that “conflict
monitoring need not be limited in this way” and that it can convey qualitative information
regarding the content of conflict.
Botvinick and colleagues (2001, 2004) proposed that conflict monitoring and
cognitive control are related based on the behavioral findings that episodes of conflict are
usually followed by superior performance in terms of RTs and accuracy. Kerns and
colleagues (2004) provided imaging evidence for the idea. They showed that the ACC
activation on a current trial was positively correlated with behavioral adjustments and
with PFC activation on the subsequent trial.
In this study, the authors found that iI trials in the Stroop task activated the ACC
significantly less than cI trials (see also Botvinick et al., 1999), showing that the system

 

7

experiences less conflict on the incongruent trial (I) preceded by another incongruent trial
(i). Kerns and colleagues (2004) further explored the effect of the ACC activation on the
adjustments in control by dividing all iI trials into low- and high-adjustment trials. Trials
that had RTs above the median RT for all iI trials were classified as low-adjustment trials,
and trials that fell below the median were classified as high-adjustment trials. Highadjustment trials were associated with greater ACC activation on the previous trial. Thus,
the more active the ACC is on the preceding trial, the greater the benefit on the current
trial. Moreover, the authors showed that the increase in the ACC activation on the
preceding trial was significantly correlated with the increase in PFC activation on the
current trial.
These findings supported the idea of Botvinick and colleagues (2001, 2004) that
the conflict monitoring function presumably performed by the ACC communicates the
information about the occurrence of conflict to control mechanisms implemented in the
PFC. However, the question of whether the ACC merely modulates the strength of the
PFC or provides a content-specific signal remains to be addressed.
1.1.3 Reinforcement learning theory of ACC function.
In the reinforcement learning view, the ACC receives rather then generates an
error signal (e.g. Holroyd & Coles, 2002). The errors are detected by mesencephalic
dopamine system, which sends an error-related signal to the ACC to train it to select an
appropriate motor controller (e.g., amygdala, dorsolateral prefrontal cortex – DLPFC
cortex). Thus, in this view, the ACC serves as a motor control filter.

 

8

The reinforcement learning (RL) theory is an attempt to unify two bodies of
literature: reinforcement learning and error detection (or conflict monitoring). According
to the theory, errors are detected by basal ganglia that compute a difference between an
expected value of the ongoing events and the outcome. When ongoing events are worse
than expected basal ganglia issue a signal to the ACC. This signal trains the ACC to
select the motor controller that has greater chances to be successful at carrying out a task.
The motor controllers mentioned in the model are amygdala, DLPFC, and orbitofrontal
cortex (OFC). Each of the motor controllers could solve a problem or perform a task in a
different fashion. The ACC decides which of the motor controllers will have an influence
on the motor system. In the model, the ACC does not necessarily have to choose one of
the controllers over the others; it can also combine input of several controllers in a
particular manner. Thus, in this view, the ACC performs an action selection function
rather than a performance monitoring function.
Experimental support for the RL theory mainly comes from an ERN-like signal
produced by the medial frontal cortex in response to error (or negative) feedback. The
signal was first observed by Miltner and colleagues (1997) and received the name
feedback error-related negativity (fERN). In this study, participants performed a time
interval production task. They were asked to press a key one second after the onset of a
trial. Feedback indicated whether the accuracy was in an acceptable range. The authors
found a negative going deflection in ERPs after the onset of negative feedback at about
250 ms. Dipole modeling suggested a medial frontal generator for the feedback errorrelated negativity. Based on the similarity in the ERN and the fERN locus, Miltner and
colleagues (1997) concluded that the same cognitive process underlies both signals.
 

9

Whereas the fERN represents a problem for the conflict monitoring theory as
there is no competition at the feedback point, the RL theory accounts nicely for it.
Moreover, it explains why the ERN propagates from feedback to response with learning
as was shown by Holroyd and Coles (2002). In this study, participants were asked to
press the left or right button in response to a stimulus. The stimulus-response mapping
was not given to the participants and had to be inferred in a course of the experiment.
There were six stimuli in total. One stimulus was mapped onto the left button and another
was mapped onto the right button – 100% mapping condition. For the other two stimuli,
the feedback was delivered independently of participants’ performance: on half of the
trials, they received a positive feedback and, on another half of the trials, they received a
negative feedback – 50% mapping condition. On the trials when the fifth stimulus was
presented, participants always received a positive feedback – always correct condition.
Finally, on the trials when the sixth stimulus was presented, participants always received
a negative feedback – always incorrect condition.
At the beginning of the experiment, the ERN was registered following the
negative feedback for all conditions. However, as participants progressed through the
session and learning occurred, the ERN moved from feedback to the point of response in
certain conditions. Thus, on trials with 100% mapping, the amplitude of ERN was larger
at the response than at the feedback after learning has occured. As the system acquired
the stimulus-response mapping, it began to rely more on its own representations of what
the correct response should be rather than the accuracy feedback. In contrast, on trials
with 50% mapping, the ERN continued to be elicited primarily by the feedback
throughout the session. The results of the simulation were in agreement with ERP data.
 

10

The RL theory of ACC function (Holroyd & Coles, 2002) has advantages over the
conflict monitoring and the mismatch theories as it accounts for both the ERN and the
fERN. However, the theory does not model any explicit error-detection mechanism.
Information about accuracy is provided to the model by the experimenter.
In the model, basal ganglia send a non-specific signal to the ACC about the
occurrence of an error in the form of a phasic reduction of dopaminergic input to the area.
The signal punishes the ACC for its choice of a motor controller on the current trial. It is
not content specific itself, although it is translated into a preference for one motor
controller – DLPFC – over another – amygdala. However, the signal is not informative in
terms of what particular representation or task set should be strengthened within the
DLPFC, for example.
*

*

*

To summarize, the models described above – the mismatch model, the conflict
monitoring model, and the reinforcement learning model – differ in terms of a particular
function the ACC supports –mismatch detection, conflict monitoring, or selection for
action. However, all three models agree that the ACC issues a signal in response to an
error occurrence. The error-related signal, as it is currently implemented in the models, is
not task-specific. The present work aims to extend the models by addressing the question
of whether the signal that the ACC issues is content specific. Answering this question
could shed light on to the larger problem of how people decide what went wrong and
what behavioral adjustments are appropriate in the current situation. If the ACC does

 

11

carry qualitative information, that would suggest that an error signal itself is informative
in terms of the type of the failure that has occurred.
1.2 Overview of Current Research
The aim of the current work is to examine neural substrates of an error detection
mechanism in the context of a complex task in which different types of errors can
potentially occur and the system is faced with a decision regarding what went wrong and
how to adjust behavior appropriately. In particular, I am interested to see whether an area
in the medial frontal cortex often associated with error detection -- the ACC –
discriminates between different types of error and/or discriminates between the decisions
people make regarding the type of an error. If the ACC discriminates between error types
then it can potentially be one of the sources of information that higher-level areas (e.g.,
PFC) use to make decisions regarding what went wrong. However, if the ACC
discriminates between decision types, or different interpretations of an external error
signal, then it can potentially be the locus of such a decision process itself.
To address these questions I developed a paradigm in which a memory task and a
multiple object tracking task are combined. This paradigm creates a situation where
multiple things can go wrong – memory or tracking failures. Participants receive
feedback on the accuracy of their performance only after both tasks are completed.
Negative feedback is ambiguous as it informs participants that they were inaccurate but
does not specify what type of error has occurred (memory or tracking error). Participants
are given a chance to correct a mistake if it occurs. The corrective actions suggest how

 

12

participants interpreted the negative feedback—whether they believed they made a
memory or tracking error.
In this paradigm, there are two independent variables: error type and
interpretation type. There are trials on which participants correctly interpret negative
feedback. For example, they make a memory error and classify it as a memory error.
There are also trials on which participants incorrectly interpret negative feedback. For
example, they make a memory error, but classify it as a tracking error. The same applies
to tracking errors. This dissociation between error type and interpretation type allows me
to address the question regarding the place of the ACC in the error processing stream the question of whether the computations that the ACC supports occur prior to the
judgment about the type of error or after the judgment is made. If the ACC responds to
the actual error type then it is likely to be one of the sources of information that higherlevel areas (PFC) use to make the judgment about the error type. Other sources of
information could be error history (e.g., memory errors are more frequent in this task), or
various explicit heuristics (e.g., “classify an error as tracking if the tracking task was
long”). However, if the ACC responds to interpretation type, then the computations that it
supports occur either after the judgment is made (and the ACC potentially selects
corrective actions) or it is one of the areas that actually make such a judgment.
In a behavioral experiment – Experiment 1 in this document – I test the paradigm.
In particular, I address the question of whether people can make accurate judgments
about the error type in the complex task I developed. In the imaging experiment –
Experiment 2 in this document – I address two questions: The first concerns the neural
source of the error signal that differentiates between error types. Does the ACC respond
 

13

differentially as a function of error type? The second question concerns the neural source
of participants’ interpretation of negative feedback. Does the ACC respond differentially
as a function of interpretation type?
In the present research, I assume that there are potentially different kinds of errors
that the mind can make. Hypothetically, there are as many types of error as basic
cognitive and perceptual processes. I chose here memory and tracking errors, primarily
because memory and tracking processes are well-studied. We know that relatively
different neural networks support memory retrieval and tracking suggesting that these are
two different processes (e.g. Jonides, Lewis, Nee, Lustig, Berman, et al., 2008 – for
memory; Culham, Cavanagh, & Kanwisher, 2001 – for tracking). So, the assumption is
that different processes produce qualitatively different errors. However, it is also possible
that in spite of the variability in the functions that the mind performs, all errors are
detected similarly, for example, as mismatches between two or more representations (e.g.
Gehring, Goss, Coles, Meyer, Donchin, 1993). Therefore, the question of what is the
universe of errors undoubtedly requires future research.

 

14

2 ERROR DETECTION IN COMPLEX TASKS
2.1 Experiment 1: Behavioral Study of Error Detection
In Experiment 1 I addressed the question of whether people are able to make
informed decisions about what kind of error they make. For that purpose, I developed a
paradigm in which a working memory task and a multiple object tracking task were
combined to create a complex situation in which multiple things can go wrong (i.e.
working memory or tracking failures). In this task, delayed error feedback indicates only
generically that something went wrong. Participants had to decide on their own what kind
of error had occurred and make corresponding behavioral adjustments.
If people have access to the information regarding the type of error they have
made on each trial, then I would expect participants’ interpretations of a negative
feedback would correspond to the error type. That is, participants are expected to correct
working memory errors as working memory errors and tracking errors as tracking errors.
If, however, people do not have access to the information regarding the type of
error, then they might either guess what the error was or they might attempt to match
probabilities. In the case of guessing, the distribution of interpretation types should not
differ from random distribution. In the case of matching probabilities, the frequency of
interpretation types are expected to match the frequency of error types regardless of what
particular kind of error has occurred. For example, if working memory errors are more
frequent than tracking errors, then participants might correct most working memory
errors as working memory errors and, more importantly, most tracking errors as working
memory errors.
 

15

2.1.1 Method
Participants. Fourteen undergraduate students from the Michigan State University
Psychology Department subject pool participated in exchange for course credit. All
participants had normal or corrected-to-normal vision.
Apparatus. Stimuli were presented on a 19″ CRT display monitor with a
resolution of 1024 x 768 pixels and a refresh rate of 117 Hz. Participants were seated at a
viewing distance of 80 cm from the computer display, with their head positioned on a
chin rest. A viewable screen subtended an angle of 28° by 21°. Stimulus presentation and
response collection were controlled by Experiment Builder software (SR Research Ltd.,
Mississauga, ON, CA).
Materials. In each trial, nine dots were displayed on the screen. The dots
consisted of blue outline rings with red interiors and were displayed on a white
background. The dots were opaque so that when two dots intersected each other, one was
shown to be in front of the other. The interior of a dot was 1.363 degrees of visual angle;
the outline ring was 0.055 degrees thick. Four of the dots formed a square, three dots
formed a triangle, and two dots were placed elsewhere on the screen (Figure 1). The dots
within the square and the triangle were separated by approximately 4.226 degrees of
visual angle. The location of the shapes – left or right from the center of the display –
varied randomly from trial to trial.
During the tracking phase of a trial, each dot moved with a velocity and direction
that were changed at random. The velocities of the dots ranged from 1 to 9 degrees per
second. The directions were chosen from among eight equal divisions of the compass.

 

16

The rate of velocity and direction changes ranged from 150 to 300 ms. When a dot
reached an edge of the screen, the x or y velocity vector reversed its value, so that the dot
appeared to bounce off the edge. The dots’ trajectories were generated in real time during
each trial.
The accuracy feedback was indicated by a colored square (Figure 1, Feedback
panel). For half of the participants, a blue square meant “correct”, and a green square
meant “incorrect”; for the other half of the participants, the meaning of the color
feedback was reversed. The square appeared in the center of the screen for 2000 ms and
subtended 1.772 degrees of visual angle.
Procedure. Participants were tested individually. At the start of the one-hour
session, they read through a description of the task. Then they were given five practice
trials. After the practice trials, the participants had an opportunity to ask questions about
the task. Eighteen blocks of eight trials each were then presented, totaling 144 trials per
participant.
Each block of eight trials started with an encoding phase, during which eight
randomly chosen shapes – squares and triangles - were presented sequentially (Figure 1).
Each shape was displayed for 500 ms. The participant’s task was to remember the
sequence. The sequence instructed participants what shape would be relevant on each of
the eight trials of the block: the first shape in the series was relevant in the first trial, the
second shape in the second trial, and so on.
The beginning of each trial was marked with a fixation cross, which was
displayed for 2,500 ms. Then the participant was shown a display consisting of nine dots:
three dots arranged in a triangle, four dots arranged in a square and two dots placed

 

17

elsewhere in the display. That was the memory retrieval phase. The participant’s task was
to retrieve the shape that was relevant on that trial, and attend to the dots that formed that
shape. For example, if participants saw the sequence of shapes displayed in Figure 1, then
on the first, second, fifth, seventh, and eighth trials, they had to attend to dots that formed
a triangle; and on the third, forth, and sixth trials, they had to attend to dots that formed a
square. The participant had 4 s to choose the correct shape, after which all dots began
moving randomly on the screen. During the tracking phase, the participant’s task was to
keep his/her attention only on the relevant dots. The tracking phase was randomly
assigned a duration of 5.0 s, 7.5 s, or 10 s. After all dots stopped moving, the participant
was asked to indicate with mouse clicks which dots they tracked. As the participant
clicked on a dot, an “x” sign appeared, covering the selected dot. The sign stayed on
display until the end of the trial, telling the participant which dots they had selected. A
color feedback then appeared, indicating whether they were accurate. After a positive
feedback, the trial ended. After a negative feedback, the participant had a chance to
correct the mistake by choosing dots one more time.

 

18

Figure 1. An example of a trial with a negative feedback in a modified multiple object
tracking task. (For interpretation of the references to color in this and all other figures, the
reader is referred to the electronic version of this dissertation.)
2.1.2 Results and Discussion
In this experiment, two types of error occurred: working memory and tracking
errors. I classified a response as a working memory error if participants clicked on three
dots during Response 1 (Figure 1) when a square was relevant on that trial; or when they
clicked on four dots when a triangle was relevant on the trial. A response was classified
as a tracking error when participants clicked on the correct number of dots, but one or
more dots were not a part of the to-be-tracked subset. Participants were considered to
have made both types of errors when they clicked on the wrong number of dots and at
least one of those dots did not belong to the to-be-ignored shape.
The participants made errors on 33% of the trials. Memory errors were almost
twice as frequent as tracking errors. Participants forgot what shape they had to attend to
on 18% of the trials (middle bar in Figure 2). They failed to track dots accurately on 10%
of the trials (left most bar in Figure 2). On five percent of the trials, participants made
both types of error (right-most bar in Figure 2). First, they chose a wrong subset of dots

 

19

to track during the retrieval phase and then they lost a dot or several dots during the
tracking phase.
Following negative feedback, participants had a chance to correct their errors by
choosing dots one more time during Response 2 (Figure 1). The corrections participants
made were informative in terms of the error interpretations they made. Two types of
interpretations were possible: working memory interpretations and tracking
interpretations. A working memory interpretation means that a participant realized they
had tracked a wrong shape and decided to switch to a different shape. An interpretation
was classified as working memory if the number of clicked dots differed between
Response 2 and Response 1. In other words, a working memory interpretation was
registered if, during Response 1, a participant chose three dots and during Response 2
they chose four dots and vice versa. A tracking interpretation occurred when a participant
believed that he or she chose the correct shape to track, but lost one or more dots during
the tracking task. An interpretation was classified as tracking if the number of clicked
dots in Response 1 and Response 2 was the same.
The corrections that the participants performed following a negative feedback
suggest that people make informed decisions regarding what kind of error has just
occurred. Figure 2 shows that 72% of the memory errors were interpreted as memory
errors (grey shading of the middle bar), and that 67% of the tracking errors were
interpreted as tracking errors (white shading of the left most bar). A Chi-square test
indicates that types of behavioral adjustments depended on types of error made χ2 (1) =
5.78, p < .05. Not surprisingly, half of the trials on which participants made both types of

 

20

error, were interpreted as trials with memory errors and half of the trials were interpreted
as trials with tracking errors.
In sum, the paradigm I developed created a complex situation where different
types of error occasionally occurred (i.e. memory or tracking errors). Given delayed error
feedback that indicated only generically that something went wrong, participants most
often correctly identified the type of error they made. The results suggest that the system
does detect and represent information about processing errors that occur in the course of
performing a complex task.
0.2
Memory interpretation
Tracking interpretation

Frequency (proportion)

0.18
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
Tracking errors

Memory errors

Both

Error Type

Figure 2. Frequency of different types of interpretations as a function of error type in
Experiment 1.
2.2 Experiment 2: Imaging Study of Error Detection
In Experiment 2, I address two questions concerning the neural substrates of an
error detection mechanism: The first question is whether the ACC represents information
about an error type (i.e. memory versus tracking errors). In the behavioral experiment, I
 

21

determined that people do have access to the information regarding the type of error they
have just committed—in more than 70% of cases they correctly identify the type of error
that has just occurred. In the imaging study, I identify whether the ACC is the brain
region that carries this information. The second question concerns participants’
interpretation of the performance on a task—does the ACC represent participants’
judgment regarding what went wrong? In the behavioral task, I found that, for example,
some of the tracking errors are classified as tracking, whereas other tracking errors are
classified as memory errors (the same is true for memory errors). In the imaging
experiment, I explore whether the ACC codes participants’ interpretation of an error
regardless of an actual error type.
2.2.1 Method
Participants. Twenty-one right-handed, healthy undergraduate students from the
Michigan State University Psychology Department subject pool participated in the
experiment. All participants gave a written consent according to the study protocol that
has been approved by the Institution Review Board at the Michigan State University. All
participants had normal or corrected-to-normal vision. The participants were rewarded
with either course credits or $20 whichever they preferred.
Given that the study focused on errors and that participants had variable error
rates (e.g. some participants did not produce enough error trials, others did not produce
enough correct trials to fill the cell of the design), participants whose accuracy was higher
than 60% or lower than 10% were excluded from the analysis. Seventeen participants out
of twenty-one met this accuracy criterion (three participants were excluded because their

 

22

accuracy was higher than 60%, and one participant was excluded because his accuracy
was lower than 10%).
Apparatus. Stimuli were presented on a rear projection screen, a 1024 x 768 32inch LCD monitor (Salvagion Design, Sausalito, CA), located in the scanner bore.
Participants viewed the screen via an angled mirror attached to the head coil. A viewable
screen subtended a visual angle of 13° by 10°. Stimulus presentation and response
collection were controlled by Experiment Builder software (SR Research Ltd.,
Mississauga, ON, CA).
Materials. In each trial, fourteen dots were displayed on the screen. The dots
consisted of blue outline rings with light purple interiors and were displayed on a white
background. The dots were opaque so that when two dots intersected each other, one was
shown to be in front of the other. The interior of a dot was 0.786 degrees of visual angle;
the outline ring was 0.026 degrees thick. Four of the dots formed a square, another four
of the dots formed a diamond, three dots formed an upward pointing triangle, and other
three dots formed a downward pointing triangle, see Figure 3. The dots in the square, the
diamond, and the triangles were separated by approximately 1.96 degrees of visual angle.
The location of the shapes – left, right, up, and down from the center of the display –
varied randomly from trial to trial.
During the tracking phase of a trial, each dot moved with a velocity and direction
that were changed at random. The velocities of the dots ranged from 2 to 3.5 degrees per
second in the easy version of the task and from 2 to 4 degrees per second in the difficult
version. The directions were chosen from among eight equal divisions of the compass.
The rate of velocity and direction changes ranged from 500 to 1000 ms. When a dot

 

23

reached an edge of the screen, the x or y velocity vector reversed its value, so that the dot
appeared to bounce off the edge. The dots’ trajectories were generated in real time during
each trial.
The accuracy feedback was indicated by a colored square (Figure 3, Feedback
panel). For half of the participants, a red square meant “correct”, and a green square
meant “incorrect”; for the other half of the participants, the meaning of the color
feedback was reversed. The square appeared in the center of the screen and subtended
0.886 degrees of visual angle.
Procedure. Participants were tested individually in sessions lasting about 150
min. At the start of the session, they read through a description of the task. The
experimenter then presented three trials and demonstrated performance on a correct trial,
a tracking error trial, and a memory error trial, describing what participants need to do
during each phase of the task (e.g., make a judgment after negative feedback). Then
participants were given two blocks of five practice trials, totaling 10 practice trials per
participant. The practice trials were from an easy version of the task meaning that the rate
of shape presentation and dots’ velocity were lower. If participants made four or fewer
errors during the practice block then they were presented with a difficult version of the
task in the scanner. That allowed me to adjust the difficulty of the task for each
participant individually. Seven out of seventeen participants performed an easy version of
the task in the scanner.
The task I used for imaging was a modification of the task used in Experiment 1.
The modifications were made with two aims: to increase the number of errors and
decrease the likelihood of correctly guessing a shape during the memory retrieval phase

 

24

of the task. The modified task is presented in Figure 3. In this task, trials were presented
in blocks of five trials each. Each block of five trials started with an encoding phase,
during which five randomly chosen shapes – squares, diamonds, upward pointing
triangles, downward pointing triangles – were presented sequentially (one shape was
repeated to create a five-shape sequence). Each shape was displayed for 550 ms in the
easy version of the task and for 500 ms in the difficult version of task. The participants’
task was to remember the sequence. The sequence instructed participants what shape
would be relevant on each of the five trials of the block as in the Experiment 1.
The beginning of each trial was marked with a fixation cross. Then the participant
was shown a display consisting of fourteen dots: four dots were arranged in a square, four
dots were arranged in a diamond, three dots were arranged in an upward pointing
triangle, and three dots were arranged in a downward pointing triangle. This was the
memory retrieval phase. The participant’s task was to retrieve the shape that was relevant
on that trial, and attend to the dots that formed that shape, as participants did in
Experiment 1. After that, all dots began moving randomly on the screen. During the
tracking phase, the participant’s task was to keep their attention only on the relevant dots.
After all dots stopped moving, the participant was asked to indicate with mouse clicks
which dots they tracked. As the participant clicked on a dot, an “x” sign appeared,
covering the selected dot. A color feedback then appeared, indicating whether the
participant was accurate. After positive feedback, the trial ended. After negative
feedback, the participant was asked to interpret the negative feedback and indicate what
type of error they thought they had made. This was the judgment phase. The participants
were asked to click with the mouse on the “Shape” button if they believed they had made

 

25

a memory retrieval error, and on the “Tracking” button if they believed they had made a
tracking error. The locations of “Shape” and “Tracking” buttons were counterbalanced,
such that for the half of the participants “Shape” button was displayed above “Tracking”
button, and for the other half of the participants “Tracking” button was displayed above
“Shape” button.
The length of all phases—Retrieval task, Tracking task, Response, Feedback,
Judgment—was predetermined by the experimenter and varied to create jittering. The
phases were randomly assigned to the following durations: Fixation cross—2.5 s, 5 s, or
7.5 s; Retrieval phase—2.5 s or 5 s; Tracking phase—5 s or 7.5 s; Response—7.5 s or 10
s; Feedback—2.5 s or 5 s; Judgment—always 5 s. Participants performed 18 blocks of
five trials in the scanner, totaling 90 trials per participant.

Figure 3. An example of a trial with a negative feedback in a modified task used for
imaging study.
Image acquisition. Magnetic resonance imaging was conducted on a 3T Signa®
HDx MR scanner (GE Healthcare, Waukesha, WI) with an eight-channel head coil.
During each session, images were first acquired for the purpose of localization, and first
and higher order shimming procedures were then carried out to improve magnetic field

 

26

homogeneity (Kim et al., 2002). T2*-weighted echo planar imaging was performed with
the following parameters: 36 interleaved slices, slice thickness = 3 mm, TE = 27.7 ms,
TR = 2,500 ms, flip angle = 80°, FOV = 22 cm, matrix size = 64 x 64, ramp sampling, the
first four data points were discarded. Each volume of slices was acquired 206 times
during each of the six functional runs, totaling 1236 volumes of images over the course of
the entire experiment. After functional data acquisition, high-resolution volumetric T1weighted fast spoiled gradient-recalled (FSPGR) images with cerebro-spinal fluid
suppressed were obtained to cover the whole brain. 180 1-mm sagittal slices were
acquired with flip angle of 8°, FOV = 256 mm, and matrix size of 256 x 256 with a voxel
size of 1 mm x 1 mm x 1mm.
fMRI data pre-processing and analysis. All fMRI data pre-processing and
analysis were conducted with AFNI software (Cox, 1996). For each subject, with the last
volume of functional images as a reference, rigid-body motion correction was done in
three translational and three rotational directions. The amount of motion in these
directions was estimated and then the estimates were included as regressors in general
linear model (GLM) during data analysis. For each subject, spatial blurring with a full
width half maximum of 4 mm was applied to reduce random noise (Parrish et al., 2000),
and also to reduce inter-subject anatomical variation and Talairach transformation
variation during group analysis. For the group analysis, all images were converted to
3

Talairach coordinate space (Talairach & Tournoux, 1988) with an interpolation to 1 mm .
Individual subject fMRI analysis. For the data analysis of each individual subject,
each phase of the task was modeled separately except Fixation cross displayed between
trials in a block and the Memory retrieval task, which were included in the model
 

27

together as memory related regressors. Three regressors were included for the Memory
retrieval phase: correct retrieval (Mcor), consistently interpreted memory error (Merr_m),
and inconsistently interpreted memory error (Merr_t). Three regressors were included for
the Tracking phase: correct tracking (Tcor), consistently interpreted tracking error
(Terr_t), inconsistently interpreted tracking error (Terr_m). Six regressors were included
for Feedback phase: feedback on correct trials (Fcor), feedback given Merr_m (FMer_m),
feedback given Merr_t (FMer_t), feedback given Terr_t (FTer_t), feedback given Terr_m
(FTer_m), and feedback give both types of errors (i.e., memory and tracking) were made
(Fboth). One regressor for the Response phase, one regressor for the Judgment phase, one
regressor for the Encoding phase were also included in the model. Therefore, 15
conditions were modeled, and the hemodynamic response (HDR) at each voxel with
respect to each condition was resolved with multiple linear regression using the
“3dDeconvolve” software in AFNI (Ward, 2000). The MRI signal model also included
the subject motion estimations in three translational and three rotational directions, and
the constant, linear, and quadratic trends for each of the six functional runs. The HDRs
were resolved to seven points from zero to 15 s at the resolution of 2.5 s (TR). The
BOLD signal change was calculated based on the area under the HDR curve. The
equivalent BOLD percentage signal change relative to the baseline state was then
calculated.
Whole-brain analysis for fMRI. After the percentage signal change was estimated
for each condition for each participant, an ANOVA was performed over the 17-subject
dataset for group analysis with mixed-effect two-factor model. The stimulus condition
was the first factor and was modeled as a fixed effect. Subject was the second factor and

 

28

was modeled as a random effect. Monte Carlo simulation of the effect of matrix and
voxel size of the imaging volume, spatial blurring, voxel intensity thresholding, masking,
and cluster identification was used to estimate the overall statistical significance with
respect to the whole brain (Ward, 2000). Based on these estimations, the ANOVA results
above were further corrected for multiple comparisons based on the following criteria:
The active voxel selection criteria required that the voxels had voxel-based p ≤ 1 × 10

-3

3

and were nearest-neighbor and within a cluster size of 177 mm . Based on application of
these criteria to the whole brain, the voxel-based p ≤ 1 × 10

-3

was corrected to be an

equivalent whole-brain corrected p ≤ 0.027.
2.2.2 Behavioral Results
In this experiment, two types of errors occurred: memory retrieval and tracking
errors. The participants made errors on 63% of the trials. Tracking errors were more
frequent than memory errors. Participants failed to track dots accurately on 30% of the
trials (Figure 4, the left most bar). They forgot what shape they had to attend to on 16%
of the trials (Figure 4, the middle bar). On 18% of the trials, participant made both types
of error (Figure 4, the right most bar). First, they chose a wrong subset of dots to track
during the retrieval phase and then they lost a dot or several dots during the tracking
phase.
Following negative feedback, participants were asked to interpret the feedback
and to make a judgment regarding what type of error they thought they had made. Two
types of interpretations were possible: memory interpretations and tracking
interpretations. The observed judgments suggest that people make informed decisions
regarding what kind of error has just occurred. Figure 4 shows that 83% of tracking errors
 

29

were interpreted as tracking errors (grey shading of the left most bar), and 75% of
memory errors were interpreted as memory errors (white shading of the middle bar). Not
surprising, half of the trials, on which participants made both types of error, were
interpreted as trials with memory errors and half of the trials were interpreted as trials
with tracking errors (the right most bar). A Chi-square test indicates that the distribution
of types interpretations differed from 50% x 50% for both memory (χ2 (1) = 25, p <
0.001) and tracking errors (χ2 (1) = 43.56, p < 0.001). These results suggest that
participants did not merely guess about type of error they had made. Thus, I replicated
here the results that were obtained in Experiment 1. The system does represent the
information regarding error type, since participants most often identify errors correctly.
0.35
Memory interpretation
Tracking interpretation

Frequency (proportion)

0.3
0.25
0.2
0.15
0.1
0.05
0
Tracking errors

Memory errors

Both

Error Type

Figure 4. Frequency of different types of interpretations as a function of error type in
Experiment 2.

 

30

2.2.3 fMRI Results
2.2.3.1 Whole brain analysis: all error versus correct trials during Task phase.
The tasks I used in the present study – memory retrieval and multiple object
tracking tasks – are not commonly used in error detection research. Therefore, first, I
wanted to determine if errors in these tasks activate dACC as they do in more traditional
tasks such as the Eriksen, Stroop, or go/no go tasks. Contrasting activation on all error
trials and correct trials (i.e. (Terr_t + Terr_m + Merr_t +Merr_m) – 2(Tcor +Mcor)), I
identified a single region in the left dACC, in BA 32, which is more active on error trials
compared to correct trials, see Figure 5 and Table 1. The region with a significant effect
extends from BA 32 dorsally into BA 8, which is consistent with previous studies of error
detection and conflict monitoring (e.g., Brown, 2009, Carter et al., 1998). This finding
suggests that errors in a memory retrieval task and a multiple object tracking task activate
the dACC (the implications of this finding are considered in the General Discussion
section).

Figure 5. Error-related activation. Regions identified in a whole brain analysis
contrasting all error trials and all correct trials (Terr_t + Terr_m + Merr_t +Merr_m) – 2
(Tcor + Mcor). The arrow points to a region in left dorsal ACC, which is more active for
error trials.
 

31

Table 1. Group analysis results for the contrast: “all errors > all correct”.

Cluster
Active
volume

L dACC

mm3

Max t
values
Max t value
locations

255

5.735 

RL, AP, IS

6,-24, 40

Merr_t
Merr_m 

0.17(0.08)

Terr_t 

0.59(0.10)

Terr_m 

0.47(0.09)

Mcor 

-0.08(0.10)

Tcor 

 

-0.03(0.07)

0.23(0.06)

32

2.2.3.2 ROI analysis: dACC during Task phase.
Having identified a region in dACC responding to errors, I want to explore this
region further and address the main questions of the present study: (1) Does the dACC
respond differentially as a function of error type? and (2) Does the dACC respond
differentially as a function of interpretation type?
To address these questions, I compared ACC activation during memory retrieval
and tracking phases of a trial. I conducted a repeated-measures ANOVA (2 x 2) with
Error Type (Tracking errors versus Memory errors) and Interpretation Type (Tracking
interpretations versus Memory interpretations) as within subject factors, see Figure 6.
The main effect of Error Type was significant, F(1, 16) = 8.919, p = 0.009. Thus, the left
dACC region was more active on Tracking error trials compared to Memory error trials.
The main effect of Interpretation Type was not significant, F(1,16) = 0.295, p = 0.594.
Activation in the left dACC did not differentiate between types of interpretation. More
importantly, the interaction of Error Type and Interpretation Type approached
significance, F (1,16) = 3.321, p = 0.087. Thus, tracking errors that were interpreted as
tracking errors (Terr_t) and memory errors that were interpreted as memory errors
(Merr_m) activated the dACC more than tracking errors that were interpreted as memory
errors (Terr_m) and memory errors that were interpreted as tracking errors (Merr_t). This
result suggested an effect of Consistency of interpretations. Estimated hemodynamic
response functions (HDR) are presented in Figure 7.

 

33

0.7
Tracking interpretation
Memory interpretation

% BOLD signal change

0.6
0.5
0.4
0.3
0.2
0.1
0
-0.1
Tracking errors

Memory errors

-0.2
Error Type

Figure 6. ROI analysis: the left dACC. The region shows an Error Type effect: tracking
errors activate the region greater than memory errors, consistent with the idea that the
ACC represents information regarding the types of errors people make. There is no effect
of Interpretation Type. The interaction of Error Type and Interpretation Type approached
significance, suggesting an effect of Consistency of interpretation: the ACC responds
greater to consistently interpreted errors (Terr_t and Merr_m) compared to inconsistently
interpreted errors (Terr_m and Merr_t). The error bars show SEM.

 

34

Figure 7. Estimated Hemodynamic Response Functions (HDR) for different types of
error during the Task phase. The graph shows the effect of Error type – blue lines above
red lines, and the effect of consistency – solid lines above dashed lines.

To explore the consistency effect further I re-represented the interpretation
information as Consistent and Inconsistent instead of Tracking and Memory. Figure 8
shows the same data as Figure 6, but here, Terr_t and Merr_m are coded as consistently
interpreted errors, and Terr_m and Merr_t are coded as inconsistently interpreted errors.
To follow up the marginally significant Consistency effect, pair-wise two-tailed t-tests
were conducted on Tracking and Memory errors separately (all t-tests presented in the
dissertation are two-tailed). The Consistency effect was significant for Memory errors, t
(16) = 3.534, p = 0.003, such that consistently interpreted Memory errors led to greater
dACC activation compared to inconsistently interpreted Memory errors. The Consistency
effect was not significant for Tracking errors, t (16) = 0.779, p = 0.436. After excluding
one outlier (i.e. 2 SDs above the mean for T_inconsistent), who did show greater

 

35

activation for T_consistent compared to T_inconsistent, but had extremely high values in
both conditions, which contributes quite a lot of variability, I found an approaching
significance effect of Consistency, t (15) = 1.916, p = 0.076.
0.7

Consistent
Inconsistent

% BOLD signal change

0.6
0.5
0.4
0.3
0.2
0.1
0
-0.1

Tracking errors

Memory errors

-0.2

Error Type

Figure 8. ROI analysis: the left dACC. The region shows Consistency effect: Errors
interpreted consistently lead to a greater dACC activation than errors interpreted
inconsistently. There is also an effect of Error Type. The interaction between Consistency
of Interpretation and Error Type is not significant. The error bars show SEM.
A whole brain analysis supports the finding of the ROI analysis. The dACC
shows greater activation for consistently interpreted errors compared to correct trials
(Terr_t + Merr_m) – (Tcor + Tcor), see Figure 9 and Table 2. At the same time,
inconsistently interpreted errors do not show activity in this region. Consistently
interpreted errors also activated the left medial frontal gyrus (MFG), and the left superior
frontal gyrus (SFG). The analysis and the discussion of the activation in these regions are
presented in Appendix A.

 

36

Figure 9. Consistently interpreted errors. Regions identified in a whole brain analysis
contrasting consistently interpreted errors and correct trials (Terr_t + Merr_m) – (Tcor +
Mcor). The arrows point to the regions more active for consistently interpreted errors.
They are two regions in the left dACC (big and small), the left MFG, and the left SFG.

Table 2. Group analysis results for the contrast: “Consistently interpreted errors >
correct”.

Cluster

Active
volume

Max t
values

mm3

Max t
value
locations
RL, AP, IS

Merr_m

Terr_t

Mcor

Tcor

L dACC
(large)

625

7.334 

6,-24, 41

.18(.08)

.54(.17)

-.03(.06)

.22(.09)

L dACC
(small)

204

6.116 

4, -29, 31

-.03(.05)

.30(.05)

-.14(.06)

.02(.08)

L&R
SFG

274

7.123 

-9, -13, 56

.04(.09)

.64(.18)

.02(.10)

.33(.15)

L MFG 

283

6.359 

40, -32, 30

.28(.07)

.39(.09)

.08(.05)

.10(.07)

 

37

Each trial in this study consisted of task performance (i.e., a memory retrieval
task, and a tracking task) and feedback processing. So far I have considered dACC
activation during task performance, the activation during a memory retrieval phase and a
tracking phase. I found that dACC responded to errors that would be consistently
interpreted later. It is reasonable to assume that consistently interpreted errors are the
errors that people actually detect and/or are aware of, and that people are more likely to
interpret an error inconsistently when they miss it. Making this assumption, I can
conclude that dACC responds to errors that are actually detected as the system performs a
task.
2.2.3.3 ROI analysis dACC: Task phase versus Feedback phase.
Next, I consider the activation in the left dACC during the feedback phase of a
trial and compare it to the activation during task performance. Figure 10 shows the
activation in the left dACC as function of consistency of interpretation and the phase of
the task when activation is measured – during task performance (Task) or during
feedback processing (Feedback). A repeated measures ANOVA (2 x 2) with Consistency
(Consistent versus Inconsistent) and Time Point (Task versus Feedback) as within subject
factors showed a significant interaction between factors, F(1, 16) = 6.021, p = 0.026.
Neither of the main effects were significant, p > 0.4. To follow up the interaction, pairwise t-tests were conducted. As expected, during Task, consistently interpreted errors led
to greater dACC activation compared to inconsistently interpreted errors, t(16) = 4.001, p
= 0.001. However, during Feedback, the pattern was reversed such that the dACC was
more active on inconsistently interpreted error trials than on consistently interpreted error
trials, the effect was approaching significance, t(16) = 1.703, p = 0.108.

 

38

0.7
Consistent
Inconsistent

% BOLD signal change

0.6
0.5
0.4
0.3
0.2
0.1
0
-0.1
Task

Feedback

-0.2
Time point

Figure 10. ROI analysis: the left dACC. The region shows Consistency by Time Point
interaction: During task performance, errors interpreted consistently lead to a greater
dACC activation than errors interpreted inconsistently. However, during feedback
processing, errors interpreted inconsistently lead to greater dACC activation than errors
interpreted consistently. Neither of the main effects is significant. The error bars show
SEM.
To follow up the approaching significance consistency effect during feedback
processing, I consider tracking and memory errors separately, see Figure 11. The
consistency effect was not significant for memory errors, t (16) = 0.687, p = 0.502.
Tracking errors showed a marginally significant consistency effect, t (16) = 1.983, p =
0.065, meaning that inconsistently interpreted tracking errors (FTer_m) activate the
dACC region greater than consistently interpreted tracking errors (FTer_t). The lack of
reliability in this test could be explained by high variance in FTer_m condition. The high
variability in the FTer_m is probably due to the fact that the parameter was estimated
based on a few trials (i.e. M = 4.8 SD = 2.8). To explore the effect of consistency I
examined outliers. After excluding two outliers – one scoring 2 SDs above the mean,

 

39

another scoring 2 SDs below the mean on FTer_m – the marginal effect of consistency
became significant, t (14) = 2.561, p = 0.023. I interpret this result as evidence for an
effect of consistency during the Feedback phase. However, at the time of writing I am
also collecting additional data to test whether the effect raches conventional levels of
significance. The HDR functions estimated based on the data from all 17 participants are
shown in Figure 12.

0.8
Consistent
Inconsistent

0.7

% BOLD signal change

0.6
0.5
0.4
0.3
0.2
0.1
0
-0.1
-0.2

Tracking

Memory

-0.3
Error Type

Figure 11. ROI analysis: the left dACC. The region shows Consistency effect during
feedback processing only for tracking errors: Errors interpreted inconsistently lead to
greater dACC activation than errors interpreted inconsistently. The effect of consistency
is not significant for memory errors. The error bars show SEM.

 

40

0.3

% BOLD signal change

0.25
0.2
0.15
FMer_m
FMer_t
FTer_m
FTer_t

0.1
0.05
0
-0.05

0

2.5

5

7.5

10

12.5

15

-0.1
-0.15
-0.2
Time (s)

Figure 12. Estimated HDR functions for all types of error during the Feedback phase.
2.2.3.4 Whole-brain analysis: Interaction between Consistency and Time point (only
for tracking errors).
A whole-brain analysis of the interaction between Consistency of interpretation
(Consistent or Inconsistent) and Time point (Task or Feedback) for tracking errors
revealed two areas – left MFG and left Caudate – that were more active during
consistently interpreted errors compared to inconsistently interpreted errors during Task,
but showed a reversed pattern during Feedback – the areas were more active during
inconsistently interpreted errors compared to consistently interpreted errors during
Feedback, see Figure 13 and Table 3. Activity in the dACC appears in the whole-brain
analysis only at lower thresholds, p = 0.003 (voxel-wise uncorrected). Thus the wholebrain analysis does not seem to be sensitive enough to detect dACC activation when
testing for the interaction of Consistency and Time point. Whole-brain analyses of the

 

41

interaction for memory errors and memory errors combined with tracking errors did not
reveal any areas of activation.

Figure 13. Consistency by Time point interaction. Regions identified in a whole-brain
analysis contrasting (Terr_t + FTer_m ) and (Terr_m + FTer_t ). The arrows point to the
regions more active for (Terr_t + FTer_m ). They are the left MFG and the left Caudate.

Table 3. Group analysis results for the interaction between Consistency and Task Phase.

Cluster

Active
volum
e
mm3

Max t
values

Max t value
locations
Terr_t

RL, AP, IS

Terr_m

FTer_t

FTer_m

.04(.05)

.35(.13)

L
Caudate

7.237  14, -2, 21

.13(.07)

-.12(.06)

L MFG

 

343

188

6.146  36, -26, 40

.12(.06)

-.07(.11) .01(.05)

42

.76(.31)

*

*

*

To summarize the results, consistently interpreted errors evoke dACC activity
during the task performance (memory retrieval and tracking), but inconsistently
interpreted errors evoke dACC activity during negative feedback. I assume that
consistently interpreted errors are detected errors, and inconsistently interpreted errors are
undetected errors. Under this assumption, the dACC seems to respond to errors when the
system detects an error and also to negative feedback when such feedback is unexpected,
i.e. when an error is not detected. This pattern of activation is consistent with the
Reinforcement Learning theory of ACC function (Holroyd & Coles, 2002), and it is
problematic to the Conflict monitoring theory (Carter et al., 1998). The present study
confirms the Reinforcement Learning theory’s assumption that the ACC is activated
when an outcome is worse than expected: Both events, a detection of an error and
negative feedback given after undetected errors, are worse-than-expected outcomes. The
Conflict monitoring theory has difficulty explaining ACC activation after negative
feedback as there is no conflict at that point. I will elaborate on these ideas in the General
Discussion section.

 

43

3 GENERAL DISCUSSION
Several accounts of dACC function currently exist – the mismatch theory (e.g.,
Gehring et al., 1993), the conflict monitoring theory (e.g., Carter et al., 1998; Botvinick et
al., 2001), the reinforcement learning theory (e.g., Holroyd & Coles, 2002), etc. The
results of the present imaging study challenge the idea that the dACC performs a
mismatch detection or conflict monitoring function, and they are more consistent with the
idea that the dACC responds to worse-than-expected outcomes. The finding that the same
region in the left dACC responds both to the errors that people are aware of and to the
unexpected negative feedback is consistent with the reinforcement learning account and it
is problematic for the conflict monitoring and mismatch detection theories.
In the following, I first discuss the finding that the left dACC responds to worsethan-expected outcomes. I explain why this is inconsistent with the conflict monitoring
account. Then, I revisit Holroyd and colleagues’ ERP study described in the introduction
and explain how the imaging results of the present study and their ERP results converge.
Next, I introduce an error-likelihood account of ACC function (Brown & Braver, 2005)
and reinterpret the findings taken to support this account in the light of the present results.
Then, I discuss the finding that the dACC responds differentially as function of error type
and speculate about what the underlying mechanisms could be. Finally, I discuss the
implications of the finding that the dACC responds to errors in tasks that do not involve
response conflict.
3.1 The dACC in response to worse-than-expected outcomes

 

44

In the imaging experiment, I used a complex task, in which participants first
performed two sub-tasks (a memory retrieval task and a multiple object tracking task)
then processed feedback on the performance. This design allowed me to conduct a within
subject comparison of dACC activation during task performance and during feedback
processing. Another important feature of the task is participants’ judgments after negative
feedback. These judgments allowed me to distinguish between detected and undetected
errors. On a single trial, participants performed two tasks – a memory retrieval task and a
multiple object tracking task. The feedback participants received after performing these
tasks was ambiguous, since it told them if they were accurate, but negative feedback did
not indicate on which of two tasks they made an error. After negative feedback,
participants were asked to make a judgment regarding the type of error. Most of the time
they made correct judgments – when they failed to retrieve a correct item from memory
they indicated that they made a memory error, and when they lost a track of a dot they
indicated that they made a tracking error. However, sometimes they made incorrect
judgments – when they made a memory error they indicated that it was a tracking error,
and when they made a tracking error they indicated that it was a memory error. My
assumption here is that participants are more likely to make a correct/consistent judgment
when they actually detect an error, and they are more likely to make an
incorrect/inconsistent judgment when they miss an error. Thus, the judgments allowed
me to separate all errors into detected and undetected errors.
Exploring the activity in the left dACC region identified in “all errors versus
correct” contrast I found an interaction between error detection and phase of the trial. The
region was more active to detected errors than undetected errors during task performance

 

45

but during feedback the pattern was reversed such that the region was more active to
negative feedback after undetected errors than to negative feedback after detected errors.
This pattern of results is problematic for the conflict monitoring theory of ACC function,
as there is no conflict during feedback processing. The results are, instead, consistent
with the reinforcement learning theory of ACC function, according to which the ACC
responds to worse-than-expected outcomes.
3.1.1 Evidence concerning the Conflict monitoring theory
The conflict monitoring theory can accommodate the fact that the ACC responds
to detected errors during task performance. During memory retrieval task, there could be
a conflict between multiple items competing for the retrieval similar to what has been
observed in a verb generation task (Barch, Braver, Sabb, & Noll, 2000).
In a verb generation task, participants are presented with a noun and their task is
to generate a verb associated with the noun. Some of the nouns have multiple verbs
equally strongly associated with them. For example, the noun “ball” is approximately
equally associated with multiple verbs, such as “hit”, “throw”, “bounce”, and “play”. In
contrast, other nouns have a few or one strongly associated verb. For example, the noun
“bell” is strongly associated with one verb -- “ring ” (Eighty three percent of subjects
produced that verb in response to the noun in a norming study). Barch and colleagues
(2000) found that nouns like “ball” lead to greater ACC activation compared to the nouns
like “bell”, presumably because “ball” primes a number of verbs that compete for the
retrieval. The finding that the dACC is more active in response to errors during memory
retrieval can be explained in a similar vein. On memory error trials, there is no a single
dominant item and multiple items create a competition that the ACC detects. On

 

46

undetected error trials, one of the items could erroneously gain greater activation
compared to other items, which could lead to participants being more confident in
accuracy of retrieval.
The conflict monitoring theory could also explain dACC activation during errors
in the multiple object tracking task. People are more likely to lose a dot when it intersects
with another dot. These intersections create competition between dots for attention. The
intersections are probably more frequent on error trials compared with correct trials. On
undetected error trials, the system perhaps did not engage a conflict resolution
mechanism as much as it did on detected error trials. Therefore, the dACC responds more
to detected errors compared to undetected errors.
Although the conflict monitoring theory can accommodate dACC activation to
errors and especially to detected errors during performance on both tasks (i.e. memory
retrieval and multiple object tracking tasks), it has difficulty explaining dACC activation
in response to negative feedback. There are no competing responses or items during
feedback processing. Previous research has suggested ACC involvement in negative
feedback processing (e.g., Miltner et al., 1997; Holroyd & Coles, 2002; Gehring et al.,
2002); however these studies employed ERPs, which are known to be imprecise in terms
of spatial resolution. Some conflict monitoring theory proponents (e.g. Carter & van
Veen, 2007) doubt that the ACC is the origin of feedback-related ERN, because the ACC
involvement in feedback processing was not confirmed by imaging studies (van Veen et
al., 2004; Nieuwenhuis et al., 2005).
In the fMRI studies (van Veen et al., 2004; Nieuwenhuis et al., 2005), participants
performed a time estimation task similar to the one used in the original ERP study by

 

47

Miltner and colleagues (1997), which was described in the introduction. Recall that in
this task, participants were asked to press a button one second after the onset of a trial,
and feedback indicated whether the response was within an acceptable range.
Nieuwenhuis and colleagues did not find a single region in the medial frontal cortex
responding more to negative feedback compared to positive feedback, and they
concluded that the ACC is not the neural generator of the feedback-related negativity. In
the present study, I did find that a region in the left dACC responds to negative feedback.
The discrepancies between the results of our study and the previous fMRI studies of
feedback-related responses in the ACC (e.g., Nieuwenhuis et al., 2005) likely involve
distinguishing between expected and unexpected negative feedback in the current study. I
found an ACC response only for the unexpected negative feedback after undetected
errors. When I collapsed all negative feedback, including detected errors, and compared it
to all positive feedback, I did not find a significant difference, p > 0.05, similar to
previous work. Therefore, expectations regarding the accuracy of the performance rather
than merely the negative valence of the feedback seem to be a key to the ACC
engagement. In addition to collapsing between expected and unexpected negative
feedback, Nieuwenhuis and colleagues employed a task that did not encourage the system
to form any kind of performance expectations. In their task, the range of acceptable
responses changed within a session to keep the participants’ accuracy at 50% level. As
participants were calibrating their 1 sec estimations throughout the session, the time
window of correct response narrowed. So what was considered to be a correct estimation
at the beginning of a session was penalized as an incorrect response at the end of the

 

48

session. This procedure could potentially discourage the system from forming any strong
expectations regarding the accuracy of the performance.

3.1.2 Evidence concerning the Reinforcement learning theory
The theory that cannot only accommodate ACC activation during feedback
processing but actually predicts the observed interaction between error detection and task
phase is the reinforcement learning theory (Holroyd & Coles, 2002). In this theory, the
ACC gets activated when the system detects that an outcome of an action is worse-thanpredicted. When the error detection fails because the system does not have a criterion for
what constitutes a correct action or due to fatigue or noise in the system, it waits until
some external feedback provides an evaluation of performance. However, when error
detection occurs because the system learned the criteria for correct actions, and the
system functions properly then it does not need to wait for external feedback to issue an
evaluative signal.
The results of the present imaging study map well onto the results of the ERP
study by Holroyd and Coles (2002), which was discussed in the introduction. Recall that,
in the ERP study, participants were asked to press a left or right button in response to a
stimulus. The stimulus-response mapping was not given to participants but had to be
learned throughout the experiment. One stimulus was mapped onto the left button,
another stimulus was mapped onto the right button. That was the 100% mapping
condition. Two stimuli were not mapped onto a particular button, and participants
received negative feedback on half of the trials when the stimuli appeared – 50%
mapping. The fifth stimulus was always penalized, and the sixth stimulus was always

 

49

rewarded. At the beginning of a session, all conditions elicited ERN at the point of
feedback. However, as learning occurred, the ERN moved from feedback to the point of
response but only for stimuli in the 100% mapping condition. Holroyd and Coles (2002)
concluded that when the system learns a representation of a correct response it does not
need to rely on external feedback to evaluate the performance. However, when such
representations are not available as in 50% mapping condition, the system waits until the
feedback.
Similarly, in the present study, I found that when the system is able to detect an
error, it does not wait until the feedback to issue an error signal or a signal indicating that
the performance outcome was worse-than-expected. However, when such detection fails
(in this case not because it was impossible as in Holroyd and Coles’ study, but due to
some other factors such as fatigue, noise, etc.), the system does not issue the error signal
until it processes external feedback. This is exactly the pattern of the results that would
have been predicted by the reinforcement learning theory.
3.1.3 Evidence concerning the Error-likelihood theory
Recently, Brown and Braver (2005) proposed yet another theory of ACC
function. In this view, the ACC represents a prediction of the likelihood of an error. I
describe this study here, because it seems to provide converging evidence for the idea
that expectations of a particular outcome influence ACC activation. First, I present the
hypothesis of the theory regarding ACC function and the main empirical results taken to
support it. Then I discuss a recent finding from this research group, which explored
individual differences in forming expectations of performance outcomes (Brown &
Braver, 2007). Finally, I try to reconcile and integrate the results of the present research,

 

50

the results of Brown and Braver studies (2005, 2007), and the results of Holroyd and
Coles ERP study (2002) that was described earlier.
Brown and Braver (2005) observed that the ACC responds more on a high-error
likelihood trial compared to a low-error likelihood trial even when there is no error or
conflict on that trial. This result was taken to show that the ACC learns and stores errorlikelihood information, rather than monitoring conflict. The authors employed a modified
go/no go task, called a change signal task. In this task, participants were presented with a
color cue at the beginning of a trial – blue corresponded to a high-error condition and
white corresponded to a low –error condition. Then an arrow appeared pointing to the left
or to the right. On 66% of the trials, the arrow indicated the required button-press – left or
right index finger. These were go trials. However, on the 33% of the trials, a second
arrow appeared signaling that the response had to be left-right reversed – if the first arrow
pointed to the left, the second arrow would point to the right and the response should be
made with the right index finger. These were change trials. The change trials were
equally frequent in blue and white conditions. What made the blue condition harder or
more error-prone was a longer interval between the first and the second arrow. The
longer interval made it more difficult to cancel the prepared response since the second
arrow appeared when the point of no return in response generation had already been
reached.
Brown and Braver found that, as expected, the ACC responds more on error trials
compared to correct trials – the error effect. They also found that the ACC responded
more to correct-change trials compared to correct-go trials – the conflict effect.
Importantly, they found that the ACC also responded more to high-error-likelihood

 

51

correct trials compared to low-error-likelihood correct trials. The effect was observed for
both change and go trials. The finding that the ACC responds to high-error-likelihood
trials even in the absence of any conflict (i.e., go trials) led to the conclusion that the
ACC represents the information about error-likelihood. The authors also found that the
pattern of ACC activation was reversed on error trials: the ACC responded greater to
low-error-likelihood trials compared to high-error-likelihood trials. Brown and Braver
claimed that the finding was consistent with their model, but did not provide detailed
explanation apart from saying that “when error likelihood is lower, actual errors more
highly suppress phasic dopamine cell activity.”
I believe that the results of Brown and Braver (2005) study show that the ACC
responds to violations of expectations rather than that the ACC responds to high-error
likelihood context. I believe that the cue at the beginning of a trial creates expectations
regarding performance on that trial: in a high-error likelihood condition, people expect to
make an error, and in low-error likelihood condition, people expect to perform correctly.
In Brown and Braver’s study (2005), the ACC responded more to correct trials in the
high-error likelihood condition than to correct trials in the low-error likelihood condition;
and it responded more to error trials in the low-error likelihood condition than to errors in
the high-error likelihood condition. These results are consistent with the idea that the
ACC is activated when expectations regarding the outcome of the performance are
violated regardless of the valence of the violation (i.e. better- or worse-than-expected).
In more recent studies, Brown and Braver (2007, 2008) investigated individual
differences in learning error-likelihood information. They discovered an inverse
relationship between ACC activation to error-likelihood (blue versus white cue

 

52

conditions) and ACC activation to conflict (change versus go trials). Risk-taking
individuals in their study lacked an error-likelihood effect, but at the same time, they
showed a greater conflict effect compared to controls. The authors concluded that risktaking individuals were acquiring the error likelihood information at a lower rate than the
control group. The results of the study are consistent with the idea that expectations of a
particular performance outcome influence ACC activation. The risk-taking individuals
did not learn to associate a cue with a particular error-likelihood condition; therefore,
they did not show any error-likelihood effect. At the same time, if the risk-taking
individuals expect positive outcomes on most of the trials then errors and conflicts would
always be unexpected, which would lead to greater ACC activation on these trials.
3.1.4 Integrating the finding from ERP and neuroimaging studies
The results of all three studies, Holroyd and Coles (2002), Brown and Braver
(2005), and the current study, are consistent with the idea that the ACC responds to
unexpected performance outcomes. Holroyd and Coles (2002) and the present study show
that the ACC responds to worse-than-expected outcomes. The Brown and Braver study
(2005) shows that the ACC responds not only to worse-than-expected, but also to betterthan-expected outcomes (although the authors did not interpret their results in this
manner).
Figure 14 summarizes the finding from these three studies. The information about
a negative outcome or the possibility of a negative outcome can become available at three
time points: during a cue/stimulus presentation (future), during task performance or
response (present), and during feedback (past). A red cross marks the point when the
information about a negative outcome or an outcome that was different-than-expected is

 

53

presented. Describing this figure, I refer to the red cross as a worse-than-expected
outcome or negative outcome for simplicity, but I believe that the same is true for betterthan-expected outcomes. The upper line in the Figure 14 represents a scenario in which
participants learn about negative outcome at the point of feedback. The middle line
represents a scenario in which participants learn about negative outcome during task
performance or as they issue a response. The lower line represents a scenario in which
participants learn that the performance is likely to be worse-than-expected during a cue or
a stimulus presentation. The red curves above each time point indicate the ACC response
to an event in each of the scenarios: high or low.

Figure 14. Summary of the results from three studies: Holroyd & Coles, 2002 (A),
present study (B), and Brown & Braver, 2005 (C).
Holroyd and Coles (2002) investigated all three scenarios, see Figure 14 the grey
box A. The upper line represents their 50% mapping condition, in which participant had
to wait until feedback to learn about the outcome of their actions. The middle line
 

54

represents a 100% mapping condition after learning has occurred – ERN was observed at
the response phase. The lower line represents their “always incorrect” condition, in which
stimulus itself was informative in terms of the outcome of the performance.
The yellow box C in Figure 14 shows the scenarios investigated in the Brown and
Braver (2005) study. The middle line represents errors after receiving a low-error
likelihood cue, and correct trials after receiving a high-error likelihood cue; in both cases,
outcomes of performance (positive and negative) are inconsistent with developed
expectations. The lower line represents errors after receiving a high-error likelihood cue
and correct trials after receiving a low-error likelihood cue; in both cases, the outcomes
are consistent with developed expectations.
The blue box B in Figure 14 represents the scenarios considered in the current
study. The upper line represents undetected errors (or inconsistently interpreted errors);
the system learns about the negative outcome of the performance only at feedback. The
middle line represents detected errors (or consistently interpreted errors); the system
registers that the performance is worse-than-expected during task performance.
The results of the three studies seem to be consistent with the idea that the ACC
responds to unexpected performance outcomes. The system can learn about the
outcome of the performance at any point in the trial: before the task, during the task, and
after the task. Importantly, the results depicted in Figure 14 are obtained with two
different methodologies – ERP and fMRI, which allows us to conclude that the signal
obtained during task performance and feedback processing has a common source – the
ACC. The question of whether the ACC responds during cue processing remains to be
addressed.

 

55

3.2 Content-specific signal in the ACC
The imaging results showed that an error related signal within the ACC is content
specific. Previous studies have demonstrated that the ACC represents information about
the occurrence of an error (e.g., Carter et al., 1998). The present results extend our
knowledge regarding the ACC involvement in error monitoring by showing that the ACC
represents information about the type of error that participants make, but not the type of
error they think they made (i.e., type of interpretation).
Comparing the trials on which participants failed to retrieve correct information
with the trials on which participants lost track of dots, I found an area within the left
dACC that is differentially activated to the memory retrieval and tracking failures. This
area was more active when participants lost track of a dot than when they failed to
retrieve a correct item from memory. This differential response was observed during task
performance but not during feedback processing. Comparing the trials on which
participants interpreted errors as tracking and the trials on which participants interpreted
errors as memory failures, I did not find an area within the left ACC that responds
differentially to the memory and tracking interpretations.
The differences in the activation observed for memory and tracking errors could
also be due to the difference in visual stimulation during memory and tracking phases. In
the analysis reported above, in section ROI analysis: dACC during Task phase, I
compared activation during memory and tracking errors without subtracting the
activation during correct trials. Thus, the observed differences in activation could
potentially be caused by differences in task demands. To test this possibility I compared

 

56

the activation during correct memory and correct tracking phases, see Appendix B. The
analysis revealed a marginally significant effect of Task phase – the tracking phase
seemed to result in greater ACC activation compared to the memory retrieval phase even
when no mistakes were made. These results suggest that differences in the task demands
(e.g., visual stimulation) could at least partially be responsible for differences in
activation between memory and tracking errors. To test whether there are also differences
in error-related activation above and beyond the differences on correct memory and
correct tracking phases, I would have to subtract correct trials from error trials, ((Terr_t +
Terr_m) – 2 x Tcor) – (Merr_m + Merr_t) – 2 x Mcor). Unfortunately, that would have
been a case of double dipping (e.g., Kriegeskorte, Simmons, Bellgowan, & Baker, 2010),
because I identified the region in the left dACC by comparing all errors versus correct
trials. One way of overcoming this problem would be to use tasks that are better matched
in terms of their demands on visual system.
Assuming that the left dACC does differentiate between types of error, I see
various explanations to this finding. One possibility is that the ACC contains neurons
tuned to a particular type of error, similar to the feature detectors in visual cortex that
respond to bars of light with a particular orientation. The neurons in the ACC could
hypothetically respond to different types of error. These neurons could have different
anatomy and physiology, or could differ in terms of their connections with the rest of the
brain. A single cell recording study would be needed to address the question regarding
the existence of various subsets of neurons tuned to different error types.
Another possibility is that people have differential sensitivity to the performance
of different subsystems in the mind. For example, they could be more likely to detect a

 

57

tracking error and miss a memory error. A lower detection rate of memory errors could
hypothetically lead to lower activation of the ACC region when we are looking at the
responses averaged across all error trials. Behavioral studies have the potential to address
this question. For example, participants could be presented with a task similar to one used
in the present study with one difference – they would not be given feedback. In the
present study, when participants received negative feedback, they could potentially
employ a strategy, such as “if I did not make a tracking error, but I get negative feedback
then it must be a memory error.” In this case, detection of a memory error during a
retrieval attempt is not necessary for correctly identifying a trial as a memory error trial.
Without feedback participants could no longer rely on strategies or heuristics to infer the
accuracy of their performance, but they would have to rely on an error signal generated
during task performance. This paradigm would allow the comparison of an error
detection rate for different types of errors.
Finally, the greater ACC response to tracking errors compared with memory
errors could be explained by the differences in expectations regarding the accuracy of the
performance on the two tasks. The memory task could be too difficult for participants to
form strong expectations of accurate performance, whereas the tracking task could be
perceived as more manageable. So when participants fail to retrieve a correct item their
expectations are not violated as much as when they lose a track of a dot. This hypothesis
could be tested by parametrically varying the strength of expectations of positive
outcomes. If a difficult version of a tracking task leads to lower ACC activation in
response to an error compared to an easy version of a tracking task then it would suggest
that ACC activation varies with the strength of expectations.

 

58

3.3 ACC activation in tasks without response conflict
Most of the research on error detection and conflict monitoring employs tasks that
involve response conflict such as the Eriksen task, the Stroop task, the go/no go task, etc.
In these tasks, errors or conflicts arise from stimuli activating several responses
simultaneously (i.e. response conflict). Research addressing the question of whether the
ACC responds to conflict at all levels of processing or only to conflict at the response
phase has produced mixed results. Some studies suggest that the ACC is activated only
during response conflict (e.g., Milham, Banich, Webb, Barad, Cohen, Wszalek, &
Kramer; 2001), whereas other studies suggest that the ACC responds to conflict at
semantic level as well (e.g., Barch et al., 2000). The findings of the present study support
the latter idea, that the ACC can be activated by errors that do not involve response
conflict. I find that errors in memory retrieval and in the multiple object tracking tasks
activate a region in the left dACC. There is no response conflict during tracking or
memory retrieval and yet ACC activation is greater when performance fails on these
tasks. In the context of the finding that the ACC responds to worse-than-expected
outcomes, the lack of ACC activation during non-response conflict trials (i.e. Milham et
al., 2001) could suggest that the system did not register the event as worse-than-expected,
or the non-response conflict violated the system’s expectations to a lesser degree than
response conflict.

 

59

4 FUTURE DIRECTIONS
4.1 Does the ACC respond to better-than-expected outcomes?
One of the most interesting findings of the present research is an interaction
between error detection and the phase of the trial. I found that the ACC responds to errors
when they are detected and to negative feedback when it is unexpected. I concluded that
the ACC responds to worse-than-expected outcomes. In a follow up study, I plan to
investigate whether the ACC responds to outcomes that are worse-than-expected or to
unexpected outcomes regardless of the valence of that outcome– negative or positive.
In this experiment, I will present participants with only a multiple-object tracking
task. After the tracking phase of the trial, I would present feedback on performance
negative or positive. After feedback, participants will be asked to make a judgment of
whether the feedback was congruent with their evaluations of the performance. There
will be four types of trials: (1) expected negative feedback – when participants correctly
detect a tracking failure; (2) unexpected negative feedback – when participants fail to
detect a tracking error; (3) expected positive feedback – when participants do not detect
an error and there is no error; and (4) unexpected positive feedback – when participants
believe they made an error, but they receive a positive feedback.
First, I expect to replicate the results of the current study and find that unexpected
negative feedback evokes greater ACC activity compared with expected negative
feedback. More importantly, I am interested in examining the contrast between expected
positive feedback and unexpected positive feedback. The finding that the ACC is more
active during the processing of unexpected positive feedback than during the processing
of expected positive feedback would suggest that the ACC responds to violations of

 

60

expectations of performance outcome regardless of the valence of the expectations –
worse- or better-than-expected.
4.2 Violations of expectations in the context of anxiety.
Detection of an error is accompanied with an emotional reaction. People vary in
terms of the intensity and duration of the emotion they experience in relations to an error
commission: anxiety prone people are known to overreact to an error and experience a
higher level of frustration compared to their non-anxious counterparts. Neuroimaging
studies show a heightened level of activation in the ACC and greater ERN in people
scoring high on anxiety questionnaires (e.g. Fitzgerald, Welsh, Gehring, Abelson, Himle,
Liberzon, & Taylor, 2005; Hajcak, McDonald, & Simons, 2003). In future studies, I plan
to investigate whether anxiety-prone people overreact to negative outcomes of their
actions or to unexpected outcomes regardless of the valence of the performance (betteror worse-than-predicted).
In this study, participants first will fill out the STAI-trait and state inventory,
Beck anxiety scale, and BIS/BAS questionnaires. The participants will be assigned to
high or low anxiety group based on the results of these questionnaires. Then each subject
will participate in one fMRI session. During this session, they will perform the task
described in the previous section, “Does the ACC respond to better-than-expected
outcomes?” If people from the high anxiety group show greater ACC activity only to
unexpected negative feedback then this finding would confirm the idea that cognitive
processing in the context of anxiety is biased toward negative outcomes. However, if
people from high anxiety group show greater ACC activity in both the unexpected
negative and unexpected positive feedback conditions, then it would suggest that the

 

61

system is not biased toward negative outcomes, but instead is more generally sensitive to
violations of expected outcomes. Also, it would be interesting to explore if there are
subgroups within the high anxiety group: those who show heightened activation to
unexpected negative feedback only, and those who show heightened activation to both
unexpected feedback – negative and positive. Thus, people who would commonly be
diagnosed with anxiety disorder could in principle suffer from failures in different parts
of the system, and may require differential treatments.
4.3 Negative emotions and rate of learning.
I would also like to explore the idea that emotions influence the rate with which
the system learns error likelihood information. Detection of an error is usually
accompanied with a negative emotional reaction. The functional significance of this
reaction is poorly understood. Recent findings (e.g., Fitzgerald, et al., 2005) show that
heightened emotional reactions to errors are associated with greater activities in the brain
areas implicated in error-detection, pain, and reward processing – ACC, insula, basal
ganglia. However, no clear behavioral consequences of this increased activity and
associated emotions have been established yet (e.g., Hajcak et al., 2003). In future work, I
plan to explore the effect of emotions on learning subsequent to errors, in particular,
learning of error-likelihood information.
From studies on decision making, we know that negative affect can be beneficial
in learning outcome probabilities (Mueller, Nguyen, Ray, & Borkovec, 2010). In a recent
study, Mueller and colleagues (2010) showed that people scoring high on anxiety
questionnaires perform better than those scoring low on anxiety questionnaires on the
Iowa gambling task. In this task, people are presented with four decks of cards. Drawing

 

62

from a deck of cards can result in a particular combination of gains and losses. The decks
differ in the net gains/losses. With time people learn to draw from advantageous decks
and avoid drawing from disadvantageous decks. People scoring high on anxiety were
shown to learn the contingencies faster than people scoring low on anxiety (Mueller et
al., 2010). This enhancement may be attributed to stronger emotional signals after errors
in anxious individuals.
In a recent study, Brown and Braver (2008) investigated individual differences in
learning error-likelihood information. They showed that the amplitude of ACC activity
depended on the rate of learning error likelihood information. The authors found that
risk-taking individuals did not show the error-likelihood effect (blue versus white cue) in
ACC activation when compared to controls, but they did show a greater conflict effect
(change versus go trials). Brown and Braver hypothesized that risk-taking individuals
were acquiring the error likelihood information at a lower rate than the control group.
This assumption was confirmed by a simulation, in which a computational model
behaved similarly to risk-taking individuals when the rate of learning was lowered
(Brown & Braver, 2008).
Therefore, in future work, I plan to address the question of whether high anxiety
is associated with faster learning rates and greater error-likelihood effects in the ACC. In
addition, I would also like to test whether an inverse relationship between errorlikelihood and conflict effects would be observed in high anxiety individuals, who are
traditionally characterized as risk-avoiding.

 

63

4.4 Lingering effect of emotions
Emotions are not just generated in response to a cognitive process (e.g., errordetection), but some components of emotions may last for some time after the cause is
gone. The question I plan to address in the future is whether negative emotions generated
on a current trial can spillover onto a subsequent trial and influence its evaluation.
To address this question, I plan to modify the change signal task (Brown and
Braver, 2005). In this modified version, there will be three conditions – one high-error
likelihood condition (blue cue) and two equally low-error likelihood conditions (yellow
and white cues). In the original task employed by Brown and Braver (2005), conditions
were equiprobable and randomly intermixed. In the modified version of the task, one
low-error likelihood condition (e.g., yellow cue) will follow high-error likelihood
condition more often than another low-error likelihood condition (e.g., white cue). Brain
activity will be recorded as participants perform the task and learn how error-prone each
condition is. At the end of the session, participants will be asked to rate the conditions in
terms of their difficulty/number of errors they produced.
As in the original study by Brown and Braver (2005), I expect to find greater
ACC activity in high-error likelihood condition compared to low-error likelihood
conditions (blue > yellow or white). Importantly, out of two equally low error likelihood
conditions, the one that follows high-error likelihood condition more often is expected to
be associated with greater ACC activity than the one that rarely follows the high-error
likelihood condition (yellow > white). These results would suggest that error-generated
emotions linger over time and influence the recording of error-likelihood information
associated with the subsequent performance. Participants’ rating at the end of a session

 

64

will suggest if subjects are aware of the error-likelihood information they learn (as
reflected in their brain activity). The lingering effect would inform models of errordetection. If found it would suggest that detection of an error/conflict per se is not
necessary for learning of error-likelihood information. The affective reaction generated in
response to an error is sufficient for such learning to occur.
Negative affect can potentially not only spillover to a subsequent action, but it
could also travel up a hierarchy of actions. The lingering effect of error-generated
emotions could be a mechanism by which the system reinforces an action at a more
abstract level. For example, spilling sugar on the floor, while trying to add it to a cup of
coffee, can potentially negatively reinforce not only “adding sugar” (an elementary
action) but the whole “coffee-making” procedure (an action at a higher level of
abstraction). Follow-up experiments addressing these questions would have important
implication for theories of action control in the normal population (e.g. Cooper &
Shallice, 2000; Botvinick & Plaut, 2004) and in patients suffering from action
disorganization syndrome and apraxia (e.g., Humphreys & Forde, 1999; Schwartz, Reed,
Montgomery, Palmer, & Mayer, 1991).
4.5 Lingering effect of emotions in a context of anxiety
In this study, I would like to explore the individual differences in the lingering
effect of error-generated emotions. In particular, I would like to address the question of
whether people scoring high on negative affect (e.g. anxiety) would produce a greater
lingering effect or a more lasting effect compared to people scoring low on anxiety.
First, participants will fill out the STAI-trait and state inventory, Beck anxiety
scale, and BIS/BAS questionnaires. They will be assigned to high or low anxiety group

 

65

based on the results of these questionnaires. Then each subject will participate in one
fMRI session. During this session, they will perform the modified version of change
detection task described in the previous session “Lingering effect of emotions”. Anxiety
prone people tend to overreact to errors and experience more intense and prolong
negative emotions. Findings that the lingering effect can be modulated by the level of
anxiety (train or induced) would confirm the assumption that error-generated emotions
are the primary source of the learning error-likelihood information.

 

66

5 CONCLUSIONS
The results of the present work challenge the idea that the dACC performs
conflict monitoring, and they are more consistent with the idea that the dACC responds to
unexpected performance outcomes. In an fMRI study, I found that the same region in the
left dACC responds both to detected errors and to unexpected negative feedback.
Although the conflict monitoring theory can accommodate the finding that errors in a
memory retrieval task and in a tracking task evoke ACC activity, it is unclear how the
theory would explain the finding that the ACC responds to unexpected negative feedback
since there is no conflict during feedback processing. The results of the present study are
consistent with the reinforcement learning account, since both detection of errors and
processing of unexpected negative feedback are worse-than-expected outcomes. Future
studies are needed to address the question of whether the ACC responds to worse-thanexpected outcomes, or whether it responds to unexpected performance outcomes
regardless of valence – positive or negative.
The imaging results also showed that an error related signal within the left dACC
is content specific. Tracking errors led to greater ACC activity compared to retrieval
errors. Several explanations are possible and future work is needed to distinguish
between them. One possibility is that the ACC contain neurons that are tuned to a
particular type of error. Alternatively, participants might have differential expectations
regarding the accuracy of their performance on two tasks. The errors on the task that has
greater expectations associated with it result in greater ACC activation.

 

67

APPENDIX A
Analysis of regions other than the left dACC
Although the contrast “all errors versus all correct trials” revealed only one
region, the left dACC, to be active more for error trials, other important contrasts showed
multiple brain regions to be involved in processing of errors and unexpected feedback.
Here I address the question whether these areas perform a function similar to the left
dACC. In particular, whether they exhibit the same pattern of activation as the left dACC
does – responding to consistently interpreted errors during the Task phase and to
inconsistently interpreted errors during the Feedback phase.
Consistently interpreted errors versus correct trials
Contrasting consistently interpreted errors and correct trials (Terr_t + Merr_m) –
(Tcor +Mcor), I found a greater activation for consistently interpreted errors (Terr_t +
Merr_m) in the following regions: the left dACC (two adjacent regions: a small inferior
region and a large superior region), the left medial frontal gyrus (MFG), and the left
superior frontal gyrus (SFG), see Figure 9. Inconsistently interpreted errors do not show
activity in these regions when compared to correct trials in the whole brain analysis.
In the ROI analysis, I explored activation in these regions during feedback
processing. I was interested to see if the regions would show a pattern of activation
similar to what I found for the left dACC in “all errors versus all correct trials” contrast,
that is if they will be more active on inconsistently identified errors compared to
consistently identified errors during feedback processing. I expected to see a large
superior region in the left dACC showing this pattern, as it encompasses the left dACC
identified in “all errors vs. correct trials” contrast. I did not have predictions for other

 

68

regions as they might perform a function different from the left dACC and show
consistency effect only during error detection but not during negative feedback
processing.
ROI analysis: a large region in the left dACC.
Figure A1 shows percentage signal change in the large superior dACC region as a
function of consistency of interpretation and error type. A repeated-measures ANOVA
with Consistency (consistent versus inconsistent) and Error type (tracking versus
memory) factors revealed a marginally significance main effect of consistency, F (1, 16)
= 3.317, p = 0.087. The main effect of Error type and the interaction were not significant,
p > 0.7. To explore the marginally significant effect of Consistency further, pair-wise ttests were conducted on tracking and memory errors separately. The effect of
Consistency was not significant for memory errors, p = 0.398. The effect of Consistency
was small but significant for tracking errors in a one-tailed t-test, t(16) = 1.996, p =
0.032. After excluding two outliers (2 SDs from the mean in FTer_m condition), I find
the effect of Consistency to be significant even in a two-tailed t-test, t(14) = 2.592, p =
0.021. Thus, the results show that the activation pattern in the large region of the left
dACC in “consistently interpreted errors vs. correct trials” is similar to the pattern of
activation observed in the left dACC in “all errors vs. correct trials”.

 

69

0.8
Consistent
Inconsistent

0.7

% BOLD signal change

0.6
0.5
0.4
0.3
0.2
0.1
0
-0.1
Tracking

Memory

-0.2
-0.3
Error Type

Figure A1. ROI analysis: a large region of the left dACC. During negative feedback
processing, the region shows Consistency effect for tracking errors, but not for memory
errors. The error bars show SEM.
ROI analysis: a small region in the left dACC, the left SFG, and the left MFG.
In the ROI analysis, I explored activation in a small region of the dACC, the left
SFG, and the left MFG during feedback processing, see Table A1. Repeated-measures
ANOVAs with Consistency (consistent versus inconsistent) and Error Type (tracking vs.
memory) as within subject factors showed neither of the main effects nor the interaction
of the factors was significant, p > 0.05. In addition, pair-wise t-tests on tracking and
memory errors did not show consistency effects either, p > 0.05. Thus, the small region in
the left dACC, the left SFG, and the left MFG did not exhibit the pattern of activation
observed in the left dACC in the “all errors vs. correct trials” contrast. Inconsistently

 

70

interpreted errors did not lead to greater activation compared to consistently interpreted
errors.
Table A1. ROI analysis: Percentage BOLD signal change observed during feedback
processing for consistently and inconsistently interpreted tracking and memory errors in a
small region in the dACC (a) and in the left SFG (b). Repeated measures ANOVAs
showed neither the main effects (Consistency or Error Type) nor the interaction between
factors.

Tracking Errors

Memory Errors

Consistent
small L dACC
L SFG
L MFG

Inconsistent

Consistent

Inconsistent

Mean (SD)

Mean (SD)

Mean (SD)

Mean (SD)

-0.19 (0.38)

0.07 (1.11)

-0.10 (0.74)

-0.43 (0.94)

-0.02 (0.56)

-0.18 (0.72)

-0.18 (0.83)

-0.47 (1.40)

0.04 (0.51)

0.22 (0.90)

-0.15 (0.58)

0.23 (1.91)

To summarize, contrasting consistently interpreted errors with correct trials, I
identified a number of regions more active for consistently interpreted errors. They are
two regions in the left dACC (big and small), a region in the left SFG, and a region in the
left MFG. ROI analysis in these areas showed that only the big region in the left dACC
exhibited the pattern of activation similar to the pattern found in the left dACC from the
“all errors versus all correct trials” contrast. The other regions – a small inferior region in
the left dACC, the left SFG, and the left MFG – although active for consistently
interpreted errors during task performance, do not show a consistency effect during
negative feedback processing. The results suggest that these regions perform a function
 

71

different from the left dACC. Although they are involved in error processing, they do not
seem to respond to worse-than-expected outcomes as the left dACC does.
Consistently interpreted tracking errors versus correct tracking trials
Contrasting consistently interpreted tracking errors and correct tracking trials
(Terr_t – Tcor), I found a greater activation for consistently interpreted tracking errors
(Terr_t) in the following regions: the dACC bilaterally, the left medial frontal gyrus
(MFG), the left superior frontal gyrus (SFG), and the inferior parietal gyri bilaterally, see
Figure A2. Inconsistently interpreted tracking errors (Terr_m) do not show activity in
these regions when compared with correct tracking trials.

Figure A2. Consistently interpreted tracking errors. Regions identified in a whole brain
analysis contrasting consistently interpreted tracking errors and correct trials (Terr_t –
Tcor). The arrows point to the regions more active for consistently interpreted tracking
errors. They are the dACC bilaterally, the left medial frontal gyrus (MFG), the left
superior frontal gyrus (SFG), and the inferior parietal gyri (IPG) bilaterally.
 

72

ROI analysis.
In the ROI analysis, I explored activation in these regions during feedback
processing. In particular, I was interested to see if these regions would show the same
pattern of activation as the left dACC identified in all errors versus correct trials contrast
– greater activation for inconsistently interpreted errors during feedback processing.
Contrasting the activation on consistently interpreted trials with activation on
inconsistently interpreted trials (FTer_t – FTer_m), I did not find a consistency effect for
neither of the regions, p > 0.05, see Figure A3 and Table A2. This result indicates that the
regions shown in Figure A2 perform a function different from left dACC identified in all
errors versus all correct contrast. The regions also do not show greater activation for
errors compared with correct trials during feedback (i.e., negative feedback versus
positive feedback) p > 0.05. This pattern of results suggests that these areas participate in
error detection, but not in evaluation of the performance outcome. They can potentially
carry the information about occurrence of an error to other areas such as dACC or basal
ganglia, which later issue an evaluative signal suggesting that performance is worse than
expected.

 

73

Figure A3. ROI analysis: consistently interpreted tracking errors versus inconsistently
interpreted tracking errors during feedback processing in brain regions identified with the
contrast (Terr_t – Tcor). 1 = RIP, 2 = LIP, 3 = LMF, 4 = RACC, 5 = LACC. Neither of
the comparisons reaches significant level. Thus the bilateral IP, the left MFG, and the
bilateral ACC do not show a consistency effect during feedback processing. (The region
in the left ACC described here is more anterior and inferior compared to the regions in
the left ACC identified in “all errors vs. correct trials” contrast and in “consistent
interpretations vs. correct trials” contrast.)
Table A2. Group analysis for the contrast: “Consistently interpreted tracking errors >
correct tracking” during the Task phase.

Cluster

Active
volume

Max t
values

mm3

Max t value
locations
RL, AP, IS

Terr_t

Tcor

R IPG

788

7.82

-30, 44, 47

1.01(0.20)

0.79(0.21)

L IPG

487

7.625

38, 30, 41

0.77(0.13)

0.57(0.12)

SFG

465

7.392

7, -14, 50

0.45(0.18)

0.18(0.14)

L & R dACC

345

6.605

9, -29, 29

0.37(0.14)

0.08(0.10)

 

74

L MFG

248

6.446

38, -30, 25

0.33(0.06)

0.05(0.05)

Inconsistently interpreted errors versus consistently interpreted errors during
Feedback
Contrasting inconsistently interpreted errors with consistently interpreted errors
(FTer_m + FMer_t) – (FTer_t + FMer_m), I found a single region in the left IPG to be
active more for inconsistently interpreted errors, see Figure A4 and Table A3. ROI
analysis did not show a consistency effect in this region during task performance. A
repeated-measures ANOVA with Consistency (consistent vs. inconsistent) and Error type
(tracking vs. memory) showed neither the main effects nor the interaction between
factors, p > 0.7. Pair-wise t-tests did not show consistency effects for tracking errors (for
consistently interpreted errors, M = -0.15, SD = 0.34; for inconsistently interpreted errors,
M = -0.13, SD = 0.83). Pair-wise t-test did not show consistency effect for memory errors
also (for consistently interpreted errors, M = -0.15, SD = 0.34, for inconsistently
interpreted errors, M = -0.11, SD = 0.39). Thus, this region seems to respond to
unexpected negative feedback, but not to detection of an error during task performance.
This pattern of results suggests that this region performs a function different from the left
dACC identified in “all error versus all correct trials” contrast.

 

75

Figure A4. Feedback: Inconsistently interpreted errors. Contrasting inconsistently
interpreted errors and consistently interpreted errors during Feedback (FTer_m + FMer_t)
– (FTer_t + FMer_m), I find a single area in the left IPG responding greater to
inconsistently interpreted errors.
 
Table A3. Group analysis results for the contrast: “Inconsistently interpreted errors >
Consistently interpreted errors” during the Feedback phase.

Cluster

Active
volume

Max t
values

mm3
L IPG 

186

Max t value
locations
RL, AP, IS

5.914 

FMer_t

FMer_m

FTer_t

FTer_m

52, 47, 25

0.65(.09)

0.05(.14)

0.07(.07)

0.45(.20)

*

*

*

Exploring the contrasts “consistently interpreted errors versus all correct trials”
and “consistently interpreted tracking errors versus correct tracking trials”, I found
several brain areas, other than the ACC, being active for consistently interpreted errors
during the Task phase. They were the left SFG, left MFG, right IPG and left IPG. The
ROI analysis of these areas revealed that these areas do not show the same pattern of
activation as the left dACC does. Although, during the Task phase, these areas were
active selectively for consistently interpreted errors, during the Feedback phase, the effect
 

76

of the consistency of interpretation was not significant. At the same time, a brain area,
identified in the contrast “inconsistently interpreted errors versus consistently interpreted
errors during feedback processing”, the left IPG, exhibited an opposite pattern. It showed
a consistency effect during the Feedback phase but not during the Task phase.
The presence of the areas that respond to detected errors but not to unexpected
negative feedback, and the areas that respond to unexpected negative feedback but not to
detected errors suggests that error detection and processing of unexpected negative
feedback are different phenomena. However, these phenomena might have common
processes, for example, evaluation of performance outcome. Left dACC might be
involved in this process, since this area responds to both error detection and unexpected
negative feedback.
The present study does not allow making conclusions regarding the functions of
the other areas involved in processing of errors and negative feedback. Although from
previous studies we know that the left MFG, for example, is involved in working
memory, especially it is important for storage of spatial information (e.g. Leung &
Goldman-Rakic, 2002). The fact that we find it to be active for detected errors and not
during feedback processing is consistent with the idea that this region supports memory
processes. Possibly, after detecting an error the system needs to retain the information
about the occurrence of the error, so later it can make a correct judgment about the type
of the error. In the present study, we find that the left and right IPG are active during
detected errors compared to correct trials during the Task phase. In the previous
literature, these areas were also found to be active during correct tracking (e.g. Culham,
Brandt, Cavanagh, Kanwisher, Dale, Roger, & Tootell, 1998). Hypothetically, the areas,

 

77

performing tracking function, work harder when tracking becomes difficult and errors are
likely to occur. Thus the additional recruitment of these areas could serve as a signal to
high-level areas that tracking is likely to go wrong. These conclusions are merely
suggestive since they are based on reverse inferences. Further research is required to
determine the role of these areas in processing errors and negative feedback.

 

78

APPENDIX B
Correct Memory phase versus Correct Tracking phase
The present study shows that the left dACC responds greater to tracking errors
compared to memory errors. One of the possible explanations is that memory and
tracking phases impose differential demands on visual system. In this case, we should
observe a differential ACC activation to tracking and memory phases even in the absence
of an error. A pair-wise t-test, comparing Tcor with Mcor (see Figure A5), showed a
marginally significant effect of the Trial phase, t(16) = 1.935, p = 0.071. Thus tracking
seems to engage the left dACC region greater than memory retrieval even in the absence
of an error.
0.7
Tracking correct
Memory correct

% BOLD signal change

0.6
0.5
0.4
0.3
0.2
0.1
0
-0.1
-0.2
Trial type

Figure A5. ROI analysis: the left dACC. The region shows an effect of the Trial Phase:
correct tracking phase is associated with a greater dACC activation compared to correct
memory retrieval phase.

 

79

REFERENCES

Barch, D. M., Braver, T. S., Sabb, F. W., & Noll, D. C. (2000). Anterior cingulated and
the monitoring of response conflict: Evidence from and fMRI study of overt verb
generation. Journal of cognitive Neuroscience, 12, 298-309.
Botvinick, M., Braver, T. S., Barch, D. M., Carter, C. S., & Cohen, J. D. (2001). Conflict
monitoring and cognitive control. Psychological Review, 108, 624-652.
Botvinick, M., Cohen, J. D. & Carter, C. S. (2004). Conflict monitoring and anterior
cingulate cortex: An update. Trends in Cognitive Sciences, 8, 539-546.
Botvinick, M., Nystrom, L. E., Fissell K., Carter, C. S., & Cohen, J. D. (1999). Conflict
monitoring versus selection-for-action in anterior cingulate cortex. Nature, 402,
179-181.
Botvinick, M., & Plaut, D. C. (2004). Doing without schema hierarchies: A recurrent
connectionist approach to normal and impaired routine sequential action.
Psychological Review, 111, 395–429.
Brown, J. W. (2009). Conflict effects without conflict in anterior cingulated cortex:
Multiple response effects and context specific representation. NeuroImage, 47,
334-341.
Brown, J. W., & Braver, T. S. (2005). Learned prediciotns of error likelihood in the
anterior cingulated cortex. Science, 307, 1118-1121.
Brown, J. W., & Braver, T. S. (2007). Risk prediction and aversion by anterior cingulated
cortex. Cognitive, Affective, & Behavioral Neuorscience, 7, 266-277.
Brown, J. W., & Braver, T. S. (2008). A computational model of risk, conflict, and
individual difference effects in the anterior cingulate cortex. Brain Research,
1202, 99-108.
Buckner, R.L., Koutstaal, W., Schacter, D. L., Wagner, A. D., & Rosen, B. R. (1998).
Functional-anatomical study of episodic retrieval suing fMRI. I. Retrieval effort
versus retrieval success. Neuroimage, 7, 151-162.
Cabeza, R., Dolcos, F., Graham, R., & Nyberg, L. (2002). Similarities and differences in
the neural correlates of episodic memory retrieval and working memory.
Neuroimage, 16, 317-330.
Carter, C. S, Braver, T. S., Barch, D. M., Botvinick, M. M., Noll, D., & Cohen, J. D.
(1998). Anterior cingulate cortex, error detection, and the online monitoring of
performance. Science, 280, 747-749.

 

80

Carter, C. S., van Veen, V. (2007). Anterior cingulated cortex and conflict detection: An
update of theory and data. Cognitive, Affective & Behavioral Neurocscience, 7,
367-379.
Cooper, R., & Shallice, T. (2000). Contention scheduling and the control of routine
activities. Cognitive Neuropsychology, 17, 297–338.
Corbetta, M., Miezin, F. M., Dobmeyer, S., Shulman, G. L. & Petersen, S. E. (1990).
Attentional modulation of neural processing of shape, color, and velocity in
humans. Science, 248, 1556-1559.
Courtney, S. M., Ungerleider, L. G., Keil, K., & Haxby, J. V. (1997) Transient and
sustained activity in a distributed neural system for human working memory.
Nature, 386, 608-611.
Cox, R. W. (1996). AFNI: software for analysis and visualization of functional magnetic
resonance images. Computers and Biomedical Research, 29, 162-173.
Culham, Brandt, Cavanagh, Kanwisher, Dale, Roger, & Tootell, (1998). Cortical fMRI
activation produced by attentive tracking of moving targets. Journal of
Neurophysiology, 80, 2657-2670.
Culham, Cavanagh, & Kanwisher, (2001). Attention response functions: characterizing
brain areas with fmri activation during parametric variations of attentional load.
Neuron, 32, 737-745.
Donaldson, D. I., Wheeler, M. E., & Petersen, S. E. (2009). Remember the source:
Dissociating frontal and parietal contributions to episodic memory. Journal of
Cognitive Neuroscience, 22, 377-391.
Eriksen, C. W., & Schultz, D. W. (1979). Information processing in visual search: A
continuous flow conception and experimental results. Perception &
Psychophysics 25, 249–263.
Falkenstein, M., Hohnsbein, J., Hoorman, J., & Blanke, L. (1991). Effects of crossmodal
divided attention on late ERP components. II. Error processing in choice reaction
tasks. Electroencephalography and Clinical Neurophysiology 78, 447–455.
Fitzgerald, K. D., Welsh, R. C., Gehring, W. J., Abelson, J. L., Himle, J. A., Liberzon, I.,
& Taylor, S. F. (2005). Error-related hyperactivity of the anterior cingulated
cortex in obsessive-compulsive disorder. Biological Psychiatry, 57, 287-294.
Gehring, W. J., Goss, B., Coles, M. G. H., Meyer, D. E., Donchin, E. (1993). A neural
system for error detection and compensation. Psychological Science, 4, 385-390.
Hajcak, G., McDonald, N., & Simons, R. F. (2003). Anxiety and error-related brain
activity. Biological Psychology, 64, 77-90.

 

81

Holroyd, C. B., & Coles, M. G. H. (2002). The neural basis of human error processing:
Reinforcement learning, dopamine, and the error-related negativity. Psychological
Review, 109, 679-709.
Howe, P. D., Horowitz, T. S., Morocz, I. A., Wolfe, J., & Livingstone, M. S. (2009).
Using fMRI to distinguish components of the nultiple object trakcing task.
Journal of Vision, 9, 1-11.
Huk, A.C., Dougherty, R.F., & Heeger, D.J. (2002). Retinotopy and functional
subdivision of human areas MT and MST. Journal of Neuroscience, 22, 71957205.
Humphreys, G. W., & Forde, E. M. E. (1999). Disordered action schema and action
disorganization syndrome. Cognitive Neuropsychology, 15, 771–811.
Jonides, J., Lewis, R.L., Nee, D.E., Lustig, C.A., Berman, M.G., & Moore K.S. (2008).
The mind and brain of short term memory. Annual Review of Psychology, 59,
193-224.
Jovicich, J., Peters, R. J., Koch, C. Braun, J., Chang, L., & Ernst, T. (2001). Brain areas
specific for attentional load in a motion-tracking task. Journal of Cognitive
Neuroscience, 13, 1048-58.
Kerns, J. G., Cohen, J. D., MacDonald, A. W., Cho, R. Y., Stenger, V. A., & Carter, C. S.
(2004). Anterior cingulated conflict monitoring and adjustments in control.
Science, 303, 1023-1026.
Kim, D. H., Adalsteinsson, E., Glover, G. H., Spielman, D. M. (2002). Regularized
higher-order in vivo shimming. Magnetic Resonance in Medicine, 48, 715-722.
Konishi, S., Nakajima, K., Uchida, I., Kameyama, M., Nakahara, K., Sekihara, K., &
Miyashita, Y. (1998). Transient activation of inferior prefrontal cortex during
cognitive set shifting. Nature Neuroscience, 1, 80–84.
Milham, M. P., Banich, M. T., Webb, A., Barad, V., Cohen, N. J., Wszalek, T., &
Kramer, A. F. (2001). The relative involvement of anterior cingulated and
prefrontal cortex in attentional control depends on nature of conflict. Cognitive
Brain Research, 12, 467-473.
Miltner, W. H. R., Braun, C. H.,Coles, M. G. H. (1997). Event-related brain potentials
following incorrect feedback in a time-estimation task: Evidence for a “generic”
neural system for error detection. Journal of Cognitive Neuroscience, 9, 788-798.
Mueller, E. M.,Nguyen, J., Ray, W. J., & Borkovec, T. D. (2010). Future-oriented
decision-making in generalized anxiety disorder is evident across different
versions of the Iowa Gambling task. Journal of Behavior Therapy and
Experimental psychiatry, 41, 165-171.

 

82

Nee, D.E., & Jonides, J. (2008). Neural correlates of access to short-term memory. PNAS,
15, 14228-14233.
Norman, D. A. (1981). Categorization of action slips. Psychological Review, 88, 1–14.
Nieuwenhuis, S., Slagter, H. A., von Geusau, N. J. A., Heslendfeld, D. J., Holroyd, C. B.
(2005). Knowing good from bad: Differential activation of human cortical areas
by positive and negative outcomes. European Journal of Neuroscience, 21, 31613168.
Perrish, T. B., Gitelman, D. R., LaBar, K. S., Mesulam, M. M. (2000). Impact of signalto-noise on functional MRI. Magnetic Resonance in Medicine, 44, 925-932.
Pylyshyn Z. W., & Storm, R. W. (1988). Tracking multiple independent targets: evidence
for a parallel tracking mechanism. Spatial Vision, 3, 179-197.
Runganath, C., Johnson, M. K., & D’Esposito, M. (2003). Prefrontal activity associated
with working memory and episodic long-term memory. Neuropsychologia, 41,
378-389.
Reason, J. (1990). Human error. Cambridge, England: Cambridge University Press.
Rizzolatti, G., Riggio L., & Sheliga B.M. (1994). Space and selective attention. In
Umiltà C, Moscovitch M.(eds) Attention and performance, XV Bradford Book
Cambridge MA, 231-265.
Schwartz, M. F., Reed, E. S., Montgomery, M. W., Palmer, C., & Mayer, N. H. (1991).
The quantitative description of action disorganization after brain damage: A case
study. Cognitive Neuropsychology, 8, 381–414.
Talairach, J., & Tournoux, P. (1988). Co-Planar Stereotaxic Atlas of the Human Brain: 3Dimensional proportional system: an approach to cerebral imaging. New York,
New York: Georg Thieme Verlag.
Talmi, D., Grady, C. G., Goshen-Gottstein, Y., & Moscovitch, M. (2005). Neuroimaging
the serial position curve: A test of single-store versus dual-store models.
Psychological Science, 16, 716-723.
Van Veen, V., Holroyd, C. B., cohen, J. D., Stenger, V. A., & Carter, C. S. (2004). Errors
without conflict: Implications for performance monitoring theories of anterior
cingulated cortex. Brain & Cognition, 56, 267-276.
Ward, B. (2000). Deconvolution analysis of fMRI time series data. Biophysics Research
Institute, Medical College of Wisconsin, Milwaukee, WI.
Zeki, S. (1993). A Vision of the Brain. Oxford, England: Wiley-Blackwell.

 

83