University of Pennsylvania

ScholarlyCommons
Publicly Accessible Penn Dissertations

1-1-2015

Neural Mechanisms of Episodic Memory

formation
Nicole Long
University of Pennsylvania, nmarie.long@gmail.com

Follow this and additional works at: http://repository.upenn.edu/edissertations

Part of the Neuroscience and Neurobiology Commons, and the Psychology Commons

Recommended Citation
Long, Nicole, "Neural Mechanisms of Episodic Memory formation" (2015). Publicly Accessible Penn Dissertations. 1860.
http://repository.upenn.edu/edissertations/1860

This paper is posted at ScholarlyCommons. http://repository.upenn.edu/edissertations/1860

For more information, please contact libraryrepository@pobox.upenn.edu.
Neural Mechanisms of Episodic Memory formation
Abstract
In order to remember what you had for breakfast today, you must rely on episodic memory, the memory for
personal events situated within a spatiotemporal context. In this dissertation, I use electroencephalographic
(EEG) recordings to measure the neural correlates of successful episodic memory formation. The recorded
EEG signals simultaneously sample local field potentials throughout the brain, and can be analyzed in terms of
specific time-varying oscillatory or spectral components of neural activity which are thought to reflect the
concerted activity of neuronal populations. I collected EEG recordings while participants engage in free recall,
an episodic memory task during which participants must study and then recall a list of items. In the first
chapter, I compare the spectral correlates during encoding of items later remembered to those later forgotten
using two separate recording modalities, scalp and intracranial EEG. I find that memory formation is
characterized by broad low frequency spectral power decreases and high frequency power increases across
both datasets, suggesting that scalp EEG can resolve high frequency activity (HFA) and that low frequency
decreases in intracranial EEG are unlikely due to pathology. In the next chapter, I connect these HFA increases
to memory-specific processes by comparing study items based on how they are re- called, not whether they
are recalled. I find increased HFA in left lateral cortex and hippocampus during the encoding of subsequently
clustered items, those items recalled consecutively with their study neighbors at test. The precise time course
of these results suggests that context updating mechanisms and item-to-context associative mechanisms
support successful memory formation. In the third chapter, I measure how the formation of these episodic
associations is modulated by pre-existing semantic associations by including a semantic orienting task during
the encoding interval. I find that semantic processing interferes with the formation of new, episodic memories.
In the final chapter, I show that the memory benefit for emotionally valenced items is better explained by a
contextual mechanism than an attentional mechanism. Together, my work supports the theory that contextual
encoding associative mechanisms, reflected by HFA increases in the memory network, support memory
formation.

Degree Type
Dissertation

Degree Name
Doctor of Philosophy (PhD)

Graduate Group
Psychology

First Advisor
Michael J. Kahana

Keywords
Electroencephalography, Episodic, Free recall, Memory

This dissertation is available at ScholarlyCommons: http://repository.upenn.edu/edissertations/1860

Subject Categories
Neuroscience and Neurobiology | Psychology

This dissertation is available at ScholarlyCommons: http://repository.upenn.edu/edissertations/1860

NEURAL MECHANISMS OF EPISODIC MEMORY FORMATION

COPYRIGHT

2015

Nicole Marie Long

To Jonathan.

iii
Acknowledgments

First, I would like to thank my advisor, Mike Kahana. It has been through Mike’s
guidance and support that I have become the researcher that I am today. Mike
has taught me both how to critically evaluate the details of my work and how
to take a step back and view my work in the context of the larger literature. I
would also like to thank my previous mentors, Lila Davachi and David Badre, who
gave me support and encouragement when I was just learning about memory and
neuroscience research.
I am extremely grateful to all of the current and former Kahana lab members for
their ideas, support, help with data collection, and thoughtful discussions. I would
particularly like to thank Youssef Ezzyat, Karl Healey, Lynn Lohnas, Ashwin Ra-
mayya, John Burke, Jeremy Manning, Patrick Crutchley, and Elizabeth Crutchley.
I am likewise grateful to those members of the Davachi and Badre labs from whom
I first learned the ropes of memory research, in particular, Bernard Staresina, Ilke
Öztekin, and Sophie Lebrecht.
I would also like to thank Sharon Thompson-Schill and Joe Kable for their
insightful and valuable feedback as members of my thesis committee.
I am indebted to the patients who donated their time and cooperation, as well
as the neurologists, neurosurgeons, technicians, and funding agencies who made
this work possible.

iv
 Finally, I would like to thank my husband, and Kahana-lab alumnus, Jonathan
Miller, for his love and patience as he tirelessly read draft after draft of all my
papers as well as this dissertation. I would also like to thank my parents who have
become neuroscience aficionados as a result of listening to me discuss my work,
and have always provided me with support and encouragement.

v
 ABSTRACT

NEURAL MECHANISMS OF EPISODIC MEMORY FORMATION

Nicole Marie Long

Michael J. Kahana

In order to remember what you had for breakfast today, you must rely on
episodic memory, the memory for personal events situated within a spatiotempo-
ral context. In this dissertation, I use electroencephalographic (EEG) recordings
to measure the neural correlates of successful episodic memory formation. The
recorded EEG signals simultaneously sample local field potentials throughout the
brain, and can be analyzed in terms of specific time-varying oscillatory or spectral
components of neural activity which are thought to reflect the concerted activity of
neuronal populations. I collected EEG recordings while participants engage in free
recall, an episodic memory task during which participants must study and then
recall a list of items. In the first chapter, I compare the spectral correlates during
encoding of items later remembered to those later forgotten using two separate
recording modalities, scalp and intracranial EEG. I find that memory formation is
characterized by broad low frequency spectral power decreases and high frequency
power increases across both datasets, suggesting that scalp EEG can resolve high
frequency activity (HFA) and that low frequency decreases in intracranial EEG are
unlikely due to pathology. In the next chapter, I connect these HFA increases to
memory-specific processes by comparing study items based on how they are re-
called, not whether they are recalled. I find increased HFA in left lateral cortex and
hippocampus during the encoding of subsequently clustered items, those items
recalled consecutively with their study neighbors at test. The precise time course
of these results suggests that context updating mechanisms and item-to-context

vi
associative mechanisms support successful memory formation. In the third chap-
ter, I measure how the formation of these episodic associations is modulated by
pre-existing semantic associations by including a semantic orienting task during
the encoding interval. I find that semantic processing interferes with the formation
of new, episodic memories. In the final chapter, I show that the memory benefit
for emotionally valenced items is better explained by a contextual mechanism than
an attentional mechanism. Together, my work supports the theory that contex-
tual encoding associative mechanisms, reflected by HFA increases in the memory
network, support memory formation.

vii
Contents

Acknowledgments iv

Abstract vi

Contents viii

List of tables x

List of figures xi

1 Introduction 1
1.1 Spectral signals of episodic memory encoding . . . . . . . . . . . . . 2
1.2 Retrieved context theory . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Semantic memory in the context of episodic memory . . . . . . . . . 8
1.4 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Subsequent memory e↵ect in intracranial and scalp EEG 10

2.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 iEEG Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4 Scalp Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

viii
 2.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3 Contextual encoding mechanisms in hippocampus and prefrontal cortex

support successful memory formation. 33
3.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4 Modulation of task demands suggests that semantic processing interferes

with the formation of episodic associations 48
4.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.5 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

5 Recall dynamics reveal the retrieval of emotional context 65

5.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.3 Norming study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.4 Free recall study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.5 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.6 Supplemental . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

6 General discussion 82
6.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

ix
 6.2 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.3 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

References 89

x
List of tables

2.1 Subsequent memory e↵ect in iEEG and scalp EEG studies . . . . . . 22

xi
List of figures

2.1 Maps of a priori selected regions of interest. . . . . . . . . . . . . . . . 19

2.2 Frequency analysis of the subsequent memory e↵ect . . . . . . . . . . 25
2.3 Time-frequency analysis of the subsequent memory e↵ect . . . . . . . 32

3.1 Behavioral measures of recall performance . . . . . . . . . . . . . . . 41

3.2 Subsequent memory and clustering e↵ects . . . . . . . . . . . . . . . 41
3.3 Clustering e↵ects across time . . . . . . . . . . . . . . . . . . . . . . . 43
3.4 Relation between hippocampal activation and temporal di↵erence
score . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4.1 Methods and behavioral results . . . . . . . . . . . . . . . . . . . . . . 51

4.2 Subsequent clustering e↵ects . . . . . . . . . . . . . . . . . . . . . . . 54
4.3 HFA as a function of clustering and list type . . . . . . . . . . . . . . 55
4.4 Subsequent memory e↵ect for semantically isolated items . . . . . . . 57

5.1 Conditional response probability based on valence . . . . . . . . . . . 77

5.2 Semantic-valence CRP . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

xii
Chapter 1

Introduction

Our memories are composed of personal experiences as well as knowledge for

facts. The distinction between these types of memories was formalized by Endel
Tulving (Tulving, 1972) when he coined the term “episodic memory”. Episodic
memories are unique to an individual; many people know that Paris is the capital
of France, but only I can re-experience the night when I was visiting Paris and
was faced with the dilemma: nutella crêpe or jam crêpe? (The answer is get one,
and once you’ve walked one block to the next crêpe stand, get the other flavor.)
This experiential quality, the ability to perform a sort of mental time travel when
retrieving a memory, is one of the major features that dissociates episodic from
semantic memory, the memory for facts or knowledge devoid of such contextual
details.
Episodic memory can be measured in the laboratory through the free recall
paradigm. In free recall, a participant studies a list of items, typically words or
nameable objects, and either immediately after study or after a brief delay, must
recall any items from the preceding list. Each item is considered an “episode” and
the ability to retrieve an item is based on the success of reinstating the spatiotem-

1
poral context in which that episode was first experienced. As there are no external
cues available during test, free recall mirrors aspects of real-world retrieval.
Our understanding of episodic memory processes has greatly benefited from
behavioral studies utilizing free recall; however, behavioral studies cannot disen-
tangle successful encoding from successful retrieval processes as the behavioral
data collected during retrieval may reflect success during encoding, retrieval, or
both. In order to understand the mechanisms that support successful episodic
encoding, we utilize neuro-imaging methods to measure neural signals while par-
ticipants study lists of items. This enables the comparison between encoding items
that are subsequently remembered and subsequently forgotten. This subsequent
memory e↵ect (SME) analysis is a classic technique in the study of memory encod-
ing.

1.1 Spectral signals of episodic memory encoding

Researchers began studying memory encoding processes through electroencephalo-

graphic (EEG) recordings using electrodes placed on the scalp. EEG recordings
measure voltage potentials across the brain. Whereas an event related potential
(ERP) analysis can reveal voltage di↵erences between conditions of interest and
was one of the earliest measures used to study encoding processes (Sanquist et al.,
1980; Paller et al., 1987, 1988), EEG data contains other dimensions of information.
Specifically, any given voltage trace is composed of a series of sine waves of particu-
lar frequencies, amplitudes and phase characteristics. Unlike ERP analyses, where
the interpretation of polarity is unclear, these rhythmic signals directly reflect the
concerted activity of groups of neurons or cell assemblies (Lachaux et al., 2003).
These oscillatory or spectral signals arise due to modulations in neuronal excitabil-

2
ity that a↵ect the likelihood of spike output and sensitivity to synaptic input and
are considered the dominant communication mechanism of the brain (Fries, 2005;
Schnitzler & Gross, 2005; Jacobs & Kahana, 2010).
The power, or squared amplitude, of a spectral signal reveals how much energy
of a particular frequency is contained in an EEG trace. Decades of research, begin-
ning with Berger (1929) has shown that fluctuations in power at specific frequency
bands correlate with a wide array of cognitive, perceptual, and motor processes.
Although the exact definitions tend to vary, there are five commonly measured
frequency bands which include delta (2 - 4 Hz), theta (3 - 8 Hz), alpha (8 - 14 Hz),
beta (16 - 26 Hz), and gamma (> 30 Hz).
Across these five bands, two frequencies have emerged as critical players in
the service of memory: theta and gamma. As theta represents the “on-line” mode
of the hippocampus (Buzsáki, 2002) and hippocampus is essential for episodic
memory (Scoville & Milner, 1957), it would appear that theta is critical for memory.
However, theta fluctuations can be observed across a variety of processes, includ-
ing, but not limited to, arousal, anxiety, olfaction, motivation, and movement.
Indeed, theta is so ubiquitous that, in the words of Gyorgy Buzsaki (2005),

The only firm message that can be safely concluded ... is that in an
immobile animal no theta is present, provided that no changes occur in
the environment (and the animal is not “thinking”).

It is thus clear that theta does not map onto a single cognitive process. Instead,
it likely represents a cellular or biological signal which gives rise to all of these
processes. One hypothesis is that theta is directly related to long term potentia-
tion (LTP), the cellular mechanism of memory formation (Kandel, 2001). In LTP,
neighboring neurons that successively discharge action potentials (APs) within 10

3
to 30 ms of one another will be more likely to fire synchronously in the future, due
to a cascade of protein changes enhancing the connection between those neurons.
Stimulation at the peak of the theta rhythm will induce LTP (Axmacher et al., 2006;
Fell & Axmacher, 2011). Furthermore, the peak and trough of the theta phase in
dentate gyrus/CA1 may correspond to memory encoding and retrieval processes,
respectively (Hasselmo, 2005; Hasselmo & Stern, 2014). Outside of the hippocam-
pus, theta directs passage of information around multiple cortical loops and may
underlie top-down control of the hippocampus via prefrontal cortex (T. Mitchell
et al., 2008; Nyhus & Curran, 2010; Fell & Axmacher, 2011).
Despite this evidence suggesting that theta power should increase for memory
formation, theta increases have been consistently demonstrated in studies of human
memory. Although researchers have observed theta increases for subsequently
remembered relative to forgotten items (Klimesch, Doppelmayr, Russegger, &
Pachinger, 1996; Sederberg, Kahana, Howard, Donner, & Madsen, 2003), there
is comparable evidence for theta power decreases (Sederberg et al., 2003, 2006;
Guderian, Schott, Richardson-Klavehn, & Duzel, 2009; Lega, Jacobs, & Kahana,
2011). This discrepancy might result from di↵erences in methodology and/or data
analytic methods. For instance, it appears that post-stimulus theta fluctuations
may be dependent on the level of pre-stimulus theta fluctuations (Guderian et
al., 2009; Fell et al., 2011). Thus, among other parameters, how data are baseline
corrected could alter the observed theta results.
Somewhat less controversial is the finding that gamma or high frequency ac-
tivity (HFA, 30 - 200 Hz) increases for successful memory formation (Sederberg et
al., 2003; Gruber, Tsivilis, Montaldi, & Müller, 2004; Osipova et al., 2006; Sederberg
et al., 2006). Like theta, gamma is not specific to any single cognitive process,
but fluctuates across a range of behaviors including attention, visual search, ob-

4
ject recognition, learning, language processing and emotional evaluation (Engel
et al., 2001). Changes in gamma likely reflect one of two signals. First, gamma
oscillations can induce LTP by synchronizing pre- and post-synaptic activity; the
window during which two spikes must co-occur to induce LTP precisely matches
the interval of the gamma oscillation (Axmacher et al., 2006; Jensen et al., 2007;
Nyhus & Curran, 2010; Jutras & Bu↵alo, 2010). Alternatively, broad asynchronous
increases in HFA positively correlate with neuronal firing (Manning et al., 2009),
which may reflect large scale increases in multi-unit activity (Burke et al., 2015) and
may index cortical network processing (Ray et al., 2008; Jacobs & Kahana, 2009;
Canolty & Knight, 2010; Lachaux et al., 2012).
The critical issue that arises with HFA is not whether it increases for success-
ful memory formation, but more broadly, whether high frequency signals can be
detected using scalp EEG. Unlike intracranial EEG (iEEG) which measures signals
directly from the surface of the brain and magnetoencephalography (MEG) which
measures magnetic signals, high frequency signals in scalp EEG may be attenuated
by the skull. Specifically, volume conduction through the skull results in a low pass
spatial filtering (Nunez & Srinivasan, 2006). A recent study failed to find gamma
e↵ects in scalp EEG, despite finding evidence for such e↵ects in iEEG (Morton et
al., 2013). Additionally, there is evidence that gamma responses observed through
scalp EEG are artifactual and the result of saccades (Yuval-Greenberg et al., 2008).
In sum, the extant literature paints a complicated picture as to how theta and
gamma support memory formation. Given its role in LTP and communication
across the brain, theta power is expected to increase during successful memory
formation. However, such results are not found consistently across studies. And
although gamma increases are consistently observed across studies utilizing iEEG
and MEG, there are more broad concerns about the ability of scalp EEG to resolve

5
high frequency signals. With a diversity of recording modalities and analysis
methods, it is unclear whether these discrepancies are due to selection of analysis
parameters or actual di↵erences in recording methods. In Chapter 2,I directly
address this question by analyzing the SME using matched data analytic methods
in intracranial and scalp EEG datasets.

Convergent evidence from functional Magnetic Resonance Imaging HFA is

positively related to both spiking activity and the BOLD e↵ect in functional Mag-
netic Resonance Imaging (fMRI, Logothetis et al., 2001; Logothetis, 2003; Mukamel
et al., 2005; Niessing et al., 2005; Lachaux et al., 2007; Ojemann et al., 2013),
greatly facilitating comparisons across recording modalities, in particular between
fMRI and iEEG, the latter of which has more precise spatial resolution than scalp
EEG. fMRI studies have consistently shown that increased activation in left pre-
frontal cortex (PFC) and medial temporal lobe cortex is predictive of subsequent
memory (Wagner et al., 1998; Brewer, Zhao, Desmond, Glover, & Gabrieli, 1998;
Fernández et al., 1998; Fernandez, Brewer, Zhao, Glover, & Gabrieli, 1999; Strange,
Otten, Josephs, Rugg, & Dolan, 2002; Kim, 2011).
PFC activity is thought to reflect control mechanisms which guide encoding
and retrieval processes (Blumenfeld & Ranganath, 2007; Hayama & Rugg, 2009;
Thompson-Schill et al., 1997, 2005; Badre et al., 2005; Badre & Wagner, 2007)
whereas hippocampus engages in relational processing which binds information
into a single, contextually rich, “high-resolution” representation (Paller & Wag-
ner, 2002; Hwang & Golby, 2006; Mayes, Montaldi, & Migo, 2007; Brewer et al.,
1998; Eichenbaum, 2004; Brassen, Weber-Fahr, Sommer, Lehmbeck, & Braus, 2006;
Staresina & Davachi, 2006; Shrager, Kirwan, & Squire, 2008; Yonelinas, 2013). Fur-
thermore, PFC directly interacts with hippocampus through reciprocal connections

6
(Hwang & Golby, 2006; Dickerson et al., 2007) and PFC may use top-down pro-
cessing to guide or bias encoding and retrieval in hippocampus (Simons & Spiers,
2003; Preston & Eichenbaum, 2013).
These same regions have been identified in an iEEG study examining the SME
using HFA (Burke, Long, et al., 2014), providing convergent evidence that a specific
network supports successful encoding. However, it is unclear whether the SME
reflects memory-specific processing per se. As items may be forgotten due to
failures of any number of processes, including inattention, the contrast comparing
remembered and forgotten items cannot distinguish memory-specific encoding
processes from a mnemonic attentional signal. To overcome this limitation, it is
necessary to measure di↵erences at encoding based on how items are remembered,
not whether they are remembered. To define how items are remembered, I turn to
retrieved context theory.

1.2 Retrieved context theory

Retrieved context theory originated with Estes’ (1955) stimulus sampling model.
In brief, the theory posits the existence of a context representation which is an
amalgamation of internal states and external stimuli (McGeoch, 1942; Underwood,
1957; Bower, 1972). By simulating a slowly updating context representation which
is reinstated and used as a cue during retrieval, computational models of memory
(Howard & Kahana, 2002a; Sederberg et al., 2008; Polyn et al., 2009a; Lohnas et
al., in press) can capture many of the behavioral phenomena observed during free
recall tasks.
Retrieved context theory can account for the robust behavioral finding that
participants temporally cluster their recalls: when asked to recall items in any

7
order, participants typically consecutively recall study neighbors (Kahana, 1996;
Howard & Kahana, 1999; Sederberg et al., 2010). The phenomenon of temporal
clustering arises because during study, neighboring items will be associated to
similar context representations. Therefore, when a given context representation
is reinstated at test, it will be likely to cue both a given item and that item’s
neighbors. Importantly, the degree to which an item is associated to its context
can vary, determining the likelihood that a given item will be clustered with its
neighbors at recall.
Recent neuro-imaging studies have provided evidence corroborating the exis-
tence of a context representation. By measuring the similarity of encoding and
retrieval patterns, researchers have shown that the retrieval pattern of item i is
most similar to the encoding patterns of item i’s neighbors and that this similarity
decreases as a function of lag (Manning et al., 2011; Ya↵e et al., 2014).
How does the presence of this contextual representation relate to the SME? Be-
havioral evidence has shown that recall success is positively related to the amount of
temporal clustering exhibited across participants (Sederberg et al., 2010), suggest-
ing that associating items to their context promotes successful memory formation.
Thus, the SME may reflect contextual processing mechanisms. However, cluster-
ing could be the result of retrieval, rather than encoding, mechanisms and the SME
could reflect non-specific attentional signals. In Chapter 3, I investigate the neural
correlates of clustering in order to elucidate the ambiguous findings from the SME
analysis.

8
1.3 Semantic memory in the context of episodic mem-

ory

All of the studies discussed thus far have endeavored to understand the neural
mechanisms of episodic memory. However, it is clear that no participant enters a
laboratory experiment with a blank slate. Instead, a participant comes to the study
with a large body of knowledge and pre-existing associations which are likely to
influence the formation of new episodic associations. The distributed-plus-hub
theory of the organization of semantic memory suggests that the basis of semantic
memory extends across a wide range of cortex (Martin, 2007; Patterson et al., 2007,
though see Caramazza & Shelton, 1998; Tyler & Moss, 2001). These same cortical
regions show reactivation during retrieval of episodic memories (Polyn et al.,
2005; Rissman & Wagner, 2012), suggesting that these two systems are intimately
connected.
In free recall, semantic clustering, or consecutive recall of semantic associates,
is a robust finding (Bousfield, 1953; Long et al., 2010). Like temporal clustering,
semantic clustering can be explained by retrieved context theory whereby seman-
tic information is incorporated into the slowly updating context representation.
In fact, retrieved context theory can account for clustering along a variety of di-
mensions including categorical clustering, consecutively recalling items from the
same category, as well as spatial clustering, consecutively recalling items studied
in nearby spatial locations (Manning et al., 2012; Morton et al., 2013; J. F. Miller et
al., 2012; J. F. Miller, Neufang, et al., 2013). Such a mechanism could also explain
memory benefits for emotionally valenced items, if emotional items tend to be
clustered at recall (LaBar & Cabeza, 2006),
Critically, though these results illustrate the presence of semantic information

9
during episodic tasks, they do not address how semantic associations influence the
formation of episodic memories. In Chapters 4 and 5, I will address how semantic
memory and emotional valence interact with episodic encoding.

1.4 Overview

In this thesis, I will attempt to address outstanding questions about the neural
mechanisms that support memory formation and what cognitive processes those
signals represent. Chapter 2 identifies the spectral correlates of successful mem-
ory formation across both intracranial and scalp EEG studies. Chapter 3 extends
these findings using cognitive theory to relate contextual processing signals to the
traditional subsequent memory e↵ect. Chapter 4 measures the impact of seman-
tic knowledge on the formation of episodic associations. Chapter 5 investigates
whether retrieved context theory can account for the memory benefit observed for
emotionally valenced items.

10
Chapter 2

Subsequent memory e↵ect in

intracranial and scalp EEG

Nicole M. Long, John F. Burke, & Michael J. Kahana

NeuroImage, 84, 488–494

2.1 Abstract

Successful memory encoding is marked by increases in 30-100 Hz gamma-band

activity in a broad network of brain regions. Activity in the 3-8 Hz theta band has
also been shown to modulate memory encoding, but this e↵ect has been found to
vary in direction across studies. Because of the diversity in memory tasks, and
in recording and data-analytic methods, our knowledge of the theta frequency
modulations remains limited. The di↵erence in the directionality of these theta
e↵ects could arise from a distinction between global cortical and deeper subcortical
e↵ects. To address this issue, we examined the spectral correlates of successful
memory encoding using intracranial EEG recordings in neurosurgical patients and

11
scalp EEG recordings in healthy controls. We found significant theta (3-8 Hz) power
modulations (both increases and decreases) and high gamma (44 - 100 Hz) power
increases in both samples of participants. These results suggest that (1) there are
two separate theta mechanisms supporting memory success, a broad theta decrease
present across both the cortex and hippocampus as well as a theta power increase
in the frontal cortex, (2) scalp EEG is capable of resolving high frequency gamma
activity, and (3) iEEG theta e↵ects are likely not the result of epileptic pathology.

2.2 Introduction

Memory processes during encoding that give rise to successful retrieval are collec-
tively termed subsequent memory e↵ects (SMEs, Paller & Wagner, 2002) and have
been characterized using scalp electroencephalography (EEG, Paller et al., 1987;
Klimesch et al., 1997; Sederberg et al., 2006), magnetoencephalography (MEG,
Osipova et al., 2006; Guderian et al., 2009), and intracranial EEG recorded in neuro-
surgical patients undergoing treatment for intractable epilepsy (iEEG, Fernandez,
E↵ern, et al., 1999; Fell et al., 2001; Sederberg et al., 2003). Whereas these recording
modalities have millisecond temporal resolution, scalp EEG is limited by poor spa-
tial resolution and may not reveal changes in high frequency activity due to muscle
and eye movement artifacts that generate their own high frequency electrical sig-
nals (Yuval-Greenberg et al., 2008; Muthukumaraswamy, 2013). In comparison,
iEEG o↵ers subcentimeter range spatial resolution and the ability to directly record
from deep brain structures. However, iEEG can only be recorded in neurosurgical
patients leading some to question the generalizability of these results to neurolog-
ically healthy individuals.
Both iEEG and scalp EEG have been e↵ectively used to study the spectral

12
correlates of memory encoding. Although most studies show gamma (30 - 100 Hz)
power increases for subsequent memory (Gruber et al., 2004; Sederberg et al., 2006;
Osipova et al., 2006; Serruya et al., 2014), direct comparisons cannot be easily made
because of di↵erences in experimental and data analytic methods. For example,
Morton et al. (in press) measured category-specific oscillatory patterns and found
that high gamma was more informative in iEEG than scalp EEG. However, the scalp
study included a preliminary session in which participants rated the familiarity of
the experimental stimuli. As gamma e↵ects are often observed for primacy items
(Sederberg et al., 2006; Serruya et al., 2014), pre-exposure to the items may have
dampened potential scalp gamma e↵ects.
Theta frequency (3-8 Hz) activity has exhibited both increases and decreases
during successful memory formation (Burgess & Gruzelier, 1997; Klimesch, 1999;
Sederberg et al., 2003, 2006; Osipova et al., 2006; Guderian et al., 2009; Lega et
al., 2011; Hanslmayr & Staudigl, 2013). The inconsistent patterns observed in the
theta band could arise from a number of factors including the task parameters and
the brain regions, time windows, and frequencies analyzed. For example, there
may be di↵erential e↵ects of theta power based on anatomical location, with the
hippocampus showing an increase in theta power and neocortical regions showing
decreases (Lisman & Jensen, 2013).
Our goal here is to compare the spectral SMEs measured using both intracranial
and scalp EEG by controlling as many of these variables as possible. Using identical
data analytic methods and roughly corresponding brain regions, we analyzed
data from neurosurgical patients (n=93) and healthy participants (n=102) who
participated in a free recall study. To foreshadow our results, we found very
similar patterns of results in both iEEG and scalp EEG indicating that memory
e↵ects observed in iEEG can be directly translated to healthy individuals and that

13
high frequency e↵ects can be detected by scalp EEG.

2.3 iEEG Methods

Participants

98 participants with medication-resistant epilepsy underwent a surgical procedure

in which electrodes were implanted subdurally on the cortical surface as well as
deep within the brain parenchyma. In each case, the clinical team determined
the placement of the electrodes so as to best localize epileptogenic regions. De-
mographic and electrode information are described in publications on the same
dataset (Burke et al., 2013).
Data were collected at 4 hospitals: Children’s Boston (Boston, MA), Hospital
of the University of Pennsylvania (Philadelphia, PA), Freiburg University Hospi-
tal (Freiburg, Germany), and Thomas Je↵erson University Hospital (Philadelphia,
PA). The research protocol was approved by the IRB at each hospital and informed
consent was obtained from the participants and their guardians. We restricted our
analysis to include only those patients (n=93) who were left–hemispheric language
dominant, as assessed by either the patients’ handedness or a clinically adminis-
tered intracarotid injection of sodium amobarbital (Wada test). As the electrode
placements in these 93 patients were clinically determined, each patient did not
have electrodes in all of our regions of interest (see Methods, below). Therefore,
the total number of patients per region of interest varied as a function of electrode
placement and the total number of patients for a given region of interest ranged
from 29 (left inferior prefrontal cortex) to 55 (non-hippocampal medial temporal
lobe cortex).

14
Experimental paradigm

Each patient participated in a delayed free-recall task in which they were instructed
to study lists of words for a later memory test; no encoding task was used. Lists were
composed of either 15 (67/93 patients) or 20 common nouns, chosen at random and
without replacement from a pool of high frequency nouns (either English or Ger-
man, depending on the subject’s native language; http://memory.psych.upenn
.edu/WordPools). Each sequentially presented word remained on the screen for
1600 ms, followed by a randomly jittered 800-1200 ms blank inter-stimulus interval
(ISI).
Immediately following the final word in each list, participants were given a
distraction task designed to attenuate the recency e↵ect (Kahana, 2012). The dis-
traction task was a series of arithmetic problems of the form A+B+C=??, where A,
B and C were randomly chosen integers ranging from 1-9. The distraction interval
lasted at least 20 sec, but patients were allowed to complete any problem that they
started resulting in a variable distraction interval (average duration, 25 sec).
Following the distraction period, participants were given 45 seconds to freely
recall as many words as possible from the list in any order. Vocalizations were
digitally recorded and subsequently manually scored for analysis. On average,
patients participated in two sessions yielding an average total of 14 lists. Any
session in which probability of recall was less than 15% was excluded from the
final analysis, resulting in an average of one session per patient.

Electrophysiological recordings and data processing

iEEG data were recorded using a Bio-Logic, DeltaMed, Nicolet, GrassTelefactor, or

Nihon Kohden electroencephalogram (EEG) system. Depending on the amplifier

15
and the discretion of the clinical team, the signals were sampled at 256, 400, 500,
512, 1000, 1024, or 2000 Hz. Signals were referenced to a common contact placed
either intracranially or on the scalp or mastoid process.

2.4 Scalp Methods

Participants

102 (60 female) paid volunteers (ages 18 - 29), were recruited via fliers posted
around the University of Pennsylvania campus. Participants were provided with
a base monetary compensation plus an additional performance-based monetary
incentive to ensure full e↵ort. Our research protocol was approved by the Institu-
tional Review Board at the University of Pennsylvania, and informed consent was
obtained from all participants.

Experimental paradigm

The data reported in this manuscript were collected as part the Penn Electrophys-
iology of Encoding and Retrieval Study, involving three experiments that were
sequentially administered. The data reported here come from participants who
took part in Experiment 1. The methods are briefly summarized below, and com-
plete description of the methods can be found in (Lohnas & Kahana, 2013; J. F. Miller
et al., 2012).
Each session consisted of 16 lists of 16 words presented one at a time on a
computer screen. Each study list was followed by an immediate free recall test.
The study also included an encoding task manipulation: some lists were en-
coded freely (no task lists) whereas other lists were encoded using a size or animacy
task (e.g. will the item fit in a shoebox? Is it living or non-living?). To facilitate com-

16
parison between scalp EEG and iEEG studies, we only considered the “no-task”
lists in these analyses.
Words were drawn from a pool of 1638 words (http://memory.psych.upenn
.edu/WordPools). Each item was on the screen for 3000 ms, followed by jittered
800 - 1200 ms inter-stimulus interval. After the last item in the list, there was a 1200
- 1400 ms jittered delay, after which a tone sounded, a row of asterisks appeared,
and the participant was given 75 s to attempt to recall any of the just-presented
items.
To maintain consistency with the iEEG dataset, we only analyzed the first 1600
ms of word presentation. As the scalp paradigm utilized immediate free recall,
to further constrain the comparison, we excluded words from late serial positions
(13 - 16) to minimize e↵ects of recency. Additionally, each participant completed
seven experimental sessions; however, to more closely match the iEEG dataset and
reduce the influence of practice e↵ects, we only analyzed the first 4 sessions. It was
with this number of sessions that iEEG and scalp EEG datasets had on average an
equal number of recall events.

Electrophysiological recordings and data processing

EEG measurements were recorded using Geodesic Sensor Nets (GSN; Netstation 4.3
acquisition environment, from Electrical Geodesics, Inc.). The GSN provided 129
standardized electrode placements across participants. All channels were digitized
at a sampling rate of 500 Hz, and the signal from the caps was amplified via
either the Net Amps 200 or 300 amplifier. Recordings were initially referenced
to Cz and later converted to an average reference. Channels that demonstrated
high impedance or poor contact with the scalp were excluded from the average
reference.

17
 To identify epochs contaminated with eyeblink and other movement artifacts,
electrooculogram (EOG) activity was monitored bipolarly using right and left elec-
trode pairs (electrodes 25, 127 and 8 and 126 on the GSN). An individual word
presentation event was rejected from subsequent analyses if the weighted running
average for either the right or the left EOG pair exceeded a 100 µV threshold. Ad-
ditionally, events were excluded on a per channel basis for each participant if the
voltage on a particular event/channel pair exceeded a pre-determined threshold.
The threshold was set as 4.5 times the standard deviation of the mean voltage cal-
culated across all electrodes (excluding those on the eyes, face and neck, electrodes
F10, 8, FPZ, 17, FPZ, 25, F9, T9, 56, 63, 99, 107, 113, T10, 126, 127) for a single session
for each participant.

Analysis of iEEG and scalp EEG data

Behavioral analysis We calculated probability of recall as well as the number of

intrusions, or incorrect recalls, for participants in both datasets. Intrusions were
measured for each participant and each list by calculating the number of incorrect
words recalled (words not from the preceding study list) as a proportion of the
total number of words recalled. Values were then averaged across lists for each
participant.

Oscillatory analysis To minimize confounds resulting from volume conduction

and saccades, we analyzed both the iEEG and scalp EEG with bipolar referencing
(Nunez & Srinivasan, 2006; Kovach et al., 2011). We defined the bipolar mon-
tage in our dataset based on the geometry of the iEEG and scalp EEG electrode
arrangements. For each participant and electrode, the raw EEG signal was first
downsampled to 200 Hz and a fourth order 2 Hz stopband butterworth notch filter

18
was applied at 50 or 60 Hz to eliminate electrical line noise. We isolated pairs
of immediately adjacent electrodes and found the di↵erence in voltage between
them (Burke et al., 2013). The resulting bipolar signals were treated as new virtual
electrodes and are referred to as such in the remainder of the text. The Morlet
wavelet transform (with a wave number of 6) was used to compute spectral power
as a function of time for all EEG signals during word presentation (0 - 1600 ms)
and a 1000 ms bu↵er was included on both sides of the data to minimize edge
e↵ects. Frequencies were sampled logarithmically at 46 intervals between 2 and
100 Hz. Power values were then down-sampled by taking a moving average across
100 ms time windows from stimulus onset and sliding the window every 50 ms,
resulting in 31 total time windows with 16 non-overlapping time windows. Log
transformed power values were then Z-transformed to normalize power within
participants. Power was Z-transformed according to the mean and standard devi-
ation of the power across all events within a session, separately for each participant,
electrode and frequency.

ROI selection and analysis iEEG ROIs were selected a priori by Brodmann area
or gyrus. We were interested in regions most commonly associated with mem-
ory encoding and retrieval (Wagner et al., 1998; Blumenfeld & Ranganath, 2007;
Sederberg, Schulze-Bonhage, Madsen, Bromfield, McCarthy, et al., 2007; Shrager
et al., 2008; Kim, 2011), and therefore chose the following eight ROIs: bilateral
inferior temporal cortex (BA 20, 21), bilateral inferior frontal cortex (IFC, BA 45,
47), bilateral dorsolateral prefrontal cortex (DLPFC, BA 46, 9), bilateral hippocam-
pus and parahippocampal cortex. Localizations were radiologically determined
by a neurologist at each of the four hospitals. The number of participants with at
least one electrode in each of these regions is in parentheses in Table 2.1A. Scalp

19
 Anterior
Dorsolateral Superior
PFC
Anterior
Inferior PFC Inferior

Inferior Posterior
temporal lobe Inferior

ECoG ROIs Scalp ROIs

Figure 2.1: Maps of a priori selected regions of interest.

ROIs were selected a priori (Weidemann et al., 2009) to loosely match the cortical
iEEG ROIs with bilateral anterior superior (AS, corresponding to DLPFC), bilateral
anterior inferior (AI, corresponding to IFC), and bilateral posterior inferior (PI,
corresponding to inferior temporal). Figure 2.1 shows both iEEG and scalp EEG
topographies for the current study.

Frequency e↵ects across time The data were split into five distinct bands, theta
(3 to 8 Hz), alpha (10 to 14 Hz), beta (16 to 26 Hz), low gamma (28 to 42 Hz)
and high gamma (44 to 100 Hz), by taking the mean of the Z-transformed power
in each frequency band. Z-transformed power was filtered into the conditions
being analyzed (e.g. subsequently recalled and subsequently forgotten events)
and an unpaired t-test was run comparing the two sets of events separately for
each participant, electrode and frequency band across the 1600 ms duration of
encoding word presentation. T-statistics were then averaged across electrodes
within an ROI. Averaging t-statistics means that only signals that are consistent
across an ROI will appear significant, as opposing e↵ects will be cancelled out. We
chose this method as we were interested in general e↵ects across an ROI and not
regional di↵erences within an ROI. This averaging step yielded a single t-statistic
for each participant and frequency band for a given ROI. Within an ROI and for a
particular frequency band, an unpaired t-test was calculated across the participant

20
t-statistics. All resultant t-statistics are presented in Table 2.1.
In cases in which we ran post-hoc tests on z-scored power to determine the
laterality of frequency e↵ects, we used a Bonferroni corrected p value.

Time-frequency analysis We used a modified version of the bootstrap method

detailed in previous studies (Sederberg et al., 2006; Serruya et al., 2014). We
corrected for comparisons across 16 time windows and 46 frequencies.
For each participant, electrode, time window and frequency a t-statistic was gen-
erated through an unpaired t-test comparing the z-scored power of subsequently
recalled to not recalled items. These t-statistics were averaged across electrodes
within an ROI, creating a single t-statistic for each participant, time window and
frequency for one ROI. The distribution of participant t-statistics were compared to
zero using an unpaired t-test, resulting in a single across participant t-statistic for
each time window and frequency. To correct for multiple comparisons, we ran a
bootstrap procedure in which we generated a null distribution of across participant
t-statistics. For the bootstrap analysis, we followed the same procedure as above,
however, the tests were carried out on shu✏ed data, such that the labels of sub-
sequently recalled and not recalled were randomly assigned to events. Instead of
a single across participant t-statistic being generated, 1,000 null across participant
t-statistics were generated from 1,000 iterations of the bootstrap procedure. Finally,
all null t-statistics from the 16 time windows and 46 frequencies were concatenated
into a single distribution of 736,000 values. To determine which real t-statistics, and
thus time windows and frequencies, were significant at a p = .05 level corrected for
multiple comparisons, we found the top and bottom 2.5% of the null distribution.
Any real t-statistics which exceeded those values were labeled significant and the
corresponding z-scored power di↵erences (subsequently recalled - not recalled)

21
appear in the time frequency plots in Figure 2.3.
In cases in which we ran post-hoc tests on z-scored power to determine the
temporal or regional specificity of frequency e↵ects, we used a Bonferroni corrected
p value.

2.5 Results

Before examining the spectral components of the subsequent memory e↵ect we

report the basic behavioral data for the two studies. In the scalp EEG study, partic-
ipants recalled an average of 68% of the studied items (SD=14%) and committed
an average of .31 recall errors (0.12 prior list intrusions [SD = .09] and 0.19 extra list
intrusions [SD = .17]) on each list. Neurosurgical patients who participated in the
iEEG study recalled an average of 24% of studied items (SD = 9%) and committed
an average of 4.2 recall errors (0.64 prior list intrusions [SD = .63] and 3.57 extra list
intrusions [SD = 3.58]) per list. The finding of substantially lower recall and higher
intrusion rates in the iEEG study was to be expected both because the task was in-
herently more difficult (delayed free recall for the iEEG participants vs. immediate
free recall for the scalp EEG participants) and because of the obvious di↵erences in
the populations being studied (a community sample of neurosurgical patients with
medial temporal lobe epilepsy vs. an elite college population). We also obtained
a measure of general intelligence (Wechsler Adult Intelligence Scale) for 77 of the
scalp EEG participants and 74 of the iEEG participants. As expected IQ scores for
the scalp EEG participants (M = 128, SD =10) were substantially higher than for
the iEEG participants (M = 98, SD =14).
We characterized the spectral components of the SME by comparing power in
five frequency bands (theta, alpha, beta, low and high gamma) across items subse-

22
 A. iEEG B. Scalp
DLPFC AS
Left (31) Right (42) Left (102) Right (102)
Theta -2.44* -5.56** -1.30 -2.65**
Alpha -3.99** -1.93 -3.69** -4.54**
Beta -4.01** -1.95 -2.41* -1.81
Gamma1 2.23* -0.43 0.01 -0.31
Gamma2 3.58** 1.44 1.35 1.35
IFC AI
Left (29) Right (38) Left (102) Right (102)
Theta -2.39* -4.02** -0.85 -1.58
Alpha -2.66* -2.53* -3.65** -2.97**
Beta -3.19** -1.80 -2.37* -1.86
Gamma1 -0.38 -0.35 -0.93 -0.96
Gamma2 1.70 1.49 0.20 -0.09
InfTem PI
Left (52) Right (51) Left (102) Right (102)
Theta -6.16** -3.49** -3.20** -2.29*
Alpha -6.74** -3.76** -4.08** -3.69**
Beta -4.56** -3.78** -0.37 -0.35
Gamma1 1.24 -1.48 0.84 1.08
Gamma2 4.54** 1.99 1.48 1.48
C. iEEG Subcortical
Hippocampus (47) MTL (55)
Theta -3.60** -5.92**
Alpha -4.12** -5.76**
Beta -2.63* -1.98
Gamma1 -.14 2.86**
Gamma2 2.23* 4.18**
Table 2.1: Subsequent memory e↵ect in iEEG and scalp EEG studies T-statistics for the comparison
of z-scored power for recalled - not recalled items, across the five frequency bands and for (A) iEEG
cortical, (B) Scalp, and (C) iEEG subcortical ROIs. Numbers in parentheses denote number of
participants with electrodes in each region. * p < .05, ** p < .01

23
quently recalled and subsequently forgotten across the presentation interval (1600
ms). The comparison of subsequently recalled minus not recalled items revealed
low frequency power decreases and high frequency power increases (Figure 2.2A-
C). We found significant theta and alpha power decreases across all ROIs with
the exception of a nonsignificant alpha e↵ect in RIFC (Table 2.1A-C). Beta power
was significantly decreased across most ROIs except right DLPFC, bilateral PI, and
parahippocampus.
Gamma e↵ects
were less widespread
A. iEEG B. Scalp
Left Right Left Right
with significant high
DLPFC AS
gamma increases
predominantly lo-
IFC AI

calized to the left

0.2 0.2

Delta Z−Power
Delta Z−Power

hemisphere, includ- InfTem 0 PI 0

ing left DLPFC

−0.2 −0.2
3 6 12 25 49 99 3 6 12 25 49 99
Frequency Frequency

C. iEEG Subcortical
and left inferior 0.2 0.2
Delta Z−Power

Delta Z−Power
Hippocampus 0 Parahippocampus 0
temporal cortex (Fig-
−0.2 −0.2
3 6 12 25 49 99 3 6 12 25 49 99

ure 2.2A). Gamma

Frequency Frequency

e↵ects were also

Figure 2.2: Z-scored power for recalled minus not recalled items for all
evident in both frequencies (3 to 100 Hz) and collapsed across 1600 ms for (A) iEEG Cortical,
(B) scalp, and (C) iEEG Subcortical ROIs. Shaded regions around each
hippocampal and curve are Loftus Masson 95% confidence intervals.
parahippocampal ROIs,
as well as scalp bilateral AS and PI (Figure 2.2B,C). To test the apparent gamma
laterality in iEEG we ran paired t-tests comparing high gamma power across the
three pairs of left and right ROIs in those participants with at least one electrode in
each ROI. Gamma power in left DLPFC was significantly more increased than in

24
the right (t(17) = 3.4, p = .004; critical p-value set at 0.02; Bonferroni corrected 0.05/3).

Temporal dynamics of the subsequent memory e↵ect

One of the main
benefits of EEG
A. iEEG B. Scalp
Left Right Left Right
data is its high 0.1

0.08

temporal resolution.
DLPFC
AS
0.06

0.04

Subtle e↵ects that 0.02

Delta Z−Power
IFC AI 0

may exist on smaller

−0.02

99 99 −0.04
Frequency

35

Frequency
35 −0.06

time scales could

InfTem PI
12 12
−0.08

4 4
−0.1
0 500 1000 1500 0 500 1000 1500

be obscured by 99
C. iEEG Subcortical
99
Frequency

35

Frequency
35

collapsing data across Hippocampus

12 Parahippocampus
12

4 4
0 400 800 1200 0 400 800 1200

large time inter-

vals (as above).
Figure 2.3: Time frequency spectrograms for recalled minus not recalled
z-scored power for all frequencies (3 to 100 Hz) and across 100 ms time
Though not ob-
windows spanning 0 to 1600 ms for (A) iEEG Cortical, (B) scalp, and (C)
iEEG Subcortical ROIs. All time-frequency pairs of z-power values were
served when time
tested using a bootstrap procedure (see Methods) and any area in white
did not pass the significance threshold of p = .05 corrected for multiple
intervals were col-
comparisons.
lapsed, when ex-
amining 100 ms windows between 0 and 1600 ms, we found significant theta
power increases in left DLPFC, bilateral AI, and right PI around 500 ms (Figure
2.3A-B). This theta power increase was not specifically left- or right-lateralized for
either iEEG or scalp ROIs, as revealed by paired t-tests comparing theta power
di↵erences from 400 - 600 ms in left and right DLPFC and PI in those participants
with at least one electrode in each ROI (ts < 2, ps > .1 critical p-value set at 0.03;
Bonferroni corrected 0.05/2).
All ROIs, including hippocampus, showed late theta decreases (Figure 2.3A-C).

25
We ran a 6 x 2 repeated measures ANOVA on the scalp EEG z-scored theta power
comparing all 6 ROIs and 2 time windows (early, 0 - 500 and late, 1000 - 1500 ms).
This analysis revealed no main e↵ect of ROI (F(5,505) = 1.6, p = .2), a main e↵ect
of time window (F(1,101) = 19.5, p < .0001), and no interaction (F(5,1111) = 2.2, p
= .06). A post-hoc t-test of z-scored theta power averaged across all ROIs revealed
that theta power was significantly decreased in the late time window relative to
the early time window (t(101) = 4.4, p < .0001). To assess the temporal dynamics
of z-scored theta power in the iEEG dataset we ran paired t-tests comparing theta
power from 0 - 500 ms and 1000 - 1500 ms in each ROI, as not all participants
contributed electrodes to all ROIs. Theta was significantly decreased in the late
time window for all ROIs (ts > 3.0, ps < .005, critical p-value set at .006, Bonferroni
corrected .05/8) except left DLPFC (t(30) = 2.4, p = .02).
Increased temporal precision showed that gamma power increases were present
both in iEEG and scalp EEG (Figure 2.3A-C). We ran a 6 x 2 repeated measures
ANOVA on the scalp EEG z-scored gamma power comparing all 6 ROIs and 2
time windows (early, 0 - 500 and late, 1000 - 1500 ms). This analysis revealed no
main e↵ect of ROI (F(5,505) = .91, p = .48), a main e↵ect of time window (F(1,101)
= 19.0, p < .0001), and no interaction (F(5,1111) = 1.9, p = .1). A post-hoc t-test of
z-scored gamma power averaged across all ROIs revealed that gamma power was
significantly increased in the early time window relative to the late time window
(t(101) = 4.4, p < .0001). To assess the temporal dynamics of z-scored gamma power
in the iEEG dataset we ran paired t-tests comparing gamma power from 0 - 500 ms
and 1000 - 1500 ms in each ROI , as not all participants contributed electrodes to all
ROIs. Gamma was significantly increased in the late time window for left DLPFC
and left IFC (ts > 3.5, ps < .006, critical p-value set at .006, Bonferroni corrected
.05/8). There was no significant di↵erence in gamma power for the remaining ROIs

26
(right DLPFC, right IFC, bilateral IT, hippocampus or PHC ts < 2.5, ps > .03).

2.6 Discussion

Memory formation elicited a remarkably similar pattern of results across scalp and
iEEG recordings. Across the encoding interval a general pattern of low frequency
decreases and high frequency increases was present in both datasets. While theta
power decreases and gamma power increases were evident across the encoding
interval, a more precise examination of the temporal dynamics revealed theta
increases in addition to the decreases. iEEG and scalp EEG showed significant
theta increases around 500 ms post stimulus onset. That both e↵ects predominantly
localized to the frontal region suggests that they may reflect the ‘frontal mid-line
(FM) theta’ pattern often observed during cognitive tasks (for a comprehensive
review, see D. Mitchell et al., 2008). The origin of FM theta in human EEG recording
is unclear; midline frontal areas, such as the anterior cingulate cortex (ACC), are
most commonly cited as potential sources (Gevins, Smith, McEvoy, & Yu, 1997;
Sauseng, Hoppe, Klimesch, Gerlo↵, & Hummel, 2007) and might explain the lack
of a laterality e↵ect in the current study. Additionally, there is evidence that though
generated in medial PFC, frontal midline theta extends outward into a network
of regions encompassing lateral PFC (Mizuhara et al., 2004). Our results show
that these theta increases are conserved across both iEEG and scalp EEG and are
consistent with other scalp EEG studies showing theta power increases (Klimesch
et al., 1997, 1998; Hanslmayr et al., 2011).
In addition to very circumscribed theta power increases in frontal cortex, we
also observed broad theta power decreases across iEEG and scalp EEG, including
hippocampus. There are two hypotheses to explain the decreases in theta power.

27
First, it has been suggested (Stoller, 1949) that theta decreases in iEEG reflect de-
creases in alpha power as the alpha rhythm may be slowed in epileptic brains.
However, we found highly similar theta power decreases in healthy controls in
the scalp study, suggesting that theta power decreases in iEEG are not an artifact
of the patient population. A second hypothesis is that theta power decreases and
increases are separate properties of the neocortex and the hippocampus, respec-
tively (Lisman & Jensen, 2013). However, we observe the same decreases across
both neocortical and hippocampal ROIs, which runs counter to this hypothesis.
Previous work using a subset of the data presented here (Lega et al., 2011) has
shown theta power increases in the hippocampus for subsequently remembered
items. While the current study does not show this e↵ect, it is likely due to the
fact that Lega et al. (2011) specifically regressed out broadband shifts in spectral
power in order to detect oscillations. As broadband activity is known to correlate
with local field potentials (Manning et al., 2009) and could potentially be related
to memory signals, we did not wish to bias ourselves to only detecting oscillations.
Additionally, while Lega et al. (2011) specifically focused on theta power increases,
they also observed significant theta power decreases that occurred roughly twice
as often as theta power increases (cf Figures 2A and 3 in Lega et al.), consistent
with the results reported here.
We hypothesize that these conflicting results of theta power increases and de-
creases reflect two competing e↵ects: shifts in broadband power and narrow-band
theta oscillations, leading to subsequent memory e↵ects characterized either by
theta power decreases or increases, respectively. Furthermore, broadband power
shifts appear to be much larger than narrow-band changes, resulting in the overall
decrease in theta power reported here and in other studies (Sederberg, Schulze-
Bonhage, Madsen, Bromfield, McCarthy, et al., 2007; Guderian et al., 2009; Burke et

28
al., 2013). However, when these broadband power changes are removed, as in Lega
et al. (2011), increases in theta power are more readily observed. Consistent with
the hypothesis of two separate theta e↵ects, Burke et al. (2013) recently found that
theta synchrony, presumably a more specific marker of theta oscillatory activity,
exhibits both increases and decreases during memory formation.
One limitation of the current study is the use of a single task, free recall, to
measure encoding processes. It is possible that the e↵ects observed may be specific
to free recall and thus may not be observed across other memory paradigms or
with an analysis comparing di↵erent sets of events as opposed to subsequently
recalled and not recall items (Hanslmayr & Staudigl, 2013). However, there is
some evidence that theta power decreases and gamma power increases are task
independent memory signals as this pattern has been observed in a subsequent
memory study utilizing recognition (Matsumoto et al., 2013). Our results provide
compelling evidence to motivate future memory studies and to investigate the role
of theta power in memory processes.
Although we have presented topographies of the subsequent memory e↵ect
across scalp and intracranial datasets for roughly corresponding regions of interest,
we fully recognize that scalp EEG does not permit the identification of signal gen-
erating sources with anywhere near the precision of subdural electrode recordings.
Indeed, given the di↵erence in timing of memory-related gamma-band activity
(discussed below), frontal gamma e↵ects in scalp are potentially more related to
activity in the medial temporal lobe.
In addition to late low frequency decreases, both iEEG and scalp EEG showed
high frequency increases. In iEEG, significant gamma power increases were ev-
ident across the encoding interval for left cortical and all subcortical ROIs. The
time frequency analysis revealed that these e↵ects were present across the 1600

29
ms encoding interval for all ROIs with the exception of left DLPFC and left IFC
which showed significantly greater gamma power in the late (1000 - 1500 ms) time
window. In comparison, significant gamma e↵ects were not present across the en-
coding interval for the scalp EEG dataset, although gamma e↵ects were typically
in the positive direction. The time frequency analysis revealed significant gamma
e↵ects across all ROIs for the early (0 - 500 ms) time window.
As gamma is considered a mapping signal related to the BOLD activation ob-
served with fMRI (Crone et al., 2011; Lachaux et al., 2012; Burke et al., 2013), we
would expect that the gamma e↵ects in iEEG would closely mirror the subsequent
memory e↵ects observed with fMRI. Scalp EEG, due to its low spatial resolu-
tion, would be less likely to map directly onto the signals observed in fMRI and
iEEG. The significant gamma results in scalp EEG suggest that despite concerns
about interference from eye and muscle movement (Yuval-Greenberg et al., 2008;
Muthukumaraswamy, 2013), which may still be present here, as well as general
attenuation of spectral power due to the skull (Voytek et al., 2010), scalp EEG is
able to resolve high frequency gamma e↵ects, at least up to 100 Hz.
It is clear from our results that across both intracranial and scalp EEG the
dominant electrophysiological e↵ect of successful memory encoding is an overall
skew in power toward higher-frequencies at the expense of lower-frequencies. The
meaning of this pattern vis-a-vis episodic memory is an open question, but we
note that a similar pattern of results is found across a wide variety of electrophysi-
ological recordings during behaviors ranging from motor movement (K. J. Miller
et al., 2007; Crone et al., 1998a, 1998b), to auditory tone perception (Crone et
al., 2001), among others. Indeed, this pattern is consistent with the event-related
synchronization/desynchronization (ERS/ERD) processes that have been described
outside of the memory literature (see Pfurtscheller & Lopes Da Silva, 1999 for

30
a review). Furthermore, recent studies have found that this pattern of spectral
changes correlates well with the fMRI BOLD signal (Kilner et al., 2005; Niessing
et al., 2005). Our results show that this pattern is fairly conserved across the brain,
for both iEEG and scalp EEG, given that low frequency power decreases and high
frequency power increases are observed with slight variation across all ROIs. It
is therefore possible that the spectral content of both iEEG and scalp EEG during
memory formation may not reflect a memory specific signal per se, but rather may
indicate a more non-specific underlying process of general cortical activation and
that it is the precise intersection of timing and spatial location of these e↵ects that
is important.
Beyond this characterization of memory processes, these results suggest that
while data from individual patients might not be reflective of normal functioning,
the average e↵ects across a large patient population are representative of the general
population. Additionally, these results suggest that, despite its limited spatial
resolution and potential muscle artifacts, scalp EEG measures qualitatively similar
physiological processes as more precise yet more invasive recording techniques.

31
Chapter 3

Contextual encoding mechanisms in

hippocampus and prefrontal cortex
support successful memory
formation.

Nicole M. Long & Michael J. Kahana

3.1 Abstract

To understand how memories are successfully formed, scientists have compared

neural activity during the encoding of subsequently remembered and forgotten
items. Though this approach has elucidated a network of brain regions involved
in memory encoding, this method cannot distinguish broad, non-specific signals
from memory specific encoding processes, such as associative encoding. Associa-
tive encoding, which is a key mechanism of learning, can be seen in the tendency

32
of participants to successively recall, or cluster, study neighbors. We assessed the
electrophysiological correlates of associative processing by comparing intracra-
nially recorded EEG activity during the encoding of items that were subsequently
recalled and clustered; recalled and not clustered; or not recalled. We found that
high frequency activity (HFA) in left prefrontal cortex, left temporal cortex and
hippocampus increased during the encoding of subsequently recalled items. Crit-
ically, the magnitude of this e↵ect was largest for those recalled items that were
also subsequently clustered. HFA temporally dissociated across regions, with in-
creases in left prefrontal cortex preceding those in hippocampus. Furthermore, late
hippocampal HFA positively correlated with behavioral measures of clustering.
These results suggest that associative processes linking items to their spatiotem-
poral context underlie the traditionally observed subsequent memory e↵ect and
support successful memory formation.

3.2 Introduction

To investigate the neural mechanisms of successful memory formation, scientists

compare brain activity measured during the encoding of subsequently remembered
and subsequently forgotten items. Neuro-imaging studies investigating these sub-
sequent memory e↵ects (SMEs) have revealed that increased activation in a network
of temporal and prefrontal cortical regions predict subsequent memory (Wagner
et al., 1998; Kim, 2011; Burke, Long, et al., 2014). However, these changes in ac-
tivation might be due to any number of processes, including increased attention,
use of elaborative strategies, or the formation of item-to-context associations. For-
getting an item could be caused by failures of any of these processes. Therefore, to

33
establish whether the SME is driven by memory-specific processes or a mnemonic
attentional signal, it is necessary to use a more fine-grained contrast comparing
items that vary in how they are remembered, not whether they are remembered.
In a free recall task, items that are e↵ectively encoded in relation to their context
exhibit strong temporal clustering, being recalled in close proximity to their study-
list neighbors (Kahana, 1996). By comparing brain activity during the encoding
of items that are subsequently clustered with those that are not clustered (defined
here as the subsequent clustering e↵ect, SCE) we can isolate the neural correlates
of e↵ective item-to-context associative memory encoding. To identify the memory-
specific neural mechanisms supporting memory formation, we compared the SCE
and SME. We hypothesized two potential outcomes. First, the SCE may be a
component of the SME. Activation in the SME might be driven by items that are
subsequently clustered, a prediction supported by behavioral evidence showing
that increased clustering correlates with high recall success (Sederberg et al., 2010).
Alternatively, the SCE and SME may be independent and while clustering may
correlate with probability of recall, both may be moderated by a third unknown
variable (Brown et al., 1991). In this case, the SME might instead be driven by
attentional mechanisms.
We analyzed intracranial electroencephalographic (iEEG) data from neurosur-
gical patients participating in a free recall task. The recorded iEEG signals simul-
taneously sample local field potentials throughout the brain, and can be analyzed
in terms of specific time-varying oscillatory or spectral components of neural ac-
tivity. Using brain regions selected a priori based on previous subsequent memory
studies, we measured the spectral signals during encoding of words that were later
clustered; later recalled and not clustered; or later forgotten.
To foreshadow our results, we found that high frequency activity (HFA, 44-100

34
Hz) in the memory network tracked e↵ective contextual encoding with greater
HFA for subsequently clustered compared to subsequently recalled non-clustered
items. The timing of this e↵ect dissociated across regions: LIFG clustering related
increases in HFA preceded those in hippocampus. Together, these results suggest
that prefrontal context updating mechanisms which precede item-to-context bind-
ing processes in the hippocampus are reflected in the SME and drive successful
memory formation.

3.3 Materials and Methods

Participants

136 participants (58 female; age range: 8 - 57, mean = 33) with medication-resistant
epilepsy underwent a surgical procedure in which electrodes were implanted
subdurally on the cortical surface and deep within the brain parenchyma. In
each case, the clinical team determined electrode placement so as to best local-
ize epileptogenic regions. Data were collected as part of a long-term multicenter
study; data were collected at Boston Children’s Hospital, Hospital of the Univer-
sity of Pennsylvania, Freiburg University Hospital, and Thomas Je↵erson Uni-
versity Hospital. The institutional review board at each hospital approved the
research protocol. We obtained informed consent from the participants or their
guardians. Participants were left-hemispheric language dominant as assessed
by either the participants’ handedness or a clinically administered intracarotid
injection of sodium amobarbital (Wada test). Clinical need determined the elec-
trode placements and the total number of participants contributing to each re-
gion of interest was ranged from 60 (left inferior frontal gyrus) to 86 (left inferior

35
temporal cortex). The raw, de-identified data as well as the associated codes
used in this study can be accessed at the Cognitive Electrophysiology Data Portal
(http://memory.psych.upenn.edu/Electrophysiological Data).

Intracranial recordings

iEEG data were recorded using a Bio-Logic, DeltaMed, Nicolet, GrassTelefactor, or

Nihon Kohden EEG system. Depending on the amplifier and the discretion of the
clinical team, the signals were sampled at 256, 400, 500, 512, 1000, 1024, or 2000
Hz. Signals were referenced to a common contact placed either intracranially or
on the scalp or mastoid process. Contact localization was accomplished by co-
registering the post-op CTs with the MRIs using FSL Brain Extraction Tool (BET)
and FLIRT software packages. Contact locations were then mapped to both MNI
and Talairach space using an indirect stereotactic technique. Depth electrodes
were manually localized by a neuroradiologist experienced in neuroanatomical
localization utilizing post-operative MRIs and CT images in order to accurately
identify all depth contacts located within the hippocampus. For each participant
and electrode, the raw EEG signal was downsampled to 200 Hz and a fourth order
2 Hz stopband butterworth notch filter was applied at 50 or 60 Hz to eliminate
electrical line noise.

Free recall task

Participants studied lists of 15 or 20 high-frequency nouns for a delayed free recall

task. The computer displayed each word for 1600 ms, followed by an 800 to 1200
ms blank interstimulus interval. Immediately following the final word in each list,
participants were given a series of arithmetic problems of the form A+B+C=??,

36
where A, B and C were randomly chosen integers ranging from 1-9. This distractor
interval lasted at least 20 s, but participants were allowed to complete any problem
that they started resulting an average retention interval of 25 s. After the distractor,
participants had 45 s to freely recall as many words as possible from the list in any
order. Vocalizations were digitally recorded and later manually scored for analysis.
On average, participants participated in two sessions.

Data analyses and spectral power

Two concerns when analyzing bivariate interactions between closely spaced in-
tracranial contacts are volume conduction and confounding interactions with the
reference line. We used bipolar referencing to eliminate such confounds when an-
alyzing the neural signal (Nunez & Srinivasan, 2006). We found the di↵erence in
voltage between pairs of immediately adjacent electrodes (Burke et al., 2013). The
resulting bipolar signals were treated as new virtual electrodes and are referred to
as such throughout the text. Analog pulses synchronized the electrophysiological
recordings with behavioral events.
We applied the Morlet wavelet transform (wave number 6) to all bipolar elec-
trode EEG signals from 300 ms preceding to 1600 ms following word presentation,
across 46 logarithmically spaces frequencies (2-100 Hz). We included a 1000 ms
bu↵er on both sides of the data to minimize edge e↵ects. After log transforming
the power, we downsampled the data by taking a moving average across 100 ms
time windows and sliding the window every 50 ms, resulting in 31 time intervals
(16 non-overlapping) from -300 ms to 1600 ms surrounding stimulus presentation.
Power values were then Z-transformed within session by subtracting the mean and
dividing by the standard deviation power. Mean and standard deviation power
was calculated across all encoding events and time points in a session for each

37
frequency. We split the Z-transformed power into six distinct frequency bands (✓L ,
3-4 Hz; ✓H , 6-8 Hz; ↵, 10-14 Hz; , 16-26 Hz; L, 28-42 Hz; H, 44-100 Hz; Sederberg
et al., 2006), by taking the mean of the Z-transformed power in each frequency
band.
Our conditions of interest were subsequently recalled clustered items, study
items recalled either preceding or following the recall of a study neighbor (absolute
lag between serial position of items was 1), subsequent recalled non-clustered
items, study items recalled preceding and following the recall of a non-neighboring
study item (absolute lag between serial position of items was 2 or greater), and
subsequently not recalled items. Across participants there were on average 44
clustered items, 56 non-clustered items, and 358 not recalled items. A participant
had to have a minimum of 5 items per condition to be included in the analysis.
Our two contrasts were between subsequently recalled and forgotten items
and between subsequently clustered and non-clustered items. For each contrast
of interest and for each participant, electrode and frequency band, we calculated
Z-transformed power in each of two conditions. We averaged Z-power values
across electrodes within a region of interest (ROI) as we were interested in e↵ects
consistent across an ROI and not regional di↵erences within an ROI. Therefore,
each participant contributed a single Z-power value for each of two conditions
for each ROI. Conditions were compared across participants within an ROI and
frequency using a paired t-test.

ROI selection and analysis

The three ROIs were derived from several recent large scale studies suggesting that
these are core regions in the memory network (Kim, 2011; Burke, Long, et al., 2014;
Long et al., 2014). We defined ROIs using Brodmann area or neuroradiological

38
localization and included left inferior frontal gyrus (LIFG, BA 45/47, N = 60), left
inferior temporal lobe (LIT, BA 20/21, N = 86) and hippocampus (N = 64).

Peak time analysis

Using previous methods (Burke, Long, et al., 2014), we analyzed the temporal
specificity of the subsequent clustering e↵ect. Within each ROI and each par-
ticipant, we found the time point of the maximum (peak) di↵erence in Z-power
between subsequently clustered and non-clustered items. As the resulting distri-
butions of peak times was not normally distributed (see Results), we compared
peak times across ROIs using a non-parametric Wilcoxon rank sum test.

3.4 Results

Before examining the spectral components of the subsequent memory and cluster-
ing e↵ects, we report the basic behavioral data. Participants recalled on average
23% of studied items (SD = 10%). Participants were more likely to make recall
transitions between neighboring study items than between non-neighboring study
items (Figure 3.1A), replicating the lag contiguity e↵ect (Kahana, 1996). We quan-
tified this tendency to cluster with a temporal di↵erence score: the probability of
making a transition of absolute lag of 1 minus the average probability of making
a transition of absolute lag of 3 through 5. The average temporal di↵erence score
.13 was significantly greater than zero (t(135) = 14.0, p < .01) and was positively
correlated with probability of recall (r = .19, p = .03, Figure 3.1B), replicating pre-
vious findings (Sederberg et al., 2010). These results show that iEEG patients, like
healthy controls, cluster their recalls and that the tendency to cluster is positively
related to overall performance.

39
 We character-
ized the spectral A. Contiguity Effect B. Contiguity and recall

Conditional Response Probability

0.3 0.75 r = 0.191
R
p = 0.026
components of the

Recall Probability
0.2 0.5
subsequent mem-
ory and cluster- 0.1 0.25

ing e↵ects by com- 0

-5 -1 1 5
0
-0.4 0 0.2 0.6
Lag Temporal difference score
paring Z-power across
items subsequently Figure 3.1: Behavioral measures of recall performance. (A) Conditional
response probability (CRP) as a function of lag (the di↵erence in serial po-
recalled and for- sition of two studied items). The figure shows the likelihood of making a
transition from item i to item i +/- lag. Participants are more likely to transi-
gotten (SME) or tion to nearby (small absolute lag) items than to distant (large absolute lag)
items. Errorbars are standard errors of the mean. (B) Recall probability, the
items subsequently overall proportion of items recalled, is positively correlated with temporal
di↵erence score, the di↵erence in likelihood of making nearby (+/- 1 lag)
clustered and not relative to distant (+/- 3 to 5 lag) transitions (each point is a participant).
clustered (SCE). Clus-
tered items were recalled either preceding or following the recall of a study neigh-
bor. Non-clustered items were recalled preceding and following the recall of a
non-neighboring study item. We ran a paired t-test comparing Z-power across the
encoding interval (0-1600 ms) in six frequency bands (✓L , 3-4 Hz; ✓H , 6-8 Hz; ↵,
10-14 Hz; , 16-26 Hz; L, 28-42 Hz; H, 44-100 Hz) for each contrast.
The subsequent memory analysis revealed high frequency activity (HFA, 44 -
100 Hz) increases and low frequency activity decreases in left inferior frontal gyrus
(LIFG), left inferior temporal lobe (LIT), and hippocampus (ts > 2.0, ps < .05; Figure
3.2A), replicating previous results (Burke, Long, et al., 2014; Long et al., 2014). The
subsequent clustering analysis revealed HFA increases in LIFG (t(54) = 2.0, p =
.053), LIT (t(79) = 2.2, p = .03), and hippocampus (t(60) = 2.3, p = .02; Figure 3.2B).
We have shown that subsequent clustering is characterized by HFA increases
in the SME network. Although we hypothesize that these signals reflect contextual

40
encoding, they could instead be driven by a mnemonic attentional signal that is not
memory specific. Such an attentional signal would likely be present throughout
the encoding period, including the pre-stimulus interval (Osipova et al., 2006).
In contrast, a memory specific signal should be restricted to the post-stimulus
interval, especially if this signal reflects the formation of a bound item-to-context
representation. Therefore, we tested the post-stimulus specificity of the SCE in
our ROIs. Using a paired t-test, we compared pre-stimulus (-300 to 0 ms) to
post-stimulus (0 to 1600 ms) HFA di↵erences between subsequently clustered and
non-clustered items. Pre- and post-stimulus intervals did not di↵er in LIFG (t(54)
= .68, p = .50) or LIT (t(79) = .38, p = .70), but the SCE was significantly greater in
the post- than pre-stimulus interval in hippocampus (t(60) = 3.1 p < .01).
We have identified an SCE characterized by HFA increases across our three
ROIs during the post-stimulus interval. Figure 3.3 illustrates the time course of
that e↵ect in each region. Because LIFG has been hypothesized to play a critical
role in retrieval and selection processes (Thompson-Schill et al., 1997, 2005; Gold
et al., 2006; Badre & Wagner, 2007; Blumenfeld & Ranganath, 2007), and because
these processes, which may update context (Polyn & Kahana, 2008), have to unfold
prior to hippocampally-mediated item-to-context binding in order to support sub-
sequent clustering, we hypothesized that clustering-related activity in LIFG should
precede activity in hippocampus.
To test this hypothesis, for each participant and ROI we identified the time
interval during which the SCE was maximal (see Methods). We observed that the
median peak interval in hippocampus (900 - 1000 ms) was reliably later than the
median peak interval in LIFG (600 - 700 ms; non-parametric Wilcoxon rank sum
test (z = -2.27, p = .02). There was no reliable di↵erence in peak time between
hippocampus and LIT (median peak time in LIT, 500 - 600 ms; z = 1.65, p = .10) or

41
between LIFG and LIT (z = -.66, p = .51).

A. Subsequent memory effect, Recalled > Not recalled

LIFG LIT Hippocampus
0.2 0.2 0.2
* * * * * * * * * * * * * * * * *
Z-Power difference

Z-Power difference

Z-Power difference
0 0 0

-0.2 -0.2 -0.2

θL θH α β γL γH θL θH α β γL γH θL θH α β γL γH
Frequency Frequency Frequency

B. Subsequent clustering effect, Clustered > Not clustered

LIFG LIT Hippocampus
0.2 0.2 0.2
* * * *
Z-Power difference

Z-Power difference

Z-Power difference
0 0 0

-0.2 -0.2 -0.2

θL θH α β γL γH θL θH α β γL γH θL θH α β γL γH
Frequency Frequency Frequency

Figure 3.2: Subsequent memory and clustering e↵ects. The figure shows aver-
age across participant Z-power di↵erence for six frequency bands (✓L , 3-4 Hz; ✓H ,
6-8 Hz; ↵, 10-14 Hz; , 16-26 Hz; L , 28-42 Hz; H , 44-100 Hz) and three regions
of interest (ROI; left inferior frontal gyrus, LIFG; left inferior temporal lobe, LIT,
and hippocampus). Z-power values were generated by comparing the frequency
band-specific spectral power of two conditions and averaging those di↵erence val-
ues across electrodes within an ROI. Errorbars are standard errors of the mean.
Asterisks denote power di↵erences that significantly di↵ered (p < .05) from zero. (A)
Subsequent memory e↵ect, comparison of subsequently remembered and forgotten
items. (B) Subsequent clustering e↵ect, comparison of subsequently clustered and
non-clustered items, where items recalled preceding or following a study neighbor
(+/- 1 lag) are considered clustered and all others are not clustered.

We have shown that the peak LIFG SCE precedes the peak hippocampal SCE,
supporting the hypothesis that hippocampus drives item-to-context binding. If
hippocampal HFA is indicative of such a process, then the amount of late HFA
in the hippocampus should directly relate to behavioral measures of clustering.
We selected the modal hippocampal peak time (1300 - 1400 ms, Figure 3.3) as our
interval of interest. For each participant, we extracted hippocampal HFA during

42
this interval across all encoding trials and correlated the average signal with each
participant’s temporal di↵erence score. We found a significant positive correlation
(r = .38, p = .002) whereby increased HFA in the hippocampus was associated
with increased temporal clustering (Figure 3.4). There was no correlation with the
median peak time (r = .03, p = .83).

LIFG LIT Hippocampus

0.5 0.5 0.5
γH Z-power

0 0 0 C
NC
NR
-0.5 -0.5 -0.5
Peak SCE

-300 0 500 1000 1500 -300 0 500 1000 1500 -300 0 500 1000 1500
Time (ms) Time (ms) Time (ms)

Figure 3.3: Clustering e↵ects across time. The figure shows average high gamma Z-
power across the -300 to 1600 ms presentation interval where 0 is when the word is
presented on screen, separately for each ROI. Z-power values are plotted separately for
clustered (black), not clustered (dark gray) and not recalled (light gray) conditions. Error
bars are standard error of the mean. Below each time course is a box-and-whisker plot
showing the interquartile range of peak SCE time intervals. For each participant, we
found the time interval during which the HFA di↵erence between subsequently clustered
and subsequently non-clustered items was maximal. The solid black circle is the mode
peak time interval and the white circle is the median peak time interval. For LIT, the
median and mode peak time intervals were the same.

3.5 Discussion

The goal of the current study was to measure the neural correlates of memory
encoding as they relate to subsequent clustering. Our study demonstrates three key
findings. First, there is a subsequent clustering e↵ect (SCE) whereby high frequency
activity (HFA) in left inferior frontal gyrus (LIFG), left inferior temporal lobe (LIT),
and hippocampus is greater for subsequently clustered compared to subsequently
recalled, but not clustered items. Second, clustering related HFA increases in LIFG
precede those hippocampus, and the hippocampal SCE is specific to the post-

43
stimulus interval. Finally, HFA in the hippocampus during the late interval (1300
- 1400 ms) across all encoding trials positively correlates with behavioral measures
of clustering. Together, these results suggest that the core memory network is not
simply driven by a mnemonic attentional signal and that item-to-context binding
in the hippocampus supports successful memory formation.
We found in-
creased HFA across
0.25 rr = 0.384
the core memory
p = 0.002
Temporal difference score

network (Kim,
2011) for the en-
coding of items 0
subsequently clus-
tered compared to
items subsequently -0.25
recalled but not -0.1 0.4
Late hipp. γH Power
clustered. Though
classic memory anal-Figure 3.4: Relation between hippocampal activation and temporal dif-
ference score. For each participant, we calculated the average high gamma
yses (Wagner et Z-power across all encoding trials specifically for the late, 1300 - 1400 ms
post-stimulus interval in hippocampus, as this interval was the modal peak
al., 1998; Kim, time across participants (see Results and Figure 3). We found a significant
positive correlation between late hippocampal activation and temporal
2011; Burke, Long, di↵erence score (each point is a participant).
et al., 2014) have
shown region-specific activation increases by comparing subsequently remem-
bered and forgotten items, this activation need not be a memory signal per se.
By directly comparing items based on how they were recalled, instead of whether
they were recalled, we were able to relate the SME to a memory-specific mecha-
nism of contextual encoding. Although there is a wealth of evidence showing that

44
hippocampus is involved in both associative (Davachi & Wagner, 2002; Kirwan &
Stark, 2004; Mayes et al., 2007; K. J. Mitchell & Johnson, 2009) and order memory
(Jenkins & Ranganath, 2010; Ezzyat & Davachi, 2014; Davachi & DuBrow, 2015),
previous studies have assessed encoding di↵erences between items by explicitly
manipulating task demands at either encoding or retrieval. An open question
is whether the same mechanisms support memory formation when there are no
requirements to encode or retrieve associations, as such demands could induce ex-
plicit strategy use which might obscure other processes (Carr et al., 2010). Without
using an associative encoding or retrieval task, our results show that HFA in the
SME network increases with e↵ective contextual encoding, suggesting that hip-
pocampus may readily associate items and contexts even in the absence of explicit
task demands (Eichenbaum, 2004).
We found evidence that clustering related HFA increases in LIFG precede
those in hippocampus. Substantial evidence has shown that LIFG engages in
controlled retrieval, semantic elaboration and selection processes (Demb et al.,
1995; Thompson-Schill et al., 1999; Badre & Wagner, 2007; Martin, 2007). Fur-
thermore, communication between prefrontal and medial temporal lobe cortex
has been shown to support memory formation (Dickerson et al., 2007; Preston
& Eichenbaum, 2013). In our study, an encoding item’s pre-existing associations
could be retrieved by LIFG and integrated with the current context representation.
This context representation would then be bound to the current encoding item by
the hippocampus (Polyn & Kahana, 2008). Although we found early LIFG HFA
increases, previous work (Burke, Long, et al., 2014) has suggested that late HFA
increases in LIFG might retrieve or select stimulus-relevant information in the ser-
vice of item memory (Blumenfeld & Ranganath, 2007; Kim, 2011). Therefore, the
relative time course of LIFG and hippocampus may dictate whether or not an item

45
is bound to its spatiotemporal context.
Finally, late hippocampal HFA is associated with the degree to which a partici-
pant will consecutively recall study neighbors during test. If late HFA in hippocam-
pus is indicative of item-to-context binding, as we inferred from the time course
of the SCE, then the amount of late hippocampal HFA should directly relate to
participants’ tendency to cluster responses during recall. We found that increases
in late (1300 - 1400 ms) hippocampal HFA are positively correlated with temporal
di↵erence scores. This finding, coupled with the temporal dissociation of SCE
across regions, suggests that hippocampus engages in item-to-context associative
processes during encoding.
Two major questions not addressed by the present work include the role of theta
in contextual encoding and how HFA is modulated during contextual retrieval.
One may have made the a priori prediction that theta (3 - 8 Hz) would show a
reliable clustering e↵ect. Theta may signal the on-line state of the hippocampus
(Buzsáki, 2002), the theta phase relates to long term potentiation and may provide
a temporal context for events (Buzsáki, 2005; Hasselmo & Stern, 2014), and theta
has been implicated in tasks which manipulate context (Summerfield & Mangels,
2005; Staudigl & Hanslmayr, 2013). Theta e↵ects in the current study may be
obscured by broad asynchronous power fluctuations. The most consistent pattern
associated with the SME is increased HFA coupled with decreased low frequency
activity (Burke, Long, et al., 2014; Long et al., 2014); a somewhat similar pattern
is present in the SCE as well (see Figure 3.2). Such a pattern may be indicative of
a “general activation” mechanism as it is observed outside of the memory domain
(Crone et al., 2001; K. J. Miller et al., 2007; Jerbi et al., 2009) and may obscure
narrowband theta signals. A promising future direction will be to investigate not
only narrowband e↵ects, but to also assess the role of theta phase in contextual

46
encoding (Canolty et al., 2006; Axmacher et al., 2006; Nyhus & Curran, 2010;
Rutishauser et al., 2010; Lega et al., in press).
Additionally, the current study focused on contextual processing as an indicator
of e↵ective encoding based on predictions of context models (Howard & Kahana,
2002a; Polyn et al., 2009a). Context models also posit that context is reinstated
during retrieval, a prediction supported by recent neuro-imaging work showing
evidence for content and context reinstatement (Polyn et al., 2005; Manns et al.,
2007; Manning et al., 2011, 2012; Morton et al., 2013; J. F. Miller, Neufang, et
al., 2013; Ya↵e et al., 2014). A critical open question is how the univariate HFA
increases observed here manifest during retrieval and how HFA increases interact
with multivariate representations of items and context. One prediction is that HFA
may increase prior to clustering at retrieval, and that HFA signals may correlated
with the amount of similarity between encoding and retrieval patterns.

3.5.1 Conclusion

We have demonstrated that HFA increases as a function of e↵ective contextual

encoding, whereby subsequently clustered items show the greatest HFA. Addi-
tionally, we have shown that the SCE dissociates across time and regions with
the SCE in LIFG preceding the SCE in hippocampus. This result supports the
hypothesis that prefrontal cortex retrieves or selects contextual information and
that the hippocampus associates items with this contextual representation. Finally,
we have shown that late hippocampal HFA correlates with participants’ tendency
to consecutively recall study neighbors. Together, these results suggest that HFA
increases observed in the SME network are likely the result of contextual encoding
and that this contextual encoding directly supports successful memory formation.

47
Chapter 4

Modulation of task demands suggests

that semantic processing interferes
with the formation of episodic
associations

Nicole M. Long & Michael J. Kahana

4.1 Abstract

Although episodic and semantic memory share overlapping neural mechanisms,

it remains unclear how our pre-existing semantic associations modulate the forma-
tion of new, episodic associations. When freely recalling recently studied words,
people rely on both episodic and semantic associations, shown through temporal
and semantic clustering of responses. We asked whether orienting participants
toward semantic associations interferes with or facilitates the formation of episodic

48
associations. We compared electroencephalographic (EEG) activity recorded dur-
ing the encoding of subsequently recalled words that were either temporally or
semantically clustered. Participants studied words with or without a concurrent
semantic orienting task. We identified a neural signature of successful episodic
association formation whereby high frequency EEG activity (HFA, 44 - 100 Hz)
overlying left prefrontal regions increased for subsequently temporally clustered
words, but only for those words studied without a concurrent semantic orienting
task. To confirm that this disruption in the formation of episodic associations was
driven by increased semantic processing, we measured the neural correlates of
subsequent semantic clustering. We found that HFA increased for subsequently
semantically clustered words only for lists with a concurrent semantic orienting
task. This dissociation suggests that increased semantic processing of studied items
interferes with the neural processes that support the formation of novel episodic
associations.

4.2 Introduction

Episodic memory, the memory for contextually rich personal experiences, is typ-
ically distinguished from semantic memory, the memory for general knowledge
and facts (Tulving, 1972). Despite this qualitative distinction, it is clear that seman-
tic and episodic memory interact (Tulving, 1983; Squire & Zola, 1998; McClelland
& Rogers, 2003), given that episodic retrieval leads to activation of the same neural
substrates which support semantic memory (Martin et al., 1995; Polyn et al., 2005;
Martin, 2007; Patterson et al., 2007; Binder & Desai, 2011; Rissman & Wagner,
2012). Although these lines of evidence show a broad interaction between seman-
tic and episodic memory systems, it remains unclear how pre-existing semantic

49
knowledge impacts the formation of new, episodic associations.
The goal of the current study was to measure the influence of semantic associa-
tions during the encoding of items in a free recall task (Figure 4.1A). In free recall,
participants study a list of items and, either immediately or after a brief delay,
must recall those items in any order. Participants often consecutively recall study
neighbors, a phenomena known as temporal clustering (Kahana, 1996; Sederberg
et al., 2010), which may be the result of an association formed between a study item
and a slowly updating context representation, as posited by retrieved context the-
ory (Howard & Kahana, 2002a; Sederberg et al., 2008; Polyn et al., 2009a; Lohnas
et al., in press). Semantic processing could either facilitate or interfere with the
formation of these episodic associations. Semantic processing could be facilitative
by creating a richer context representation, in line with the levels of processing
theory (Craik & Tulving, 1975). In this view, any information, including semantic,
which creates a more elaborate memory trace will provide a better retrieval cue.
Behavioral evidence has shown enhanced priming for items that are both semanti-
cally and episodically related (McKoon & Ratcli↵, 1979) and increased probability
of transitioning between these items during free recall (Howard & Kahana, 2002b).
Alternatively, if semantic and episodic associations compete for position in the con-
text representation, increased semantic processing will interfere with the formation
of episodic associations. The context maintenance and retrieval model of Polyn et
al., (2009a) inversely weights the input of semantic and episodic associations in
context, which predicts that each will interfere with the other.
We measured the spectral correlates of subsequent temporal clustering, a proxy
of episodic association formation, using scalp electroencephalographic (EEG) record-
ings. These EEG signals can be analyzed in terms of specific time-varying oscilla-
tory or spectral components of neural activity. High frequency activity (HFA, 44 -

50
 A. Methods Temporally clustered (Ct)
TREE Semantically clustered (Cs)
LIME
CAT Not clustered (NC)
BALL
PIE Not recalled (NR)
Enco
ding DOG
Perio
d DOG
CAT
PIE
TREE
LIME
Rec
all P
erio
CAT d
LIME

0 ms 3000 ms 3800 ms + JITTER

Time
B. Behavioral Results C. Regions of
No-task Task Interest
0.5 0.6 * n.s.
Anterior
Response Probability

0.25

Clustering Score Superior

Conditional

0.25 0.3 0.12

Anterior
0 0 0 Inferior Posterior
-5 -1 1 5 0.015 0.435 0.855 Inferior
Lag Semantic similarity Temporal Semantic

Figure 4.1: Methods and behavioral results. (A) Methods. During the encoding period, partic-
ipants viewed words presented for 3000 ms and separated by a variable interstimulus interval.
Following the last item on the list, participants recall the study items in any order. A subset of
study items were semantic associates, e.g. cat and dog in this figure. Semantic relatedness was de-
termined using Word Association Space (WAS) values (see Methods). Encoding items were divided
into four conditions based on how they were recalled: temporally clustered (Ct , black) or recalled
preceding or following a study neighbor, e.g. tree and lime; semantically clustered (Cs , dark grey)
or recalled preceding or following a semantic associate; not clustered (NC, light grey) or recalled
preceding and following non-neighboring and non-semantically related items; or not recalled (NR,
white). (B) Behavioral results. Participants show a tendency to both temporally and semantically
cluster their recalls. The lag contiguity analysis (first panel) shows that participants are more likely
to make transitions between study neighbors, those items separated by a lag of +/- 1, than between
non-neighboring study items. Likewise, the semantic contiguity analysis (second panel) shows
that participants are more likely to make transitions between semantically associated items, where
increased association corresponds to increased WAS values. These e↵ects are consistent for both
no-task (blue) and task (orange) lists. The third panel shows the quantification of these contiguity
e↵ects. Temporal clustering scores are reliably greater on no-task compared to task lists, whereas
semantic clustering scores do not vary as a function of list type. (C) Regions of interest. We
analyzed three a priori defined ROIs, left Anterior Superior (AS), left Anterior Inferior (AI) and left
Posterior Inferior (PI).

51
100 Hz) in particular may reflect general cortical activation (Manning et al., 2009;
Jacobs & Kahana, 2009; Burke et al., 2015) and shows reliable increases during suc-
cessful memory formation (Sederberg et al., 2003; Long et al., 2014; Burke, Long, et
al., 2014).
We sought to assess the impact of semantic processing on the formation of
episodic associations by measuring these spectral signals. To manipulate semantic
processing, we included lists with and without a semantic orienting task (hence-
forth task and no-task lists). We predict that HFA will increase during the encoding
of subsequently temporally clustered items and that this HFA increase will vary
as a function of orienting task. If semantic processing is facilitative, these e↵ects
should be greater for task lists than no-task lists, as the semantic processing on task
lists should enhance contextual encoding. If semantic processing interferes with
episodic processing, subsequent temporal clustering e↵ects should be greater for
no-task lists, as semantic processing will disrupt episodic processes specifically on
task lists.

4.3 Results

The goal of our study was to assess the influence of semantic processing on episodic
memory formation. We first measured the tendency of participants to temporally
and semantically cluster their recalls (Figure 4.1B) separately for task and no-
task lists. Di↵erence in serial position, or lag, determined temporal relatedness.
Word Association Space (WAS) values (Steyvers et al., 2004), determined semantic
relatedness. Participants recalled on average 66% (SD = 13%) of studied items on
no-task lists and 58% (SD = 11%) of studied items on task lists. Participants were
more likely to make recall transitions between neighboring than distal study items

52
(Figure 4.1B, first panel) and were more likely to make recall transitions between
semantically related study items than non-semantically related study items (Figure
4.1B, second panel). We quantified the tendency to cluster study neighbors with
a temporal clustering (tc) score: the probability of making a transition of absolute
lag of 1 minus the average probability of making a transition of absolute lag of 3
through 5 (Kahana, 1996). Likewise, we quantified the tendency to cluster semantic
associates with a semantic clustering (sc) score: the probability of making a transition
to another item with a WAS value of .4 or greater minus the average probability
of making a transition to an item with a WAS value less than .4. Across both list
types, tc and sc scores were reliably greater than zero (Figure 4.1B, third panel, ts >
13.0, ps < .001). tc scores reliably di↵ered between no-task and task lists (t(151) =
6.3, p < .001), but sc scores did not (t(151) = 1.0, p = .32).
The behavioral results suggest that the semantic orienting task might interfere
with episodic encoding as temporal clustering was reliably decreased in task lists.
However, clustering is likely the result of both encoding and retrieval processes,
thus the decrease in temporal clustering might be due to interference at retrieval.
As behavioral measures alone cannot dissociate these two processes and our goal
was to measure how semantic processing impacts episodic association formation,
we turn to the neural data to measure the processes at encoding. We extracted
spectral signals across six frequency bands, low theta (3-4 Hz), high theta (6-8 Hz),
alpha (10-14 Hz), beta (16-26 Hz), low gamma (28-42 Hz) and high gamma (44-100
Hz) and three regions of interest (ROIs), left Anterior Superior (AS), left Anterior
Inferior (AI) and left Posterior Inferior (PI; Figure 4.1C). We made comparisons
across items subsequently temporally clustered, but not semantically clustered
(Ct ), items subsequently semantically clustered, but not temporally clustered (Cs ),
and items subsequently recalled, but not clustered (NC) separately for no-task and

53
task lists. Ct items were recalled preceding or following a study neighbor, Cs items
were recalled preceding or following a semantic associate, and NC items were
recalled preceding and following non-neighboring and non-related study items.
Semantic associates were pairs of study list items with WAS values of .4 or greater.
Our first goal was to identify a neural subsequent temporal clustering e↵ect.
In a previous study (Long & Kahana, submitted) we found that high frequency
activity (HFA, 44 - 100Hz) increases as a function of subsequent temporal cluster-
ing; however, as we did not control for semantic relatedness among study words,
temporally clustered words could also have been clustered based on semantic as-
sociations. Therefore, we were unable to dissociate the contributions of episodic
and semantic processing. To measure the signals unique to episodic association
formation, we analyzed the subsequent temporal clustering e↵ect (SCEt ) by com-
paring Ct items to NC items on no-task lists. All semantically clustered items were
excluded. This SCEt analysis revealed several power increases across ROIs (Figure
4.2, top row, blue line). Beta power increases were significant in all ROIs (AS, t(151)
= 3.0, p < .01; AI, t(151) = 2.8, p< .01; PI, t(151) = 2.3,p = .02). Low and high gamma
power increases were significant in AS (low gamma, t(151) = 2.7, p < .01; high
gamma, t(151) = 2.0, p= .04). These results suggest that the formation of episodic
associations is characterized by increases in HFA over left prefrontal regions.
Next, we sought to test the hypothesis that orienting participants to semantic
associations could facilitate or interfere with the formation of episodic associations.
If semantic processing facilitates episodic memory formation, we should observe
an increased SCEt on task relative to no-task lists. Semantic processing may in-
crease contextual encoding by creating a more elaborative encoding context (Craik
& Tulving, 1975). Alternatively, if semantic processing interferes with episodic
memory formation, we should observe a decreased SCEt on task relative to no-task

54
lists. The task may decrease contextual encoding by directing processing toward
semantic associations at the expense of episodic associations. We tested these com-
peting hypotheses by comparing Ct and NC items on task lists. This SCEt analysis
revealed no consistent e↵ects in our ROIs (Figure 4.2, top row, orange line) and
specifically no significant HFA e↵ects (ts < 1.5, ps > .10).
Our results sug-
gest that semantic No-task
Task
processing inter-
feres with episodic 0.15
SCEt ∆Z-Power

encoding mecha-
0

nisms. How-
ever, this inter- -0.15
θL θH α β γL γH θL θH α β γL γH θL θH α β γL γH

pretation assumes
0.15

that the seman-

SCEs ∆Z-Power

tic orienting task 0

fully focuses pro-

-0.15
cessing on seman- θL θH α β γL γH θL θH α β γL γH θL θH α β γL γH

tic features. Al-

though there is a Figure 4.2: Subsequent clustering e↵ects. The top panel shows the sub-
sequent temporal clustering e↵ect (SCEt ) and the bottom panel shows the
decrease in tempo- subsequent semantic clustering e↵ect (SCEs ) for no-task (blue) and task
(orange) lists. Each line shows the di↵erence in Z-Power between subse-
ral clustering for quently clustered and subsequently recalled, but not clustered, items, for 6
frequency bands (✓L , 3-4 Hz; ✓H , 6-8 Hz; ↵, 10-14 Hz; , 16-26 Hz; L , 28-42
task lists, there is Hz; H , 44-100 Hz). Z-Power is averaged across the encoding interval (0 -
3000 ms). Errorbars are standard error of the mean.
not concurrent in-
crease in semantic clustering. It is thus unclear whether the task increases semantic
processing. If the semantic orienting task directs resources away from episodic
association formation in favor of semantic processing, there should be a neural

55
subsequent semantic clustering e↵ect for task lists. To test this hypothesis, we an-
alyzed the subsequent semantic clustering e↵ect (SCEs ) by comparing Cs and NC
items exclusively for task lists. All temporally clustered items were excluded. This
SCEs analysis revealed broad increases in power across both anterior ROIs (Figure
4.2, bottom row, orange line). High theta increases were significant in AS (t(148)
= 2.1, p = .03) and beta increases were significant in AI (t(148) = 2.0, p= .04). Low
and high gamma increases were significant in both AS (low gamma, t(148) = 2.3, p
= .02; high gamma, t(148) = 2.7, p < .01) and AI (low gamma, t(148) = 2.3, p = .02;
high gamma, t(148) = 2.9, p < .01).
We have shown
that for task lists No-task
Task

the SCEt decreases

∆ High frequency activity

0.2 0.2 0.2

whereas the SCEs

increases, suggest- 0 0 0

ing that partici-

-0.2 -0.2
-0.2
SCEt SCEs SCEt SCEs SCEt SCEs
pants are process-
ing semantic fea- Figure 4.3: HFA as a function of clustering and list type. Subsequent
temporal and semantic clustering e↵ects (SCE and SCE ) separately for
t s
tures during the no-task (blue) and task (orange) lists. Each bar shows the di↵erence in high
frequency activity between subsequently clustered (temporal or semantic)
orienting task. How-and subsequently recalled, but not clustered words. There is a reliable
interaction in AS and AI such that the SCE is specific to no-task lists and
t
ever, participants the SCE is specific to task lists. Error bars are standard errors of the mean.
s

may process those

semantic features regardless of the orienting task. Indeed, the behavioral results
suggest that semantic clustering is matched for both no-task and task lists. If se-
mantic processing interferes with episodic encoding, a reliable SCEt should not
co-occur with a SCEs . To test this hypothesis, we measured the SCEs for no-task
lists. This SCEs analysis revealed no consistent e↵ects in our ROIs (Figure 4.2,

56
bottom row, blue line) and specifically no significant HFA e↵ects (ts < 2, ps > .05).
Although we see behavioral evidence for semantic clustering, this may be driven by
the use of semantic associations exclusively during retrieval. As our neural results
suggest that semantic processing interferes with episodic processing, we should
observe an interaction between the type of clustering e↵ect and the list type. We
compared HFA for the SCEt and SCEs for no-task and task lists (Figure 4.3) using a
2 ⇥ 2 repeated measures ANOVA and found a reliable clustering e↵ect type ⇥ list
type interaction in AS (F(1,143) = 14.5, p < .001) and AI (F(1,143) = 11.7, p < .001),
but not PI (F(1,143) = .85, p = .36).
The dissociation in clustering e↵ect type and list type suggests that seman-
tic processing interferes with episodic processing. As HFA increases for both of
these processes, it might reflect a control mechanism which biases retrieval of task-
relevant features. Although HFA increases are typically associated with successful
memory formation (Sederberg et al., 2003; Burke, Long, et al., 2014; Long et al.,
2014), an implication of such a control mechanism is that when task-relevant associ-
ations are incongruent with episodic memory demands, HFA should be negatively
related to memory performance. We can test this prediction by measuring the en-
coding activity of semantic isolated items, those items with no semantic neighbors
on the study list (pairs of study list items with WAS values < .4, see Methods).
Increased semantic processing of these items, reflected through increases in HFA,
should be detrimental to their subsequent recall, as no other semantic associate
will be available as a retrieval cue, and the necessary episodic associations will
not have been formed. We measured HFA during the encoding of semantically
isolated items which were later recalled or not recalled, for both task and no-task
lists. This analysis revealed a negative subsequent memory e↵ect exclusively for
task lists (Figure 4.4). We ran a 2 ⇥ 2 repeated measures ANOVA comparing recall

57
status (recalled or not recalled) and list type (no-task or task) to test the significance
of this result. We found a reliable recall status ⇥ list type interaction in AS (F(1,151)
= 6.4, p = .01), but not AI or PI (AI, F(1,151) = 1.8, p = .19; PI, F(1,151) = .79, p = .37).

4.4 Discussion

The goal of the current study was to measure the contribution of semantic associa-
tions to the formation of episodic associations. Our study demonstrates three key
findings. First, there is a subsequent temporal clustering e↵ect (SCEt ), character-
ized by increased high frequency activity (HFA, 44 - 100 Hz) over left prefrontal
(PFC) regions, specific to lists without a concurrent semantic orienting task (no-task
lists). Second, there is a subsequent semantic clustering e↵ect (SCEs ), characterized
by increased HFA over left PFC, specific to lists with a concurrent semantic encod-
ing task (task lists). Finally, during task lists, semantically isolated items show a
negative subsequent memory e↵ect (SME). That is, HFA over PFC was greater for
subsequently forgotten compared to subsequently recalled items, specifically for
items which were only weakly semantically related to other study list items. These
results suggest that orienting processing toward pre-existing semantic associations,
as in the task lists, interferes with the formation of new, episodic associations.
We found increased HFA over left PFC regions during the encoding of subse-
quently temporally clustered items, specifically for no-task lists. Retrieved context
theory posits that clustering is the result of items forming associations with a slowly
updating context representation (Howard & Kahana, 2002a; Polyn et al., 2009a;
Lohnas et al., in press). This context representation is a weighted sum of both pre-
experimental semantic associations and newly formed episodic associations. We
hypothesize that increased HFA over PFC reflects a cognitive control mechanism

58
which determines the relative weight of each of these associations. Specifically,
PFC may alternate between maintaining the previously studied items or retrieving
semantic associations. Substantial evidence has shown that PFC is critical for a va-
riety of control processes, including maintenance and manipulation of information
in working memory (Petrides, 2000; Howard et al., 2003; Hazy et al., 2006; Chatham
et al., 2014) as well as controlled retrieval and selection processes (Thompson-Schill
et al., 1998; Thompson-Schill, 2003; Bunge et al., 2005; Badre & Wagner, 2007; Rodd
et al., 2012). Our finding that the SCEt is specific to no-task lists suggests that HFA
increases on no-task lists might reflect maintenance of previous study list items
and that the orienting task interferes with this process.
HFA over
left PFC in- Recalled
Not recalled

creased during
0.2
High Frequency Activity

the encoding
of subsequently 0

semantically clus-
-0.2
No-task Task No-task Task No-task Task
tered items specif-
ically for task Figure 4.4: Subsequent memory e↵ect for semantically isolated items. Se-
mantically isolated items are those items which do not have a strong semantic
lists. This re- study associate (WAS > .4). The figure shows HFA for semantically isolated
items that are subsequently recalled (RS0 , black) or subsequently not recalled
sult confirmed (NRS0 , white), separately for no-task and task lists. There is a reliable interaction
in AS such that HFA is increased for NRS0 relative to RS0 items specifically for
that during en- task lists. Error bars are standard errors of the mean.
coding, the se-
mantic orienting task increased semantic processing relative to the no-task lists.
Furthermore, this result is consistent with the interference hypothesis, as we would
not expect to concurrently observe both an SCEs and an SCEt on no-task lists. That
the SCEs and SCEt both show increased HFA over left PFC suggests a control

59
mechanism which directs processing based on task demands. This interpretation
is consistent with previous work showing that left PFC can bias content-specific
processing in posterior regions (E. K. Miller & Cohen, 2001; Thompson-Schill, 2003;
Noppeney et al., 2006; Bedny et al., 2008; Kuhl et al., 2013).
The dissociation of subsequent clustering e↵ect and list type suggests that the
processing of semantic and episodic associations may be inversely related as pre-
dicted by retrieved context models (Polyn et al., 2009a). However, one might have
predicted that semantic associations would facilitate the formation of episodic asso-
ciations, as there is evidence that items related both episodically and semantically
show the most facilitative priming (McKoon & Ratcli↵, 1979). Our experiment
was ideally suited to test these competing hypotheses as the design intentionally
included semantically isolated items, items paired with a study list item of low
semantic similarity (WAS < .4). According to the facilitative hypothesis, increased
semantic processing should always benefit a given study item, whether or not that
item is semantically isolated. Alternatively, according to the interference hypothe-
sis, increased semantic processing of a semantically isolated item should diminish
its probability of being recalled. That is, semantic processing is not beneficial be-
cause no other items can cue retrieval of the semantic isolate. It is furthermore
detrimental because it prevents episodic association formation. Our results sup-
port the latter hypothesis. We found increased HFA for subsequently forgotten
relative to recalled semantically isolated items, specifically for task lists. Although
HFA increases typically predict successful memory formation (Sederberg et al.,
2003; Burke, Long, et al., 2014; Long et al., 2014), when task demands emphasize
stimulus features that will not be useful at retrieval, HFA increases could predict
memory failures (Blumenfeld & Ranganath, 2007). Thus, left PFC biased processing
towards task-relevant information that was both not beneficial for later memory

60
and impaired processing that was necessary for later memory. Together our re-
sults show that accessing semantic associations can interfere with the formation of
episodic associations.
Although the majority of our e↵ects occurred in the high frequency range,
we might have expected to observe both theta (3 - 8 Hz) and beta (16 - 26 Hz)
e↵ects. Evidence has shown that theta increases during order and contextual
processing (Hsieh et al., 2011; Staudigl & Hanslmayr, 2013). We did not find
theta e↵ects, outside of a single contrast in a single ROI. As appears to be the case
with HFA, variations in task demands, e.g. explicitly encoding item order, may
modulate theta e↵ects. Also, recent work has suggested that beta decreases reflect
semantic processing (Hanslmayr et al., 2009). We found beta increases for both
subsequent temporal and semantic clustering. This discrepancy could be the result
of di↵erences in encoding single items as opposed to pairs of items or associations
(Hanslmayr et al., 2012). It is also possible that the increases we observed reflect
broadband asynchronous high frequency e↵ects (Manning et al., 2009; Burke et al.,
2015) that obscure narrowband e↵ects.
Since the term “episodic memory” was coined by Endel Tulving, researchers
have endeavored to understand how the episodic and semantic memory systems in-
teract. A critical question has been how our pre-existing associations influence our
ability to form new episodic associations. By measuring the impact of a semantic
orienting judgment in an episodic task, we have provided evidence suggesting that
the processing of semantic associations interferes with the formation of episodic as-
sociations. An important future question to address is how semantic and episodic
processes interact during retrieval. Although we did not observe encoding activity
related to semantic clustering in no-task lists, there was still behavioral evidence for
semantic clustering. Unlike episodic information, semantic information does not

61
need to be accessed during encoding in order to be available at retrieval. Thus, there
could be an interaction between episodic and semantic processing during recall as
well, potentially explaining how the most likely recall transitions are to those items
that are both semantically and episodically related (Howard & Kahana, 2002b).
An additional critical next step will be to understand how our univariate results
relate to recent multivariate work showing that episodic and semantic informa-
tion is present in context representations during encoding and retrieval (Manning
et al., 2011, 2012). Univariate HFA increases should correlate with the degree to
which patterns reflect episodic vs. semantic associations and such an e↵ect should
be modulated by task demands, given our conclusion that processing semantic
associations can interfere with the formation of episodic associations.

4.5 Materials and Methods

Participants

152 (86 female) paid volunteers (ages 18 - 29) were recruited via fliers posted
around the University of Pennsylvania campus. Participants were provided with
a base monetary compensation plus an additional performance-based monetary
incentive to ensure full e↵ort. The Institutional Review Board at the University of
Pennsylvania approved our research protocol, and informed consent was obtained
from all participants.

Free recall task

The data reported in this manuscript were collected as part the Penn Electrophys-
iology of Encoding and Retrieval Study (PEERS), involving three multi-session

62
experiments that were sequentially administered. The data reported below are
from Experiment 1.
Participants performed an immediate free recall experiment consisting of seven
sessions of 16 lists of 16 words presented one at a time on a computer screen
(Figure 4.1). Each word was drawn from a pool of 1638 words taken from the
University of South Florida free association norms (Nelson et al., 2004, available at
http://memory.psych.upenn.edu/files/wordpools/PEERS wordpool.zip). Se-
mantic relatedness was determined using the word association space (WAS) model
(Steyvers et al., 2004). WAS similarity values were used to group words into
four similarity bins (high similarity, cos ✓ > 0.7; medium-high, 0.4 > cos ✓ < 0.7;
medium-low, 0.14 >cos ✓ < 0.4; low similarity, cos ✓ < 0.14). Two pairs of items
from each of the four groups were arranged such that one pair occurred at adjacent
serial positions and the other pair was separated by at least two other items. All
randomly generated word lists conformed to this structure. The same word was
not repeated in a session.
Words were presented concurrently either with a task cue, indicating the judg-
ment that the participant should make for that word, or with no encoding task. The
two encoding tasks were a size judgment (“Will this item fit into a shoebox?”) and
an animacy judgment (“Does this word refer to something living or not living?”),
and the current task was indicated by the color and typeface of the presented item.
There were four no-task (control) lists (participants did not have to perform judg-
ments with the presented items), six single-task lists (all items were presented with
the same task, three of each task), and six task-shift lists (items were presented
with either task). List and task order were counterbalanced across sessions and
participants.
For each list, there was a 1500 ms delay before the first word appeared on the

63
screen. Each item was on the screen for 3000 ms, followed by a jittered 800 to
1200 ms interstimulus interval (uniform distribution). After the last item in the list,
there was a 1200 to 1400 ms jittered delay, after which a tone sounded, a row of
asterisks appeared, and the participant was given 75 s to attempt to recall any of
the just-presented items.
To minimize recency e↵ects (Murdock, 1962) and to facilitate comparison with
our previous analyses (Long et al., 2014, Long & Kahana, submitted), we excluded
all recency items (serial positions 13 - 16), all task-shift lists, and the last three
sessions from our analyses.

Electrophysiological recordings and data processing

EEG measurements were recorded using Geodesic Sensor Nets (GSN; Netstation 4.3
acquisition environment, from Electrical Geodesics, Inc.). The GSN provided 129
standardized electrode placements across participants. All channels were digitized
at a sampling rate of 500 Hz, and the signal from the caps was amplified via either
the Net Amps 200 or 300 amplifier. Recordings were initially referenced to Cz
and later converted to an average reference. Channels that demonstrated high
impedance or poor contact with the scalp were excluded from the average reference.
For each participant and electrode, a fourth order 2 Hz stopband butterworth notch
filter was applied to the raw EEG signal at 60 Hz to eliminate electrical line noise.
To identify epochs contaminated with eyeblink and other movement artifacts,
electrooculogram (EOG) activity was monitored bipolarly using right and left elec-
trode pairs (electrodes 25, 127 and 8 and 126 on the GSN). An individual word
presentation event was rejected from subsequent analyses if the weighted running
average for either the right or the left EOG pair exceeded a 100 µV threshold.

64
Data analyses and spectral power

We applied the Morlet wavelet transform (wave number 6) to all electrode EEG
signals from 500 ms preceding to 3000 ms following word presentation, across
46 logarithmically spaces frequencies (2-100 Hz). We included a 1000 ms bu↵er
on both sides of the data to minimize edge e↵ects. After log transforming the
power, we downsampled the data by taking a moving average across 100 ms
time windows and sliding the window every 50 ms, resulting in 69 time intervals
(35 non-overlapping) from -500 ms to 3000 ms surrounding stimulus presentation.
Power values were then Z-transformed within session by subtracting the mean and
dividing by the standard deviation power. Mean and standard deviation power
were calculated across all encoding events and time points in a session for each
frequency. We split the Z-transformed power into six distinct frequency bands (✓L ,
3-4 Hz; ✓H , 6-8 Hz; ↵, 10-14 Hz; , 16-26 Hz; L, 28-42 Hz; H, 44-100 Hz; Sederberg
et al., 2006), by taking the mean of the Z-transformed power in each frequency
band and across the 0 to 3000 ms presentation interval.
We defined four conditions of interest (Figure 4.1), items subsequently recalled
and temporally clustered (Ct ), items subsequently recalled and semantically clus-
tered (Cs ), items subsequently recalled and not clustered (NC), and items subse-
quently not recalled (NR). Ct items were study items recalled either preceding or
following the recall of a study neighbor (absolute lag between serial position of
items was 1), but not recalled preceding or following a semantic associate. Cs items
were study items recalled either preceding or following the recall of a semantic
associate (WAS value for the pair of items was >= .4), but not recalled preceding
or following a study neighbor. NC items were study items recalled preceding and
following non-neighboring and non-semantically associated study items. We an-
alyzed these conditions separately for control lists (no encoding task performed)

65
and task lists (single encoding task, either animacy or size judgment, performed
for all items in a list).

ROI selection and analysis

Z-power values were averaged across electrodes within a region of interest (ROI) as
we were interested in e↵ects consistent across an ROI and not regional di↵erences
within an ROI. Therefore, each participant contributed a single Z-power value for
each condition for each ROI. Our three ROIs (Figure 4.1 were selected a priori based
on previous scalp EEG studies (Weidemann et al., 2009; Long et al., 2014) and
were intended to cover left prefrontal (Anterior Superior, AS; Anterior Inferior, AI)
and left temporal cortex (Posterior Inferior, PI). Conditions were compared across
participants within an ROI and frequency using a paired t-test.

66
Chapter 5

Recall dynamics reveal the retrieval of

emotional context

Nicole M. Long, Michelle S. Dano↵, & Michael J. Kahana

Psychonomic Bulletin and Review

5.1 Abstract

Memory is often better for emotional rather than neutral stimuli. The benefit for
emotional items could be the result of an associative mechanism whereby items are
associated to a slowly updating context. Through this process, emotional features
are integrated with context during study, and are reactivated during test. The pres-
ence of emotion in context would both provide a stronger retrieval cue, enhancing
memory of emotional items, as well as lead to emotional clustering, whereby emo-
tionally similar items are recalled consecutively. To measure whether associative
mechanisms can explain the enhancement for emotional items, we conducted a
free recall study in which most items were emotionally neutral to minimize e↵ects

67
of mood induction and to more closely reflect naturalistic settings. We found that
emotional items were significantly more likely to be recalled than neutral items and
that participants were more likely to transition between emotional items rather than
between emotional and neutral items. Together, these results suggest that contex-
tual encoding and retrieval mechanisms may drive the benefit for emotional items
both within and outside the laboratory.

5.2 Introduction

Both the emotional quality of an item and the emotional state of an individual can
impact how an item is remembered. Emotional items are typically better remem-
bered than neutral items (Dolcos et al., 2004; LaBar & Cabeza, 2006; Kuhbandner
& Pekrun, 2013) and emotional states can enhance memory for both neutral and
emotional items (Bower, 1981; Eich, 1995; Maratos & Rugg, 2001; Erk et al., 2003).
The associative network theory of emotion (Bower, 1981) suggests that emotional
information is associated with concurrently presented stimuli and is part of the
retrieval cue. If emotion enters into such associations, then the e↵ects of emotion
on memory should extend beyond the “macro” scale of experiment level mood
e↵ects and be present on the “micro” scale of individual items. Despite this, asso-
ciative network theory has not been extended to explain the benefit for individual
emotional items in neutral contexts. Additionally, there has been conflicting evi-
dence as to whether emotion enhances (Doerksen & Shimamura, 2001) or impairs
(Kensinger, 2009; Maddock & Frein, 2009) associative memory.
Enhanced memory for items presented during a specific mood may be driven
by associations between items and context (Bower, 1981; Lewis & Critchley, 2003).
The term context here describes a mental representation, which reflects both ex-

68
ternal stimuli, such as the color of the computer screen, the experiment room, and
the internal state of the individual, including his or her mood and associations
automatically retrieved in response to external stimuli. The inclusion of emotion in
context can provide a cue for retrieving neutral items associated with the same emo-
tional context during encoding (mood dependent memory e↵ect) or for retrieving
emotional items that match the current emotional context (mood congruent mem-
ory e↵ect). A natural extension of this mechanism is that individual emotional
items should impart their emotional features to context. This emotional informa-
tion should then be present in the retrieval cue and lead to enhanced memory for
emotional items.
Alternatively, the salience theory suggests that enhanced memory for emo-
tional items is the result of more resources dedicated to processing emotional
items. Neuro-imaging evidence has shown that emotional items recruit attentional
and motivational networks (Dolcos et al., 2004; Kissler et al., 2006; Liu et al., 2012),
potentially due to modulatory signals from the amygdala. This modulation hy-
pothesis suggests that the amygdala enhances processing of emotional stimuli in
other brain networks (Anderson & Phelps, 2001; Pessoa & Adolphs, 2010). Consis-
tent with this theory, increased amygdala activation correlates with better memory
for emotional, but not neutral, items (Canli et al., 2000).
Both salience and context theories derive support from experiments using many
emotional stimuli or an emotional orienting task. Although such a design provides
ample statistical power, it does not mirror natural encounters with emotional stim-
uli. Having many emotional stimuli or an emotional orienting task could induce
explicit strategy use or mood induction, both of which may interfere with typical
organizational tendencies.
The goal of our study was to investigate memory for emotional items in a

69
largely neutral context. To accomplish this, we used a largely neutral word pool
where only a small subset of words were emotional (see below). If with such
a design, memory is still enhanced for emotional relative to neutral items, this
enhancement could be explained by two alternative hypotheses. First, if emotional
items activate emotion features in context during study, and those features are
reactivated during retrieval, then recall of an emotional item should be followed
by consecutive recall of other emotional items, since they have similar contexts
as the result of their shared emotional features. Such an “emotional” clustering
e↵ect falls naturally out of retrieved context theories (Howard & Kahana, 2002a)
and is in line with other clustering phenomena, whereby individuals organize their
memories by semantic, episodic and spatial associations (Bousfield, 1953; Kahana,
1996; J. F. Miller, Lazarus, et al., 2013). Previous work has suggested that emotional
clustering only occurs when participants are explicitly oriented to the emotionality
of items (Siddiqui & Unsworth, 2011); however, the use of many emotional items
may have altered participants’ strategies and interfered with a context mechanism.
Alternatively, if emotional features are not represented in context, then there should
be no consecutive recall of emotional items. Instead, the benefit for emotional items
may be due to increased attentional processing which does not necessarily lead to
emotional clustering.
To assess the interaction of emotion and context, we measured memory success
for emotional and neutral items in a free recall task. In free recall, participants study
and recall a list of items with no external cues, and thus, must rely on contextual
information to generate cues. To quantify emotion, we conducted an independent
experiment in which a separate group of participants rated each word for emotional
valence. The words had originally been selected to be predominantly neutral
and there was no emotion-related orienting task. We measured both memory for

70
emotional and neutral items, as well as the tendency of participants to consecutively
recall items of similar valence.

5.3 Norming study

Methods

We conducted a norming experiment using Amazon’s Mechanical Turk (MTurk,

Mason & Suri, 2012). MTurk is a crowdsourcing website used for the mass collec-
tion of data via the internet. The Institutional Review Board at the University of
Pennsylvania approved our research protocol, and informed consent was obtained
from all participants. The norming study consisted of two parts: a qualifier test to
screen for high-performing participants, and the main study in which participants
normed the free recall study word pool.
247 participants completed the qualifier study in which they rated the emotional
valence of 100 words drawn from the A↵ective Norms for English Words (ANEW)
dataset (Bradley & Lang, 1999). None of these words appeared in the free recall
study. Participants were asked to rate words on an emotional sliding scale ranging
from unpleasant to pleasant. Though not shown to the participants, these ratings
corresponded to a numerical value from 1 to 9, with 1 being unpleasant and 9
being pleasant. The instructions and scale were identical to those used in the
ANEW study. Each of the 100 words was presented twice, and in random order, to
measure internal consistency. Participants were compensated $0.30 for completing
the qualifier survey.
Qualification for the main study was determined through internal consistency,
external consistency, and response distribution. Internal consistency was measured

71
using the correlation between an individual participant’s ratings of the first and
second presentations of the same word. External consistency was measured using
the correlation between a participant’s average rating of a word and the ANEW
ratings. To qualify, both internal and external values had to exceed .7 (Cicchetti,
1994). Additionally, participants were required utilize the full range of responses
(1-9). 173 participants qualified for the main study.
120 participants completed the main study in which they rated the emotional
valence of 1638 words from the free recall study (see below). The rating procedures
were identical to the qualifier task. Participants were compensated $15. Twenty
participants did not rate all of the words and were therefore excluded from further
analysis. Ratings were averaged across the 100 participants who completed the
study. Each word was then assigned a status of negative (ratings 1-4), neutral
(ratings 4-6), or positive (ratings 6-9) valence (Baran et al., 2012).

Results and Discussion

Of the 1638 words, 140 were negative (9%), 1073 were neutral (65%), and 425 were
positive (26%). There is evidence that word properties, including word frequency,
imageability, concreteness, and semantic relatedness, can influence memory per-
formance and vary as a function of valence (Glanzer & Adams, 1990; Talmi &
Moscovitch, 2004; Bennion et al., 2013). Therefore, we attempted to simultane-
ously control for these factors across our three valence categories. We matched
average word frequency between positive and neutral words, and negative and
neutral words. We calculated which word pool (either positive or neutral, or nega-
tive or neutral) had the higher average word frequency. We then removed from this
pool the word with the highest frequency. We repeated this procedure iteratively
until the word pools did not significantly di↵er in word frequency using an un-

72
paired t-test (the p value had to exceed 0.10 to be considered non-significant). With
this new word pool, we then controlled for imageability, concreteness and semantic
relatedness with the same procedure. Imageability and concreteness ratings were
only available for 984 of the 1638 words (Wilson, 1988). Semantic relatedness
was measured using average word association space (WAS) values (Steyvers et
al., 2004). The distributions of semantic relatedness values were compared across
word pools and were not reliably di↵erent (see Supplemental Information). It is
important to note that we matched each valence pool (negative, positive) to the
neutral pool. It was impossible to directly match negative and positive pools and
still have sufficient words for analysis. Additionally, although we collected arousal
data (see Supplemental Information) and found that, on average, negative items
had higher arousal than positive items, we could not simultaneously control for all
factors and still retain a sufficient number of words for analysis.
After applying the exclusion criteria, 1268 words were available for the free
recall analyses, 66 negative (5%), 981 neutral (77%) and 221 positive (17%).
Recently published valence ratings are available for 1555 of our 1638 words
(Warriner et al., 2013). We used this dataset to test the reliability of our ratings. We
found that though their valence ratings were slightly more positive (t(3191)= 2.5,
p = .01), both ratings were significantly positively correlated (rho = .9, p < .0001).
We used our ratings for the free recall study as values were available for all 1638
words.

73
5.4 Free recall study

Methods

152 (86 female) paid volunteers (ages 18 - 29) were recruited via fliers posted
around the University of Pennsylvania campus. Participants were provided with
a base monetary compensation plus an additional performance-based monetary
incentive to ensure full e↵ort. The Institutional Review Board at the University of
Pennsylvania approved our research protocol, and informed consent was obtained
from all participants.

Experimental Paradigm

The data reported in this manuscript were collected as part of the Penn Electro-
physiology of Encoding and Retrieval Study (PEERS), involving three multi-session
experiments that were sequentially administered. The data reported below come
from participants who completed all three experiments. Additional methodologi-
cal details can be found in the supplemental information.

Experiment 1 Participants performed an immediate free recall experiment con-

sisting of seven sessions of 16 lists of 16 words presented one at a time on a
computer screen. Each word was drawn from a pool of 1638 words taken from
the University of South Florida free association norms (Nelson et al., 2004, avail-
able at http://memory.psych.upenn.edu/wordpools/PEERS wordpool.zip). Seman-
tic relatedness was determined using the WAS model (Steyvers et al., 2004). WAS
similarity values were used to group words into four similarity bins (high similar-
ity, cos ✓ > 0.7; medium-high, 0.4 > cos ✓ < 0.7; medium-low, 0.14 >cos ✓ <0.4;
low similarity, cos ✓ < 0.14). Two pairs of items from each of the four groups were

74
arranged such that one pair occurred at adjacent serial positions and the other
pair was separated by at least two other items. All randomly generated word lists
conformed to this structure. The same word was not repeated in a session.
Words were presented concurrently either with a task cue, indicating the judg-
ment that the participant should make for that word, or with no encoding task. The
two encoding tasks were a size judgment (“Will this item fit into a shoebox?”) and
an animacy judgment (“Does this word refer to something living or not living?”),
and the current task was indicated by the color and typeface of the presented item.
There were four no-task lists (participants did not have to perform judgments with
the presented items), six single-task lists (all items were presented with the same
task, three of each task), and six task-shift lists (items were presented with either
task). List and task order were counterbalanced across sessions and participants.
The task manipulation was included to assess the e↵ect of task switching on clus-
tering during recall (Polyn et al., 2009b) and was not considered here.
For each list, there was a 1500ms delay before the first word appeared on the
screen. Each item was on the screen for 3000ms, followed by jittered 800 to 1200ms
interstimulus interval (uniform distribution). If the word was associated with a
task, participants indicated their response via a keypress. After the last item in the
list, there was a 1200 to 1400ms jittered delay, after which a tone sounded, a row
of asterisks appeared, and the participant was given 75s to attempt to recall any of
the just-presented items.

Experiment 2 Experiment 2 was identical to Experiment 1 except as described

below. There were either 7 (N = 92) or 9 (N = 60) experimental sessions of 12
study lists of 16 words. Experiment 2 included a mix of immediate recall lists,
delayed recall lists (the final word was followed by a distractor), and continual

75
distractor lists (each word was followed by a distractor). Distractor tasks consisted
of answering math problems of the form A + B + C = ?, where A, B, and C
were positive, single-digit integers, and participants typed the sum as quickly as
possible. Participants performed free recall with five possible time durations for
the between-item and end-of-list distractor tasks. As listed here, the first number
indicates the between-list distractor duration and the second number indicates the
end-of-list distractor duration, both in seconds: 0-0 for immediate (identical to
Experiment 1), 0-8 or 0-16 for delayed, and 8-8 or 16-16 for continual distractor
recall. The distractor manipulation was included to attenuate between-item and
end-of-list rehearsal (Lohnas & Kahana, 2014) and was not considered here.

Experiment 3 For 92 participants, Experiment 3 used the externalized recall (ER)

procedure (Zaromb et al., 2006) to measure participants’ tendency to commit intru-
sions during recall. Participants were instructed to say aloud every time a specific,
salient word came to mind while performing free recall and to press the spacebar
immediately following recall of an intrusion or repetition. Participants performed 6
experimental sessions with methods identical to Experiment 1 with ER-specific in-
structions. For the remaining 60 participants, Experiment 3 consisted of 4 sessions
completely identical to Experiment 1.

Conditional Response Probability Analysis The order in which items are re-
called is not random; each retrieval event is influenced by items already recalled.
To determine if valence drives recall order, we can measure participants’ tendency
to recall an item of a given valence conditionalized on the prior recall item’s valence.
We evaluated recall transitions as a function of valence to measure this tendency.
For each list, separate transition probabilities were calculated for each of 9 valence

76
combinations (negative to negative, negative to positive, etc.). To account for the
low likelihood of same valence transitions, we calculated the number of actual
transitions as a function of possible transitions, based on the study items in the list
and the items that had already been recalled. Therefore, the number of possible
transitions could range from 0 to 15. The likelihood of a transition was calculated
by dividing the number of actual transitions by the number of possible transitions.
The number of actual transitions was always 0 or 1. In this way we accounted for
the lower proportion of emotional items, and less likely transitions were given a
greater value. If a neutral-neutral transition was made, and there were 7 available
neutral items, the score for that transition would be 1/7. If a negative-negative
transition was made and there were 2 available negative items, the score for that
transition would be 1/2. Lists which had no positive or negative items, thus pre-
cluding any relevant transitions, were excluded from all analyses (on average 202
lists), resulting in on average 83 lists per participant. Probabilities were averaged
across lists within participants and then across participants.

5.5 Results and Discussion

We measured probability of recall as a function of valence. On average, participants

recalled 61% of negative words, 60% of neutral words, and 62% of positive words.
We ran a one-way repeated measures ANOVA comparing valence conditions and
found a significant interaction (F(2, 302) = 11.5 p < .01; ⌘2p = .03). Post-hoc com-
parisons using the Tukey HSD test indicated that the probability of recall di↵ered
significantly between both negative (M = .61, SD = .14) and neutral words (M =
.60, SD = .14) as well as between positive (M = .62, SD = .14) and neutral words (M
= .60, SD = .14). There was no di↵erence in probability of recall for negative and

77
positive words.
We observed reliably greater recall for both positive and negative items, which
is consistent with previous work (Dolcos et al., 2004; Kensinger & Schacter, 2006).
There is some evidence that memory is di↵erentially enhanced for positive or
negative items (Erk et al., 2003; Bennion et al., 2013). Such findings may be the
result of task design as many highly emotional items could drive unintended mood
changes in participants. By using a small proportion of emotional items, our study
avoided this possible confound and showed a benefit for both positive and negative
items. However, the small proportion of emotional items might make those items
more distinctive and thus easier to remember than neutral items. To account for
this possibility, we re-analyzed the data using equally distinctive emotional and
neutral items (see Supplemental Information) and found that emotional items were
still remembered better than neutral even when controlling for distinctiveness.
Memory enhancement for positive and negative items might be driven by con-
textual or salience mechanisms. If emotional features are integrated with context
and used as a retrieval cue, then recall outputs should be clustered by valence,
akin to various clustering phenomena whereby participants organize their recalls
according to semantic, episodic and spatial associations (Bousfield, 1953; Kahana,
1996; J. F. Miller, Lazarus, et al., 2013). To explore whether participants emotionally
cluster their retrievals during recall, we calculated the likelihood of transitioning
between words as a function of valence. Participants were more likely to transition
between same valence items than di↵erent valence items and this tendency was
greater than would be expected by chance as measured by repeatedly shu✏ing
recall order and calculating transition probabilities on this permuted data (Figure
5.1). To test the significance of this valence clustering e↵ect, we ran a 3⇥3 repeated
measures ANOVA comparing valence of item transitioned from (positive, neutral,

78
 0.25
Negative
Neutral

Transition probability
Positive

0.125

0
Negative Neutral Positive
Word Recalled

Figure 5.1: Conditional response probability based on valence. For each list, the proportion
of possible and actual transitions between the three valence categories was calculated. The y-
axis shows the probability of making a transition to a particular valence item (black, negative;
grey, neutral; white, positive) as a function of the just recalled item’s valence (x-axis). Error
bars are standard error of the mean. Gray dots are the expected transition value derived from
permuting the recall order within every list, session, subject for 100 iterations. Transitions between
similar valences (e.g. negative to negative, positive to positive) were significantly more likely than
transitions between dissimilar valences (e.g. negative to neutral, positive to neutral).

or negative) and valence of item transitioned to (positive, neutral, or negative). We

found a main e↵ect of item transitioned to (F(2, 302) = 18.8, p < .01 ⌘2p = .09) and an
interaction e↵ect (F(4, 302) = 28.2, p < .01; ⌘2p = .18). Post-hoc comparisons revealed
that negative to negative transitions were significantly more likely than negative
to neutral transitions (t(151) = 4.0, p < .01; Cohen’s d = .46) and positive to positive
transitions were significantly more likely than positive to neutral transitions (t(151)
= 2.8, p < .01; Cohen’s d = .30; Bonferroni corrected for 2 tests, p = .025). This
valence clustering e↵ect was not driven by within list semantic similarity, as con-
firmed by a combined semantic-valence conditional response probability analysis
(see Supplemental Information).
Together, these results suggest that contextual encoding and retrieval mecha-
nisms may explain increased memory for emotional items. Previous research has
suggested that contextual mechanisms may drive mood dependent and mood con-

79
gruent memory e↵ects (MDM, MCM; Bower, 1981; Lewis & Critchley, 2003), but
MCM and MDM e↵ects might represent a special case only relating to the mood-
induction paradigm (Eich, 1995). Alternatively, the use of many emotional items
could induce explicit strategy use that interferes with or overrides a natural con-
text mechanism. Existing research has provided conflicting evidence as to whether
emotion enhances (Doerksen & Shimamura, 2001) or impairs (Kensinger, 2009;
Maddock & Frein, 2009; Madan et al., 2012) associative memory. Furthermore,
a related study on the organizational e↵ects of emotion on memory (Siddiqui &
Unsworth, 2011) showed emotional clustering only when a valence orienting task
(pleasantness judgment) was used. Together, these results suggest that a large per-
centage of emotional items might create subjective and variable strategies, which
may interfere with the associative mechanisms that support emotional clustering.
In the current study, we found enhanced emotional memory and emotional clus-
tering, both of which are consistent with predictions of retrieved context models
(Howard & Kahana, 2002a) and suggest that emotional information is incorporated
into a slowly updating context representation during study and is used to cue re-
trieval of items at test. That this result occurs in the absence of mood induction
suggests a general mechanism by which emotional items are better remembered
than neutral items.
A leading theory of emotional memory enhancement suggests that emotional
items receive increased attentional processing (Dolcos et al., 2004). It is possible that
there are, in addition to the context e↵ects observed here, independent attentional
e↵ects that increase the strength of the item. These attentional e↵ects may be more
directly related to arousal than valence (Kensinger, 2009). Alternatively, attention
could interact with context by promoting item-to-context binding or increasing the
salience of emotional features in context.

80
5.5.1 Conclusions

The current study suggests that enhanced emotional memory is due to an asso-
ciative mechanism by which items, including emotional features, are associated
to a slowly updating context representation. Through a free recall task with few
emotional items, we have shown that participants recall emotional items more of-
ten than neutral items and that participants consecutively recall or cluster items
of the same emotional valence. These benefits are unlikely to be driven by mood
induction or explicit strategy use that may arise when many emotional items are
present or when an emotion-orienting task is used. Such a contextual mechanism
provides a framework by which emotional memory may be enhanced in situations
outside the laboratory.

5.6 Supplemental

Free recall study methods

The first session of each experiment was a practice session to familiarize participants
with the task. All sessions ended with a recognition test. Half of the sessions
(randomly chosen) included a final free recall test before recognition in which
participants recalled words from any of the lists from the session. The recognition
and final free recall manipulations are not considered here.
Based on the results of a prior norming study, only words that were clear in
meaning and that could be reliably judged in the size and animacy encoding tasks
were included in the pool. There were three conditions: no-task lists (participants
did not have to perform judgments with the presented items), single-task lists (all
items were presented with the same task), and task-shift lists (items were presented

81
with either task). The first two lists were task-shift lists, and each list started with
a di↵erent task. The next 14 lists contained four no-task lists, six single-task lists
(three of each task), and four task-shift lists.
The math distraction task was self-paced, such that a participant may have been
presented with, but not responded to, a problem at the end of the distraction inter-
val. Participants were given a monetary bonus based on the speed and accuracy of
their responses.

Controlling semantic relatedness at the word list level

We compared the distribution of negative-negative WAS or LSA (Latent Semantic

Analysis) semantic similarity values to the distribution of negative-neutral WAS
or LSA values using a two-sample Kolmogorov-Smirnov test. Negative-negative
similarity was not reliably larger than negative-neutral similarity (WAS, ks = .0047,
p = .95; LSA, ks < .01, p = .59). Positive-positive similarity was reliably larger than
positive-neutral similarity only for WAS (WAS, ks = .05, p < .01; LSA, ks < .01, p =
.99).
While we controlled for semantic relatedness at the level of word pools, we did
not control for semantic relatedness at the list level. To address this potential con-
found, we ran a combined semantic-valence CRP analysis. The semantic-valence
CRP (Figure 5.2) shows the likelihood of a particular transition, where there are four
valence transitions (negative-negative, negative-neutral, positive-neutral, positive-
positive) and either five semantic bins (WAS, [-.15 to .08, .08 to .31, .31 to .54, .54 to
.77, .77 to 1]) or three semantic bins (LSA, [-.2 to .2, .2 to .6, .6 to 1]). Di↵erent num-
bers of bins were used because there were no negative-negative transitions for LSA
bins > .7. For each transition within each list, we calculated the number of actual
and possible transitions in a given bin (e.g. negative-negative, semantic bin > .7).

82
 A. B.

Conditional response probability

1

Conditional response probability

0.5
0.5

0 0
−0.15 0.08 0.31 0.54 0.77 −0.2 0.2 0.6
Semantic relatedness Semantic relatedness

Figure 5.2: Semantic-valence CRP. The plots show four transition types (negative-negative, solid,
black; negative-neutral, dashed, black; positive-positive solid, light gray; positive-neutral, dashed,
light gray) as a function of semantic bin. Error bars are standard error of the mean. (A) Five
semantic WAS bins (-.15 to .08, .08 to .31, .31 to .54, .54 to .77, .77 to 1). (B) Three semantic LSA bins
(-.2 to .2, .2 to .6, .6 to 1)

Importantly, as each of these calculations occurred within list, it accounts for any
imbalance in within-list semantic similarity. If negative-negative transitions were
driven exclusively by semantic relatedness, we would predict that high-semantic
negative-negative transitions would be equally likely as high-semantic negative-
neutral transitions. We found that even when controlling for semantic relatedness
within list, same valence transitions were more likely than di↵erent valence transi-
tions. These results argue against a within-list semantic relatedness e↵ect.

Distinctiveness

By virtue of being a small proportion of the total items presented, emotional items
might have been more distinctive than neutral items. To assess this possibility,
for each study list we calculated the average number of semantic neighbors for
positive, neutral and negative items. Semantic neighbors were any other items
presented on the same list with a WAS value of .2 or greater. The assumption is

83
that items with more neighbors are less distinctive and that neutral items will tend
to have more neighbors. Therefore, we iteratively removed non-distinct neutral
items until neutral items had on average .21 neighbors per list, which did not
di↵er from the average number of list neighbors for negative items, .20 (t(151) =
1.6, p = .12). Positive items had on average .24 neighbors per list, which was
significantly greater than the number of neutral item neighbors (t(151) = 8.9 p <
.01), leaving neutral items more distinctive than positive items. When controlling for
distinctiveness in this manner, probability of recall was still greater for emotional
than neutral items (negative vs. neutral, t(151) = 2.2, p = .03; positive vs. neutral
t(151) = 4, p < .01). Likewise, transitions between same-valence items was still
more likely than transitions between di↵erent-valence items (negative-negative vs.
negative-neutral, t(151) = 5.1, p < .01, positive-positive vs. positive-neutral, t(151)
= 7.5, p < .01).

84
Chapter 6

General discussion

6.1 Summary

In the previous chapters, I examined the neural mechanisms of episodic memory

formation using the free recall paradigm coupled with electroencephalographic
(EEG) recordings. My results suggest that episodic memory formation is driven
by contextual encoding processes which are supported by high frequency activity
(HFA) increases in the memory network.
In Chapter 2, I measured the spectral correlates of episodic memory formation
using matched data analytic methods in a scalp EEG and an intracranial EEG
(iEEG) study. I analyzed the subsequent memory e↵ect (SME) by comparing the
encoding signals of items subsequently recalled to those subsequently forgotten.
Although theta (3 - 8 Hz) and gamma (> 30 Hz) signals often vary during memory
formation (Nyhus & Curran, 2010; Johnson & Knight, 2015), the specific pattern
of power increases and decreases in these frequencies has been unclear (Klimesch,
1996; Sederberg et al., 2003; Gruber et al., 2004; Sederberg et al., 2006; Guderian et
al., 2009; Burke, Long, et al., 2014). This discrepancy in findings may be the result

85
of di↵erences in recording modalities (Stoller, 1949; Yuval-Greenberg et al., 2008)
or di↵erences in the choice of analysis parameters. Using matched data analytic
methods, I found that broad HFA increases and LFA decreases are associated with
successful memory formation in both iEEG and scalp EEG, suggesting that theta
e↵ects in iEEG are not likely the result of pathology and that scalp EEG is capable
of resolving HFA signals. As this pattern of HFA increases and LFA decreases has
been observed across domains outside of memory (Crone et al., 2001; Jerbi et al.,
2009; K. J. Miller et al., 2007) and correlates with the BOLD e↵ect in fMRI (Mukamel
et al., 2005; Niessing et al., 2005), this pattern may not reflect a memory specific
signal per se, but rather may indicate a more non-specific underlying process of
general cortical activation and that it is the precise intersection of timing and spatial
location of these e↵ects that is important.
Next, I connected HFA increases in the SME to contextual encoding. HFA
increases in the memory network which were observed in Chapter 2 might re-
flect memory-specific processing or a more general mnemonic attentional signal.
Although the SME analysis is a classic approach to understanding memory for-
mation (Wagner et al., 1998; Brewer et al., 1998), it is a coarse measure as an item
may be forgotten due to reasons unrelated to memory processing. In Chapter 3, I
defined items based on how they were recalled, by comparing subsequently clus-
tered and not clustered items, where clustered items are those recalled preceding
or following a study neighbor. I used clustering as a proxy for contextual encod-
ing, as retrieved context theory (Howard & Kahana, 2002a; Sederberg et al., 2008;
Polyn et al., 2009a) suggests that clustering arises due to the existence of a slowly
updating context representation that is associated to items at encoding and is used
as a retrieval cue during test. I found that HFA increases in the memory network
track contextual encoding and that the time course of HFA increases varies across

86
brain regions. Specifically, HFA increases in left prefrontal cortex (PFC) preceded
those in hippocampus. These late hippocampal clustering signals positively cor-
related with behavioral measures of clustering. These results suggest that control
processes in left PFC coupled with associative processes in hippocampus underlie
contextual encoding processes that support successful memory formation.
In Chapter 4, I measured the influence of semantic processing on the formation
of episodic associations. By including a semantic orienting task on a subset of
study lists, I was able to compare the e↵ects of subsequent temporal clustering
with and without a task that oriented participants to the semantic associations
among items. I found increased HFA over left prefrontal regions for subsequently
temporally clustered items only for no-task lists, those without a semantic orient-
ing task. I also found increased HFA over left prefrontal regions for subsequently
semantically clustered items on for task lists. These results suggest a dissociation
between semantic and episodic processing during encoding. I further confirmed
this dissociation by comparing the encoding patterns of subsequently recalled and
not recalled semantically isolated items. HFA during task lists for semantically
isolated items, those study items paired with a low semantic associate, was in-
creased when those items were subsequently not recalled. These left prefrontal
HFA e↵ects are likely to reflect a control mechanism which biases processing in
posterior regions according to task demands. The results I have observed suggest
that semantic processing interferes with the formation of episodic associations.
In Chapter 5, I used retrieved context theory to account for the memory ben-
efit observed for emotionally valenced items. Positively and negatively valenced
items are remembered better than items with a neutral valence (LaBar & Cabeza,
2006), which may be because valenced items receive more attentional processing
than neutral items. Alternatively, if emotional features, like semantic and temporal

87
features, are incorporated into the context representation, memory for emotional
items will be enhanced due to the presence of emotional features in the contextual
retrieval cue at test. Of these two hypotheses, only retrieved context theory pre-
dicts that recalls should be clustered by valence. I measured recall and transition
probabilities for items of positive, negative, and neutral valence. I found that both
positive and negative items were more likely to be recalled than neutral items and
that participants were more likely to make transitions between same valence items
than di↵erently valenced items. These results suggest that emotional features are
represented in context and provide a mechanism by which emotions might be
processed outside the laboratory.
These studies broadly provide evidence in support of a contextual encoding
mechanism whereby items are associated to a slowly updating context represen-
tation in service of memory formation. Although I consistently found high fre-
quency e↵ects, there are likely other signals associated with contextual encoding.
Furthermore, as my work has been focused on encoding, a critical next step will
be uncovering the contributions of contextual processing during retrieval. Simi-
larly, to have a complete understanding of contextual processes, we must integrate
the univariate findings I present here with the multivariate patterns that represent
contextual encoding and retrieval patterns. I discuss each of these in turn below.

6.2 Future Directions

Theta and phase

The research that I have presented in this thesis shows that contextual processing is
most clearly represented by changes in HFA. However, other studies have shown

88
theta increases associated with contextually mediated memories (Hanslmayr &
Staudigl, 2013). The majority of theta e↵ects I observe are theta decreases, rather
than increases. These are likely broadband, asynchronous e↵ects which could
potentially mask transient narrowband e↵ects.
Where theta might prove to be critical is in the phase, rather than the am-
plitude domain. Stimulation during the peak of a theta oscillation can induce
LTP (Axmacher et al., 2006) and phase locking of single unit activity to theta oscil-
lations is increased during successful memory formation (Rutishauser et al., 2010).
Phase reset of theta oscillations during memory formation is negatively corre-
lated to theta power (Rizzuto et al., 2003), as are memory related phase-synchrony
increases (Burke et al., 2013), providing compelling evidence that the power de-
creases I observe may co-occur with phase changes. Also promising is the finding
of phase-amplitude coupling (PAC) between theta phase and gamma amplitude
during memory formation (Canolty et al., 2006; Jensen et al., 2007; Lega et al., in
press). The HFA increases I have observed might be coupled to a particular theta
phase, suggesting that contextual encoding might be supported by both of these
frequency bands.

Retrieval

The goal of my thesis was to elucidate the neural correlates of episodic memory
formation. However, as stated in the introduction, successful memory is based on
success during encoding as well as retrieval. Contextual encoding is only e↵ective
insofar as context is reinstated at test. EEG analyses of retrieval has revealed
HFA increases preceding successful recall (Sederberg, Schulze-Bonhage, Madsen,
Bromfield, Litt, et al., 2007; Lega et al., 2011; Burke, Sharan, et al., 2014) which could
reflect controlled memory search processes or the reinstatement of memories.

89
 I have identified two potentially distinct signals of contextual encoding, a con-
trol mechanism supported by left prefrontal cortex and an associative mechanism
supported by hippocampus. Given fMRI results connecting left prefrontal cortex
to controlled retrieval processes (Badre et al., 2005; Badre & Wagner, 2007), one
might predict that left prefrontal HFA increases should be observed during the
retrieval of clustered compared to not clustered items. However, fMRI evidence
suggests that there may be a dissociation between encoding and retrieval such that
items which received strong associative processing during encoding might require
less controlled retrieval during test (Long et al., 2010). Thus a critical next step will
be to detail exactly when and where in the brain we observe HFA increases during
the retrieval of clustered and non-clustered items.

Multivariate analyses and patterns

The work in my thesis used univariate analyses to elucidate the neural correlates
of episodic memory formation. However, multivariate analyses of patterns of
brain activity have proven to be incredibly fruitful for addressing questions about
memory formation and retrieval (Rissman & Wagner, 2012). The patterns of ac-
tivity during encoding are reinstated during retrieval (Polyn et al., 2005) and the
strength of reinstatement predicts memory success (Kuhl et al., 2012). Multivariate
analyses of the similarity between encoding and retrieval patterns have suggested
the existence of a slowly drifting context representation (Manning et al., 2011, 2012;
Morton et al., 2013). Specifically, the similarity between encoding and retrieval pat-
terns increases as a function of temporal or semantic similarity between stimulus
items and predicts the degree to which participants will cluster their recalls.
My research has identified the control and associative processes which might
enable items to interact with context. The next step will be to measure these

90
univariate signals in tandem with the multivariate contextual signals. Changes in
HFA should predict the degree of pattern similarity between items at encoding and
at retrieval.

6.3 Concluding remarks

The work in my thesis provides a substantial contribution to our understanding

of episodic encoding mechanisms. Knowing that memory formation is driven by
contextual processing, we can move beyond free recall, to explain findings in recog-
nition, interference and inference tasks. Extending my work to other paradigms
and questions will bring us closer to a unified theory of memory function.

91
References

Anderson, A. K., & Phelps, E. A. (2001). Lesions of the human amygdala impair
enhanced perception of emotionally salient events. Nature, 411, 305–309.
Axmacher, N., Mormann, F., Fernández, G., Elger, C. E., & Fell, J. (2006). Memory
formation by neuronal synchronization. Brain research reviews, 52, 170–182.
Badre, D., Poldrack, R. A., Pare-Blagoev, E. J., Insler, R. Z., & Wagner, A. D. (2005).
Dissociable controlled retrieval and generalized selection mechanisms in ven-
trolateral prefrontal cortex. Neuron, 47, 907-918.
Badre, D., & Wagner, A. (2007). Left ventrolateral prefrontal cortex and the cognitive
control of memory. Neuropsychologia, 45, 2883–2901.
Baran, B., Pace-Schott, E. F., Ericson, C., & Spencer, R. M. (2012). Processing
of emotional reactivity and emotional memory over sleep. The Journal of
Neuroscience, 32, 1035–1042.
Bedny, M., McGill, M., & Thompson-Schill, S. L. (2008). Semantic adaptation and
competition during word comprehension. Cerebral Cortex, 18, 2574–2585.
Bennion, K. A., Ford, J. H., Murray, B. D., & Kensinger, E. A. (2013). Over-
simplification in the study of emotional memory. Journal of the International
Neuropsychological Society, 19, 953–961.
Berger, H. (1929). Über das Elektrenkephalogramm des menschen (On the human
electroencephalogram). Archiv für Psychiatry und Nervenkrankheiten, 87, 527-

92
 570.
Binder, J. R., & Desai, R. H. (2011). The neurobiology of semantic memory. Trends
in cognitive sciences, 15, 527–536.
Blumenfeld, R., & Ranganath, C. (2007). Prefrontal cortex and long-term mem-
ory encoding: An integrative review of findings from neuropsychology and
neuroimaging. The Neuroscientist, 13, 280–291.
Bousfield, W. A. (1953). The occurrence of clustering in the recall of randomly
arranged associates. Journal of General Psychology, 49, 229-240.
Bower, G. H. (1972). Stimulus-sampling theory of encoding variability. In
A. W. Melton & E. Martin (Eds.), Coding processes in human memory (p. 85-
121). New York: John Wiley and Sons.
Bower, G. H. (1981). Mood and memory. American Psychologist, 36, 129-148.
Bradley, M. M., & Lang, P. J. (1999). A↵ective norms for English words (ANEW): Stimuli,
instruction manual and a↵ective ratings (Tech. Rep.). Gainesville: University of
Florida, Center for Research in Psychophysiology.
Brassen, S., Weber-Fahr, W., Sommer, T., Lehmbeck, J. T., & Braus, D. F. (2006).
Hippocampal–prefrontal encoding activation predicts whether words can be
successfully recalled or only recognized. Behavioural brain research, 171, 271–
278.
Brewer, J. B., Zhao, Z., Desmond, J. E., Glover, G. H., & Gabrieli, J. D. (1998).
Making memories: Brain activity that predicts how well visual experience
will be remembered. Science, 281, 1185-7.
Brown, S., Conover, J., Flores, L., & Goodman, K. (1991). Clustering and recall: Do
high clusterers recall more than low clusterers because of clustering. Journal
of Experimental Psychology: Learning, Memory, & Cognition, 17, 710–721.
Bunge, S. A., Wendelken, C., Badre, D., & Wagner, A. D. (2005). Analogical rea-

93
 soning and prefrontal cortex: evidence for separable retrieval and integration
mechanisms. Cerebral cortex, 15, 239–249.
Burgess, A. P., & Gruzelier, J. H. (1997). Short duration synchronization of human
theta rhythm during recognition memory. Neuroreport, 8, 1039-1042.
Burke, J. F., Long, N. M., Zaghloul, K. A., Sharan, A. D., Sperling, M. R., & Kahana,
M. J. (2014). Human intracranial high-frequency activity maps episodic
memory formation in space and time. NeuroImage, 85 Pt. 2, 834–843.
Burke, J. F., Ramayya, A. G., & Kahana, M. J. (2015). Human intracranial high-
frequency activity during memory processing: Neural oscillations or stochas-
tic volatility? Current Opinion in Neurobiology, 31, 104-110.
Burke, J. F., Sharan, A. D., Evans, J. J., Healey, M. K., Sperling, M. R., Beck, E. N., . . .
Kahana, M. J. (2014). Theta and high–frequency activity mark spontaneous
recall of episodic memories. The Journal of Neuroscience, 34, 11355-11365.
Burke, J. F., Zaghloul, K. A., Jacobs, J., Williams, R. B., Sperling, M. R., Sharan,
A. D., & Kahana, M. J. (2013). Synchronous and asynchronous theta and
gamma activity during episodic memory formation. Journal of Neuroscience,
33, 292–304.
Buzsáki, G. (2002). Theta oscillations in the hippocampus. Neuron, 33, 325–340.
Buzsáki, G. (2005). Theta rhythm of navigation: Link between path integration
and landmark navigation, episodic and semantic memory. Hippocampus, 15,
827–840.
Canli, T., Zhao, Z., Brewer, J., Gabrieli, J. D., Cahill, L., et al. (2000). Event-
related activation in the human amygdala associates with later memory for
individual emotional experience. Journal of Neuroscience, 20, 1-5.
Canolty, R. T., Edwards, E., Dalal, S. S., Soltani, M., Nagarajan, S. S., Kirsch, H. E., . . .
Knight, R. T. (2006). High gamma power is phase-locked to theta oscillations

94
 in human neocortex. Science, 313, 1626–1628.
Canolty, R. T., & Knight, R. T. (2010). The functional role of cross-frequency
coupling. Trends in Cognitive Science, 14, 506-515.
Caramazza, A., & Shelton, J. (1998). Domain-specific knowledge systems in the
brain: The animate-inanimate distinction. Journal of Cognitive Neuroscience,
10, 1–34.
Carr, V. A., Rissman, J., & Wagner, A. D. (2010). Imaging the human medial
temporal lobe with high-resolution fmri. Neuron, 65, 298–308.
Chatham, C. H., Frank, M. J., & Badre, D. (2014). Corticostriatal output gating
during selection from working memory. Neuron, 81, 930–942.
Cicchetti, D. V. (1994). Guidelines, criteria, and rules of thumb for evaluating
normed and standardized assessment instruments in psychology. Psycholog-
ical assessment, 6, 284-290.
Craik, F. I. M., & Tulving, E. (1975). Depth of processing and the retention of words
in episodic memory. Journal of Experimental Psychology, 104, 268-294.
Crone, N. E., Boatman, D., Gordon, B., & Hao, L. (2001). Induced electrocortico-
graphic gamma activity during auditory perception. Clinical Neurophysiology,
112, 565–582.
Crone, N. E., Korzeniewska, A., & Franaszczuk, P. (2011, Jan). Cortical gamma
responses: searching high and low. International Journal of Psychophysiology,
79, 9-15.
Crone, N. E., Miglioretti, D. L., Gordon, B., & Lesser, R. P. (1998a). Functional
mapping of human sensorimotor cortex with electrocorticographic spectral
analysis. II. Event-related synchronization in the gamma band. Brain, 121,
2301.
Crone, N. E., Miglioretti, D. L., Gordon, B., Sieracki, J. M., Wilson, M. T., Uematsu,

95
 S., & Lesser, R. P. (1998b). Functional mapping of human sensorimotor cortex
with electrocorticographic spectral analysis. I. Alpha and beta event-related
desynchronization. Brain, 121, 2271–2299.
Davachi, L., & DuBrow, S. (2015). How the hippocampus preserves order: the role
of prediction and context. Trends in cognitive sciences, 19, 1–8.
Davachi, L., & Wagner, A. D. (2002). Hippocampal contributions to episodic
encoding: insights from relational and item-based learning. Journal of Neuro-
physiology, 88, 982–990.
Demb, J. B., Desmond, J. E., Wagner, A. D., Vaidya, C. J., Glover, G. H., & Gabrieli,
J. D. E. (1995). Semantic encoding and retrieval in the left inferior prefrontal
cortex: A functional mri study of task difficulty and process specificity. Journal
of Neuroscience, 15, 5870–5878.
Dickerson, B., Miller, S., Greve, D., Dale, A., Albert, M., Schacter, D., & Sperling, R.
(2007). Prefrontal-hippocampal-fusiform activity during encoding predicts
intraindividual di↵erences in free recall ability: An event-related functional-
anatomic mri study. Hippocampus, 17, 1060–1070.
Doerksen, S., & Shimamura, A. P. (2001). Source memory enhancement for emo-
tional words. Emotion, 1, 5.
Dolcos, F., LaBar, K., & Cabeza, R. (2004). Interaction between the amygdala and the
medial temporal lobe memory system predicts better memory for emotional
events. Neuron, 42, 855–863.
Eich, E. (1995). Searching for mood dependent memory. Psychological Science, 6,
67-75.
Eichenbaum, H. (2004). Hippocampus: cognitive processes and neural represen-
tations that underlie declarative memory. Neuron, 44, 109–120.
Engel, A., Fries, P., & Singer, W. (2001). Dynamic predictions: oscillations and

96
 synchrony in top-down processing. Nature Reviews Neuroscience, 2, 704–716.
Erk, S., Kiefer, M., Grothe, J., Wunderlich, A. P., Spitzer, M., & Walter, H. (2003).
Emotional context modulates subsequent memory e↵ect. Neuroimage, 18,
439–447.
Estes, W. K. (1955). Statistical theory of distributional phenomena in learning.
Psychological Review, 62, 369-377.
Ezzyat, Y., & Davachi, L. (2014). Similarity breeds proximity: Pattern similarity
within and across contexts is related to later mnemonic judgments of temporal
proximity. Neuron, 81, 1179–1189.
Fell, J., & Axmacher, N. (2011, February). The role of phase synchronization in
memory processes. Nature Reviews. Neuroscience, 12, 105-118.
Fell, J., Klaver, P., Lehnertz, K., Grunwald, T., Schaller, C., Elger, C. E., & Fernandez,
G. (2001). Human memory formation is accompanied by rhinal-hippocampal
coupling and decoupling. Nature Neuroscience, 4, 1259-1264.
Fell, J., Ludowig, E., Staresina, B., Wagner, T., Kranz, T., Elger, C. E., & Axmacher, N.
(2011). Medial temporal theta/alpha power enhancement precedes successful
memory encoding: evidence based on intracranial eeg. Journal of Neuroscience,
31, 5392-5397.
Fernandez, G., Brewer, J. B., Zhao, Z., Glover, G. H., & Gabrieli, J. D. (1999). Level
of sustained entorhinal activity at study correlates with subsequent cued-
recall preformance: a functional magnetic resonance imaging study with
high acquisition rate. Hippocampus, 9, 35-44.
Fernandez, G., E↵ern, A., Grunwald, T., Pezer, N., Lehnertz, K., Dumpelmann, M.,
. . . Elger, C. E. (1999). Real-time tracking of memory formation in the human
rhinal cortex and hippocampus. Science, 285, 1582-1585.
Fernández, G., Weyerts, H., Schrader-Bölsche, M., Tendolkar, I., Smid, H. G. O. M.,

97
 Tempelmann, C., . . . Jochen Heinze, H. (1998). Successful verbal encoding
into episodic memory engages the posterior hippocampus: A parametrically
analyzed functional magnetic resonance imaging study. Journal of Neuro-
science, 18, 1841-1847.
Fries, P. (2005). A mechanism for cognitive dynamics: neuronal communication
through neuronal coherence. Trends Cogn Sci, 9, 474–480.
Gevins, A., Smith, M. E., McEvoy, L., & Yu, D. (1997). High-resolution EEG
mapping of cortical activation related to working memory: E↵ects of task
difficulty, type of processing, and practice. Cerebral Cortex, 7, 374–385.
Glanzer, M., & Adams, J. K. (1990). The mirror e↵ect in recognition memory:
data and theory. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 16, 5-16.
Gold, C., Henze, D., Koch, C., & Buzsaki, G. (2006). On the Origin of the Extracellu-
lar Action Potential Waveform: A Modeling Study. Journal of Neurophysiology,
95, 3113–3128.
Gruber, T., Tsivilis, D., Montaldi, D., & Müller, M. (2004). Induced gamma band
responses: An early marker of memory encoding and retrieval. Neuroreport,
15, 1837–1841.
Guderian, S., Schott, B., Richardson-Klavehn, A., & Duzel, E. (2009). Medial
temporal theta state before an event predicts episodic encoding success in
humans. Proceedings of the National Academy of Sciences, USA, 106, 5365.
Hanslmayr, S., Spitzer, B., & Bauml, K. (2009). Brain oscillations dissociate between
semantic and nonsemantic encoding of episodic memories. Cerebral Cortex,
19, 1631-1640.
Hanslmayr, S., & Staudigl, T. (2013). How brain oscillations form memories–
a processing based perspective on oscillatory subsequent memory e↵ects.

98
 NeuroImage.
Hanslmayr, S., Staudigl, T., & Fellner, M. (2012). Oscillatory power decreases
and long-term memory: the information via desynchronization hypothesis.
Frontiers in Human Neuroscience, 6.
Hanslmayr, S., Volberg, G., Wimber, M., Raabe, M., Greenlee, M. W., & Bäumel,
K. H. T. (2011). The relationship between brain oscillations and bold signal
during memory formation: A combined eeg-fmri study. Journal of Neuro-
science, 31, 15674-15680.
Hasselmo, M. E. (2005). What is the function of hippocampal theta rhythm?–
Linking behavioral data to phasic properties of field potential and unit record-
ing data. Hippocampus, 15, 936–949.
Hasselmo, M. E., & Stern, C. E. (2014). Theta rhythm and the encoding and retrieval
of space and time. Neuroimage, 85, 656–666.
Hayama, H., & Rugg, M. D. (2009). Right dorsolateral prefrontal cortex is engaged
during post-retrieval processing of both episodic and semantic information.
Neuropsychologia, 47, 2409.
Hazy, T. E., Frank, M. J., & OReilly, R. C. (2006). Banishing the homunculus:
making working memory work. Neuroscience, 139, 105–118.
Howard, M. W., & Kahana, M. J. (1999). Contextual variability and serial position
e↵ects in free recall. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 25, 923–941.
Howard, M. W., & Kahana, M. J. (2002a). A distributed representation of temporal
context. Journal of Mathematical Psychology, 46, 269–299.
Howard, M. W., & Kahana, M. J. (2002b). When does semantic similarity help
episodic retrieval? Journal of Memory and Language, 46, 85–98.
Howard, M. W., Rizzuto, D. S., Caplan, J. C., Madsen, J. R., Lisman, J.,

99
 Aschenbrenner-Scheibe, R., . . . Kahana, M. J. (2003). Gamma oscillations
correlate with working memory load in humans. Cerebral Cortex, 13, 1369–
1374.
Hsieh, L.-T., Ekstrom, A. D., & Ranganath, C. (2011). Neural oscillations associated
with item and temporal order maintenance in working memory. The Journal
of Neuroscience, 31, 10803–10810.
Hwang, D. Y., & Golby, A. J. (2006). The brain basis for episodic memory: insights
from functional mri, intracranial eeg, and patients with epilepsy. Epilepsy &
Behavior, 8, 115–126.
Jacobs, J., & Kahana, M. J. (2009). Neural representations of individual stimuli in
humans revealed by gamma-band ECoG activity. Journal of Neuroscience, 29,
10203–10214.
Jacobs, J., & Kahana, M. J. (2010). Direct brain recordings fuel advances in cognitive
electrophysiology. Trends in Cognitive Sciences, 14, 162–171.
Jenkins, L. J., & Ranganath, C. (2010). Prefrontal and medial temporal lobe activity
at encoding predicts temporal context memory. Journal of Neuroscience, 30,
15558–15565.
Jensen, O., Kaiser, J., & Lachaux, J. (2007). Human gamma-frequency oscillations
associated with attention and memory. Trends in Neurosciences, 30, 317–324.
Jerbi, K., Freyermuth, S., Minotti, L., Kahane, P., Berthoz, A., & Lachaux, J. (2009).
Watching Brain TV and Playing Brain Ball:: Exploring Novel BCI Strategies
using Real-Time Analysis of Human Intracranial Data. International Review of
Neurobiology, 159–168.
Johnson, E. L., & Knight, R. T. (2015). Intracranial recordings and human memory.
Current opinion in neurobiology, 31, 18–25.
Jutras, M. J., & Bu↵alo, E. A. (2010). Synchronous neural activity and memory

100
 formation. Current Opinion in Neurobiology, 20, 150–155.
Kahana, M. J. (1996). Associative retrieval processes in free recall. Memory &
Cognition, 24, 103–109.
Kahana, M. J. (2012). Foundations of human memory. New York, NY: Oxford
University Press.
Kandel, E. R. (2001). The molecular biology of memory storage: a dialogue between
genes and synapses. Science, 294, 1030–1038.
Kensinger, E. A. (2009). Remembering the details: E↵ects of emotion. Emotion
review, 1, 99–113.
Kensinger, E. A., & Schacter, D. L. (2006). Processing emotional pictures and words:
e↵ects of valence and arousal. Cognitive, A↵ective, & Behavioral Neuroscience,
6, 110–126.
Kilner, J., Mattout, J., Henson, R., & Friston, K. (2005). Hemodynamic correlates of
eeg: a heuristic. Neuroimage, 28, 280–286.
Kim, H. (2011). Neural activity that predicts subsequent memory and forgetting:
a meta-analysis of 74 fMRI studies. NeuroImage, 54, 2446–2461.
Kirwan, C. B., & Stark, C. E. (2004). Medial temporal lobe activation during
encoding and retrieval of novel face-name pairs. Hippocampus, 14, 919–930.
Kissler, J., Assadollahi, R., & Herbert, C. (2006). Emotional and semantic networks
in visual word processing: insights from ERP studies. Progress in brain research,
156, 147–183.
Klimesch, W. (1996). Memory processes, brain oscillations and eeg synchronization.
International Journal of Psychophysiology, 24, 61–100.
Klimesch, W. (1999). EEG alpha and theta oscillations reflect cognitive and memory
performance: a review and analysis. Brain Research Reviews, 29, 169–195.
Klimesch, W., Doppelmayr, M., Russegger, H., & Pachinger, T. (1996). Theta

101
 band power in the human scalp EEG and the encoding of new information.
NeuroReport, 7, 1235–1240.
Klimesch, W., Doppelmayr, M., Russegger, H., Pachinger, T., & Schwaiger, J. (1998).
Induced alpha band power changes in the human eeg and attention. Neuro-
science letters, 244, 73–76.
Klimesch, W., Doppelmayr, M., Schimke, H., & Ripper, B. (1997). Theta synchro-
nization and alpha desynchronization in a memory task. Psychophysiology,
34, 169–176.
Kovach, C. K., Tsuchiya, N., Kawasaki, H., Oya, H., Howard, M. A., & Adolphs,
R. (2011). Manifestation of ocular-muscle EMG contamination in human
intracranial recordings. NeuroImage, 54, 213–233.
Kuhbandner, C., & Pekrun, R. (2013). Joint e↵ects of emotion and color on memory.
Emotion, 13, 375.
Kuhl, B. A., Bainbridge, W., & Chun, M. (2012). Neural reactivation reveals
mechanisms for updating memory. Journal of Neuroscience, 32, 3453–3461.
Kuhl, B. A., Johnson, M. K., & Chun, M. M. (2013). Dissociable neural mecha-
nisms for goal-directed versus incidental memory reactivation. The Journal of
Neuroscience, 33, 16099–16109.
LaBar, K. S., & Cabeza, R. (2006). Cognitive neuroscience of emotional memory.
Nature Reviews Neuroscience, 7, 54–64.
Lachaux, J. P., Axmacher, N., Mormann, F., Halgren, E., & Crone, N. E. (2012). High-
frequency neural activity and human cognition: Past, present, and possible
future of intracranial EEG research. Progress in Neurobiology, 98, 279-301.
Lachaux, J. P., Fonlupt, P., Kahane, P., Minotti, L., Ho↵mann, D., Bertrand, O., &
Baciu, M. (2007). Relationship between task-related gamma oscillations and
bold signal: New insights from combined fMRI and intracranial EEG. Hum

102
 Brain Mapping.
Lachaux, J. P., Rudrauf, D., & Kahane, P. (2003). Intracranial EEG and human brain
mapping. Journal of Physiology-Paris, 97, 613–628.
Lega, B. C., Burke, J. F., Jacobs, J., & Kahana, M. J. (in press). Slow theta-to-gamma
phase amplitude coupling in human hippocampus supports the formation of
new episodic memories. Cerebral Cortex.
Lega, B. C., Jacobs, J., & Kahana, M. J. (2011). Human hippocampal theta oscillations
and the formation of episodic memories. Hippocampus, 22, 748–761.
Lewis, P., & Critchley, H. D. (2003). Mood-dependent memory. Trends in Cognitive
Sciences, 7, 431–433.
Lisman, J. E., & Jensen, O. (2013). The theta-gamma neural code. Neuron, 77,
1002–1016.
Liu, Y., Huang, H., McGinnis-Deweese, M., Keil, A., & Ding, M. (2012). Neural
substrate of the late positive potential in emotional processing. The Journal of
Neuroscience, 32, 14563–14572.
Logothetis, N. K. (2003). The underpinnings of the BOLD functional magnetic
resonance imaging signal. Journal of Neuroscience, 23, 3963 - 3971.
Logothetis, N. K., Pauls, J., Augath, M., Trinath, T., & Oeltermann, A. (2001).
Neurophysiological investigation of the basis of the fMRI signal. Nature, 412,
150–157.
Lohnas, L. J., & Kahana, M. J. (2013). Parametric e↵ects of word frequency e↵ect in
memory for mixed frequency lists. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 39, 1943–1946.
Lohnas, L. J., & Kahana, M. J. (2014). Compound cuing in free recall. Journal of
Experimental Psychology: Learning, Memory and Cogntion, 40, 12-24.
Lohnas, L. J., Polyn, S. M., & Kahana, M. J. (in press). Expanding the scope

103
 of memory search: Intralist and interlist e↵ects in free recall. Psychological
Review.
Long, N. M., Burke, J. F., & Kahana, M. J. (2014). Subsequent memory e↵ect in
intracranial and scalp EEG. NeuroImage, 84, 488–494.
Long, N. M., Öztekin, I., & Badre, D. (2010). Seperable prefrontal cortex contribu-
tions to free recall. J Neurosci, 30, 10967 - 10976.
Madan, C. R., Caplan, J. B., Lau, C. S., & Fujiwara, E. (2012). Emotional arousal
does not enhance association-memory. Journal of Memory and Language, 66,
695–716.
Maddock, R. J., & Frein, S. T. (2009). Reduced memory for the spatial and temporal
context of unpleasant words. Cognition and Emotion, 23, 96–117.
Manning, J. R., Jacobs, J., Fried, I., & Kahana, M. J. (2009). Broadband shifts in LFP
power spectra are correlated with single-neuron spiking in humans. Journal
of Neuroscience, 29, 13613 - 13620.
Manning, J. R., Polyn, S. M., Baltuch, G., Litt, B., & Kahana, M. J. (2011). Oscillatory
patterns in temporal lobe reveal context reinstatement during memory search.
Proceedings of the National Academy of Sciences, USA, 108, 12893–12897.
Manning, J. R., Sperling, M. R., Sharan, A., Rosenberg, E. A., & Kahana, M. J.
(2012). Spontaneously reactivated patterns in frontal and temporal lobe pre-
dict semantic clustering during memory search. Journal of Neuroscience, 32,
8871–8878.
Manns, J. R., Howard, M. W., & Eichenbaum, H. (2007). Gradual changes in
hippocampal activity support remembering the order of events. Neuron, 56,
530–540.
Maratos, E. J., & Rugg, M. D. (2001). Electrophysiological correlates of the retrieval
of emotional and non-emotional context. Journal of Cognitive Neuroscience, 13,

104
 877–891.
Martin, A. (2007). The representation of object concepts in the brain. Psychology,
58, 25-45.
Martin, A., Haxby, J. V., Lalonde, F. M., Wiggs, C. L., & Ungerleider, L. G. (1995).
Discrete cortical regions associated with knowledge of color and knowledge
of action. Science, 270, 102–105.
Mason, W., & Suri, S. (2012). Conducting behavioral research on Amazons Me-
chanical Turk. Behavior research methods, 44, 1–23.
Matsumoto, J. Y., Stead, M., Kucewicz, M. T., Matsumoto, A. J., Peters, P. A.,
Brinkmann, B. H., . . . others (2013). Network oscillations modulate interictal
epileptiform spike rate during human memory. Brain, 136.
Mayes, A., Montaldi, D., & Migo, E. (2007). Associative memory and the medial
temporal lobes. Trends in Cognitive Sciences, 11, 126–135.
McClelland, J. L., & Rogers, T. T. (2003). The parallel distributed processing
approach to semantic cognition. Nature Reviews Neuroscience, 4, 310–322.
McGeoch, J. A. (1942). The psychology of human learning: An introduction. New York:
Longmans.
McKoon, G., & Ratcli↵, R. (1979). Priming in episodic and semantic memory.
Journal of Verbal Learning and Verbal Behavior, 18, 463–480.
Miller, E. K., & Cohen, J. D. (2001). An integrative theory of prefrontal cortex
function. Annual Review of Neuroscience, 24, 167-202.
Miller, J. F., Kahana, M. J., & Weidemann, C. T. (2012). Recall termination in free
recall. Memory & Cognition, 40, 540–550.
Miller, J. F., Lazarus, E., Polyn, S. M., & Kahana, M. J. (2013). Spatial clustering
during memory search. Journal of Experimental Psychology: Learning, Memory
and Cognition, 39, 773–781.

105
Miller, J. F., Neufang, M., Solway, A., Brandt, A., Trippel, M., Mader, I., . . . Schulze-
Bonhage, A. (2013). Neural activity in human hippocampal formation reveals
the spatial context of retrieved memories. Science, 342, 1111–1114.
Miller, K. J., Leuthardt, E. C., Schalk, G., Rao, R. P. N., Anderson, N. R., Moran,
D. W., . . . Ojemann, J. G. (2007). Spectral changes in cortical surface potentials
during motor movement. Journal of Neuroscience, 27, 2424–2432.
Mitchell, D., McNaughton, N., Flanagan, D., & Kirk, I. (2008). Frontal-midline
theta from the perspective of hippocampal theta. Progress in neurobiology, 86,
156.
Mitchell, K. J., & Johnson, M. K. (2009). Source monitoring 15 years later: What
have we learned from fMRI about the neural mechanisms of source memory?
Psychological Bulletin, 135, 638–677.
Mitchell, T., Shinkareva, S., Carlson, A., Chang, K., Malave, V., Mason, R., & Just,
M. (2008). Predicting human brain activity associated with the meanings of
nouns. Science, 320, 1191.
Mizuhara, H., Wang, L.-Q., Kobayashi, K., & Yamaguchi, Y. (2004). A long-range
cortical network emerging with theta oscillation in a mental task. Neuroreport,
15, 1233–1238.
Morton, N. W., Kahana, M. J., Rosenberg, E. A., Sperling, M. R., Sharan, A. D.,
& Polyn, S. M. (2013). Category-specific neural oscillations predict recall
organization during memory search. Cerebral Cortex, 23, 2407–2402.
Morton, N. W., Kahana, M. J., Rosenberg, E. A., Sperling, M. R., Sharan, A. D., &
Polyn, S. M. (in press). Category-specific neural oscillations predict recall
organization during memory search. Cerebral Cortex.
Mukamel, R., Gelbard, H., Arieli, A., Hasson, U., Fried, I., & Malach, R. (2005,
Aug). Coupling between neuronal firing, field potentials, and fMRI in human

106
 auditory cortex. Science, 309, 951–954.
Murdock, B. B. (1962). The serial position e↵ect of free recall. Journal of Experimental
Psychology, 64, 482-488.
Muthukumaraswamy, S. D. (2013). High-frequency brain activity and muscle
artifacts in meg/eeg: A review and recommendations. Frontiers in Human
Neuroscience, 7.
Nelson, D. L., McEvoy, C. L., & Schreiber, T. A. (2004). The University of South
Florida free association, rhyme, and word fragment norms. Behavior Research
Methods, Instruments and Computers, 36(3), 402-407.
Niessing, J., Ebisch, B., Schmidt, K. E., Niessing, M., Singer, W., & Galuske,
R. A. W. (2005, Aug). Hemodynamic signals correlate tightly with syn-
chronized gamma oscillations. Science, 309, 948–951.
Noppeney, U., Price, C. J., Penny, W. D., & Friston, K. J. (2006). Two distinct neural
mechanisms for category-selective responses. Cerebral Cortex, 16, 437–445.
Nunez, P. L., & Srinivasan, R. (2006). Electric fields of the brain. New York: Oxford
University Press.
Nyhus, E., & Curran, T. (2010, June). Functional role of gamma and theta oscillations
in episodic memory. Neuroscience & Biobehavioral Reviews, 34, 1023-1035.
Ojemann, G. A., Ramsey, N. F., & Ojemann, J. (2013). Relation between functional
magnetic resonance imaging (fmri) and single neuron, local field potential
(lfp) and electrocorticography (ecog) activity in human cortex. Frontiers in
human neuroscience, 7.
Osipova, D., Takashima, A., Oostenveld, R., Fernndez, G., Maris, E., & Jensen,
O. (2006). Theta and gamma oscillations predict encoding and retrieval of
declarative memory. Journal of Neuroscience, 26, 7523–7531.
Paller, K. A., Kutas, M., & Mayes, A. R. (1987). Neural correlates of encoding in an

107
 incidental learning paradigm. Electroencephalography and clinical Neurophysi-
ology, 67, 360-371.
Paller, K. A., McCarthy, G., & Wood, C. C. (1988). ERPs predictive of subsequent
recall and recognition performance. Biological Psychology, 26, 269-276.
Paller, K. A., & Wagner, A. D. (2002). Observing the transformation of experience
into memory. Trends in Cognitive Sciences, 6, 93-102.
Patterson, K., Nestor, P. J., & Rogers, T. T. (2007). Where do you know what you
know? the representation of semantic knowledge in the human brain. Nature
Reviews Neuroscience, 8, 976–987.
Pessoa, L., & Adolphs, R. (2010). Emotion processing and the amygdala: from
a ’low road’ to ’many roads’ of evaluating biological significance. Nature
Reviews Neuroscience, 11, 773–783.
Petrides, M. (2000). Dissociable roles of mid-dorsolateral prefrontal and anterior
inferotemporal cortex in visual working memory. Journal of Neuroscience, 20,
7496.
Pfurtscheller, G., & Lopes Da Silva, F. H. (1999, November). Event-related eeg/meg
synchronization and desynchronization: basic principles. Clinical Neurophys-
iology, 110, 1842-1857.
Polyn, S. M., & Kahana, M. J. (2008). Memory search and the neural representation
of context. Trends in Cognitive Sciences, 12, 24–30.
Polyn, S. M., Natu, V. S., Cohen, J. D., & Norman, K. A. (2005). Category-specific
cortical activity precedes retrieval during memory search. Science, 310, 1963–
1966.
Polyn, S. M., Norman, K. A., & Kahana, M. J. (2009a). A context maintenance and
retrieval model of organizational processes in free recall. Psychological Review,
116, 129–156.

108
Polyn, S. M., Norman, K. A., & Kahana, M. J. (2009b). Task context and organization
in free recall. Neuropsychologia, 47, 2158-2163.
Preston, A. R., & Eichenbaum, H. (2013). Interplay of hippocampus and prefrontal
cortex in memory. Current Biology, 23, R764–R773.
Ray, M. K., Mackay, C. E., Harmer, C. J., & Crow, T. J. (2008). Bilateral generic
working memory circuit requires left-lateralized addition for verbal process-
ing. Cereberal Cortex, 18, 1421–1428.
Rissman, J., & Wagner, A. D. (2012). Distributed representations in memory:
Insights from functional brain imaging. Annual Review of Psychology, 63, 101–
128.
Rizzuto, D., Madsen, J. R., Bromfield, E. B., Schulze-Bonhage, A., Seelig, D.,
Aschenbrenner-Scheibe, R., & Kahana, M. J. (2003). Reset of human neocor-
tical oscillations during a working memory task. Proceedings of the National
Academy of Sciences, USA, 100, 7931–7936.
Rodd, J. M., Johnsrude, I. S., & Davis, M. H. (2012). Dissociating frontotemporal
contributions to semantic ambiguity resolution in spoken sentences. Cerebral
Cortex, 22, 1761–1773.
Rutishauser, U., Ross, I., Mamelak, A., & Schuman, E. (2010). Human memory
strength is predicted by theta-frequency phase-locking of single neurons.
Nature, 464, 903–907.
Sanquist, T. F., Rohrbaugh, J. W., Syndulko, K., & Lindsley, D. B. (1980). Elec-
trocortical signs of levels of processing: perceptual analysis and recognition
memory. Psychophysiology, 17, 568-576.
Sauseng, P., Hoppe, J., Klimesch, W., Gerlo↵, C., & Hummel, F. C. (2007). Dissocia-
tion of sustained attention from central executive functions: local activity and
interregional connectivity in the theta range. European Journal of Neuroscience,

109
 25, 587–593.
Schnitzler, A., & Gross, J. (2005). Normal and pathological oscillatory communica-
tion in the brain. Nature reviews neuroscience, 6, 285–296.
Scoville, W. B., & Milner, B. (1957). Loss of recent memory after bilateral hippocam-
pal lesions. Journal of Neurology, Neurosurgery, and Psychiatry, 20, 11-21.
Sederberg, P. B., Gauthier, L. V., Terushkin, V., Miller, J. F., Barnathan, J. A., &
Kahana, M. J. (2006). Oscillatory correlates of the primacy e↵ect in episodic
memory. NeuroImage, 32, 1422–1431.
Sederberg, P. B., Howard, M. W., & Kahana, M. J. (2008). A context-based theory
of recency and contiguity in free recall. Psychological Review, 115, 893-912.
Sederberg, P. B., Kahana, M. J., Howard, M. W., Donner, E. J., & Madsen, J. R.
(2003). Theta and gamma oscillations during encoding predict subsequent
recall. Journal of Neuroscience, 23, 10809–10814.
Sederberg, P. B., Miller, J. F., Howard, W. H., & Kahana, M. J. (2010). The temporal
contiguity e↵ect predicts episodic memory performance. Memory & Cognition,
38, 689–699.
Sederberg, P. B., Schulze-Bonhage, A., Madsen, J. R., Bromfield, E. B., Litt, B.,
Brandt, A., & Kahana, M. J. (2007). Gamma oscillations distinguish true from
false memories. Psychological Science, 18, 927–932.
Sederberg, P. B., Schulze-Bonhage, A., Madsen, J. R., Bromfield, E. B., McCarthy,
D. C., Brandt, A., . . . Kahana, M. J. (2007). Hippocampal and neocortical
gamma oscillations predict memory formation in humans. Cerebral Cortex,
17, 1190–1196.
Serruya, M. D., Sederberg, P. B., & Kahana, M. J. (2014). Power shifts track serial
position and modulate encoding in human episodic memory. Cerebral Cortex,
24, 403–413.

110
Shrager, Y., Kirwan, C., & Squire, L. (2008). Activity in both hippocampus and
perirhinal cortex predicts the memory strength of subsequently remembered
information. Neuron, 59, 547.
Siddiqui, A. P., & Unsworth, N. (2011). Investigating the role of emotion during
the search process in free recall. Memory & cognition, 39, 1387–1400.
Simons, J. S., & Spiers, H. J. (2003). Prefrontal and medial temporal lobe interactions
in long-term memory. Nature Reviews Neuroscience, 4, 637–648.
Squire, L. S., & Zola, S. M. (1998). Episodic memory, semantic memory and amnesia.
Hippocampus, 8, 205-211.
Staresina, B. P., & Davachi, L. (2006). Di↵erential encoding mechanisms for sub-
sequent associative recognition and free recall. Journal of Neuroscience, 26,
9162–9172.
Staudigl, T., & Hanslmayr, S. (2013). Theta oscillations at encoding mediate the
context-dependent nature of human episodic memory. Current Biology, 23,
1–6.
Steyvers, M., Shi↵rin, R. M., & Nelson, D. L. (2004). Word association spaces for
predicting semantic similarity e↵ects in episodic memory. In A. F. Healy (Ed.),
Cognitive psychology and its applications: Festschrift in honor of Lyle Bourne, Wal-
ter Kintsch, and Thomas Landauer. Washington, DC: American Psychological
Association.
Stoller, A. (1949). Slowing of the alpha-rhythm of the electro-encephalogram and
its association with mental deterioration and epilepsy. The British Journal of
Psychiatry, 95, 972–984.
Strange, B. A., Otten, L. J., Josephs, O., Rugg, M. D., & Dolan, R. J. (2002). Dissociable
human perirhinal, hippocampal, and parahippocampal roles during verbal
encoding. Journal of Neuroscience, 22, 523-528.

111
Summerfield, C., & Mangels, J. A. (2005). Coherent theta-band eeg activity predicts
item-context binding during encoding. NeuroImage, 24, 692-703.
Talmi, D., & Moscovitch, M. (2004). Can semantic relatedness explain the enhance-
ment of memory for emotional words? Memory & Cognition, 32, 742–751.
Thompson-Schill, S. L. (2003). Neuroimaging studies of semantic memory: infer-
ring “how” from “where”. Neuropsychologia, 41, 280–292.
Thompson-Schill, S. L., Bedny, M., & Goldberg, R. (2005). The frontal lobes and the
regulation of mental activity. Current Opinion in Neurobiology, 15, 219–224.
Thompson-Schill, S. L., D’Esposito, M., Aguirre, G. K., & Farah, M. J. (1997, Dec).
Role of left inferior prefrontal cortex in retrieval of semantic knowledge: a
reevaluation. Proceedings of the National Academy of Sciences, USA, 94, 14792–
14797.
Thompson-Schill, S. L., D’Esposito, M., & Kan, I. P. (1999). E↵ects of repetition
and competition on activity in left prefrontal cortex during word generation.
Neuron, 23, 513–522.
Thompson-Schill, S. L., Swick, D., Farah, M. J., D’Esposito, M., Kan, I. P., & Knight,
R. T. (1998). Verb generation in patients with focal frontal lesions: A neuropsy-
chological test of neuroimaging findings. Proceedings of the National Academy
of Sciences, 95, 15855–15860.
Tulving, E. (1972). Episodic and semantic memory. In E. Tulving & W. Donaldson
(Eds.), Organization of memory. (p. 381-403). New York: Academic Press.
Tulving, E. (1983). Elements of episodic memory. New York: Oxford.
Tyler, L. K., & Moss, H. E. (2001). Towards a distributed account of conceptual
knowledge. Trends in cognitive sciences, 5, 244–252.
Underwood, B. J. (1957). Interference and forgetting. Psychological Review, 64,
49-60.

112
Voytek, B., Secundo, L., Bidet-Caulet, A., Scabini, D., Stiver, S. I., Gean, A. D., . . .
Knight, R. T. (2010). Hemicraniectomy: a new model for human electrophys-
iology with high spatio-temporal resolution. Journal of cognitive neuroscience,
22, 2491–2502.
Wagner, A. D., Schacter, D. L., Rotte, M., Koutstaal, W., Maril, A., Dale, A. M., . . .
Buckner, R. L. (1998). Building memories: remembering and forgetting of
verbal experiences as predicted by brain activity. Science, 281, 1188–1191.
Warriner, A. B., Kuperman, V., & Brysbaert, M. (2013). Norms of valence, arousal,
and dominance for 13,915 English lemmas. Behavior research methods, 45,
1191–1207.
Weidemann, C. T., Mollison, M. V., & Kahana, M. J. (2009). Electrophysiological
correlates of high-level perception during spatial navigation. Psychonomic
Bulletin & Review, 16, 313–319.
Wilson, M. D. (1988). The MRC psycholinguistic database: Machine readable
dictionary, version 2. Behavior Research Methods, Instruments & Computers, 20,
6-11.
Ya↵e, R. B., Kerr, M. S., Damera, S., Sarma, S. V., Inati, S. K., & Zaghloul, K. A.
(2014). Reinstatement of distributed cortical oscillations occurs with precise
spatiotemporal dynamics during successful memory retrieval. Proceedings of
the National Academy of Sciences, 111, 18727–18732.
Yonelinas, A. P. (2013). The hippocampus supports high-resolution binding in the
service of perception, working memory and long-term memory. Behavioural
brain research, 254, 34–44.
Yuval-Greenberg, S., Tomer, O., Keren, A. S., Nelken, I., & Deouell, L. Y. (2008).
Transient induced gamma-band response in EEG as a manifestation of minia-
ture saccades. Neuron, 58, 429–441.

113
Zaromb, F. M., Howard, M. W., Dolan, E. D., Sirotin, Y. B., Tully, M., Wingfield, A.,
& Kahana, M. J. (2006). Temporal associations and prior-list intrusions in free
recall. Journal of Experimental Psychology: Learning, Memory, and Cognition, 32,
792–804.

114