Modelling distributed working memory in a

large-scale network model of Human brain

Mengli Feng

A Thesis under the supervision of Jorge F Mejias

for the Degree of

Mres in Brain and Cognitive Science

Instituut voor Interdisciplinaire Studies

University of Amsterdam

Nov, 2022
©Copyright by Mengli Feng 2022
All Rights Reserved
 Abstract
Recent scientific development has led to opportunities to study the computational
mechanisms of cognitive functions which involve multiple interacting brain areas.
One of the cognitive functions brought to the centre of the stage by this trend of
research is working memory. Evidence from neural recordings has shown co-occurrent
activities of different brain regions during working memory. To better determine the
role of the coordination of different brain areas in working memory and its underlying
mechanism, a computational model of the human cortical network was built. The
model is constrained by anatomical data in both long-range connections and local
circuitry. By adding gradient excitation across cortical hierarchy, attractor states were
observed within the large-scale cortical model, which explains distributed working
memory. The model opens up opportunity for future research of working memory
mechanisms in human brain and clinical applications.
Acknowledgments
I would like thank Dr. Abhirup Bandyopadhyay for working with me on this
project, and members of CSN lab for their support and help.

i
ii
Contents

1 Introduction 1
1.1 Prior research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Research Questions/ Hypothesis/ Intended Results . . . . . . . . . . 3

2 Method 5
2.1 Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Cortical Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Feedforward/Feedback Projections . . . . . . . . . . . . . . . 6
2.1.3 Spine Counts . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.4 FLN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.5 SLN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.6 CIB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.7 Gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.8 Receptor Density . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 anatomical data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.1 connectivity strength . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.2 hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.3 feedforward/feedback projections . . . . . . . . . . . . . . . . 7
2.2.4 Gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Model overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 Simulations and analysis . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4.1 Sensory input . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

iii
 2.4.2 working memory tasks . . . . . . . . . . . . . . . . . . . . . . 13
2.4.3 Hierarchy linearity . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4.4 functional connectivity . . . . . . . . . . . . . . . . . . . . . . 13
2.4.5 macaque model . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5 Model Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.1 Local Circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.2 Gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.5.3 Interareal connection . . . . . . . . . . . . . . . . . . . . . . . 16
2.5.4 Hemodynamic transformation . . . . . . . . . . . . . . . . . . 18

3 Results 19
3.1 Distributed working memory in human brain . . . . . . . . . . . . . . 19
3.2 𝐺 − 𝛼 parameter space . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3 Robust working memory to distractors . . . . . . . . . . . . . . . . . 22
3.4 Stronger functional connectivity in higher areas . . . . . . . . . . . . 22
3.5 Comparison to macaque model . . . . . . . . . . . . . . . . . . . . . 24
3.6 Effect of hierarchy linearity . . . . . . . . . . . . . . . . . . . . . . . 25

4 Discussion 29
4.1 Limitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.2 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

A Appendix 33

B Appendix 35

iv
List of Figures

2-1 FF ratio derivation process. A. macaque FF strength. The left panel

shows the ratio of FF projections for all macaque cortical areas, sorted
by hierarchy values. The right panel plots FF ratio against the hierar-
chy values for each target areas. color coding indicating the hierarchy
values of for different source areas (red-green-purple: low to high) B.
different FF derivation scheme from human T1:T2 data. The first pair
shows the convergent scheme with a linear hierarchy. The second pair
shows the convergent scheme with a non-linear hierarchy. The third
pair shows the divergent scheme with a nonlinear hierarchy. C. final
FF derivation, which is a divergent scheme with linear hierarchy . . . 9

2-2 Model Schematics for human model and demonstration of neural dy-
namics from previous macaque model. A. Local Circuit with two ex-
citatory and one inhibitory population B. Interareal connections with
excitation-dominated feedforward projections and inhibition-dominated
feedback projections. C. Bifurcation diagram for local circuit plotted
against different 𝐽𝑆 values and the neural dynamics of the circuit for a
𝐽𝑆 value above the bifurcation point. D. Firing rates temporal dynam-
ics across cortical areas (Macaque model). E. Macaque cortical areas.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

v
2-3 Anatomical Data for human model. A. T1:T2 ratio. B. NMDA recep-
tor density. left panel shows the map for parcellation in [14]. Right
panel shows the map we generated for the parcellation we are using
in the current study [7]. C. parcellation used for current study. D.
Correlation between T1:T2 ratio and NMDA receptor density. E. con-
nectivity strength for 180 human cortical areas. . . . . . . . . . . . . 12

3-1 Distributed working memory in human brain. A. spatial activity map

during visual stimulation (left) and delay period (right). B. Firing rates
across areas sorted by hierarchy. C. Activity of selected areas during
working memory tasks. D. Activity of areas in eight different functional
network, including three sensory (AUD, auditory; VIS, visual; and
SOM, somatomotor) and five association (DAN, dorsal attention; FPN,
frontoparietal; VAN, ventral attention; DMN, default mode; and CON,
cingulo-opercular) networks. . . . . . . . . . . . . . . . . . . . . . . 20

3-2 Activity spatial map across parameter space. A. activity map with
different 𝛼 (0 - 1) and 𝐺 (0.5 - 1) values. B. number of areas in
working memory states during delay period across parameter space
(𝐺0.6 = 1.2, 𝛼 = 0 − 1). C. map of experimental evidence. . . . . . . 21

3-3 Working memory tasks with distractors. A. working memory task with
an auditory-go-signal. B. working memory task with a visual distrac-
tor. The top panel shows the scheme of the task. The middle panel
shows the temporal dynamics of the input areas (A1 and V1). The
bottom panel shows the temporal dynamics of the higher areas (PGS) 23

3-4 Functional connectivity for resting and working memory states. A.

scheme of hemodynamic transformation. B. functional connectivity in
resting state. C. functional connectivity iin working memory state . . 24

vi
3-5 Comparison between macaque and human data. A. macaque anatomi-
cal data. B. human anatomical data. From left to right are: 1) connec-
tivity strength matrix, 2) FF strength matrix, 3) FF strength plotted
against hierarchy values of the target areas, 4) Hierarchy values sorted
from low to high . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3-6 Comparison between macaque and human simulation results. A. sim-
ulation results for macaque model. B. simulation results for human
model. C. box-plot of firing rates in five cortical lobes from the simula-
tions results of the two models. PRE: prefrontal, FON: frontal, PAR:
parietal, TEM: temporal, OCC: occipital . . . . . . . . . . . . . . . . 26
3-7 Effect of hierarchy linearity on simulation results. A. hierarchy without
nonlinear transformation. B. hierarchy with non-linear transformation
(𝐻 = 𝑙𝑜𝑔10(ℎ + 𝑏) for 𝑏 = 0.1). C. hierarchy with non-linear transfor-
mation (𝑏 = 0.05) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

A-1 Input to V1 or multiple visual areas. A. Input to V1. B. Input to

multiple visual areas. From top to bottom: 1. location of the input
areas, 2. hierarchy of the input areas, 3. temporal dynamics across
areas with inputs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

B-1 Literature collected for generating the evidence map . . . . . . . . . . 36

vii
viii
List of Tables

ix
x
Chapter 1

Introduction

1.1 Prior research

With neurophysiological recording advances such as Neuropixels, neural activity
can be more easily recorded simultaneously across different areas in animal brain
[15, 26]. This has led to opportunities to study the computational mechanisms of
cognitive functions which involve multiple interacting brain areas. One of the cogni-
tive functions brought to the centre of the stage by this trend of research is working
memory, which is defined as the cognitive ability to retain task-related information
temporarily [13, 30, 12]. Evidence from neural recordings has shown co-occurrent ac-
tivities of different brain regions during working memory for both humans and other
primates [5].
Recent experimental and/or computational work has identified distributed work-
ing memory phenomena in mice [28] and macaque [19]. However, it is not clear how
distributed working memory mechanisms would emerge in the human brain. In this
work, I present a mechanistic model of distributed working memory in the human
brain. The model includes a local circuits with sustained neural activity after re-
ceiving brief stimuli [32] and is constrained by anatomical data in both long-range
connections and local circuitry. By adding gradient excitation across cortical hierar-
chy, attractor states were observed across cortical areas in simulations, which explains
distributed working memory. Based on the success of the macaque model, the cur-

1
rent project aims at building a large-scale cortical network model to simulate human
working memory.

1.2 Objectives

The fist objective of the project is to build the human brain model using biological
data to constrain it. For basic elements of the model, such as local circuit dynamics,
we follow the steps of a previous model of distributed working memory in macaques
[19]. However, the different nature of the data in other aspects, such as the large-scale
connectivity, leads to several challenges in the model building compare to the case of
the macaque. To begin with, there is no direct data on gradient of excitatory synap-
tic strength across cortical areas for human. One possibility is to use the hierarchy
inferred by T1/T2 weighted MRI data [7], which was proven to be an accurate proxy
to anatomical hierarchy in primate’s brain [3], and then assume that this anatomical
hierarchy is a good proxy for the gradient of excitatory strength [4, 19]. Another op-
tion is to approximate the gradient using recently measured NMDA densities cortical
across areas [14]. Second, there are much more areas in human brain compared to
macaque brain, with very different topology, which may require different treatment
to the model. In the end, we expect the model to be adjusted to a working regime,
where distributed working memory patterns emerge. The second objective is to ap-
proach the experimental evidence of fMRI patterns during working memory tasks,
which may requires further treatment of the data and tuning of the parameters. The
third objective is to use the model to explore the mechanism of how working memory
of an object can be decoded from activity of early visual areas. Previous experimen-
tal study showed late enhancement of the sensory response after receiving a stimuli
[27]. Specifically, we would like to see how long-range feedback from higher areas
contribute to this phenomenon.

2
1.3 Research Questions/ Hypothesis/ Intended Re-
sults
The main research question in the project is whether and how a large-scale cortical
model can account for distributed working memory in human brain. This leads to the
following sub-questions: 1. how human data affect the model behaviour and whether
the model would behave similarly to the macaque model. 2. Do the same hypothesis
in macaque model hold for human model?
It is hypothesized that with hierarchical inter-areal and local connections based
on neuroimaging data, the model can simulate distributed neural activities during
working memory tasks. The expectation is that the simulations from the model should
be able to produce results that are comparable to experimental results, although
we also expect some discrepancies because of limitations in knowledge of human
connectomes.

3
4
Chapter 2

Method

2.1 Concepts

Before introducing the model, I would like to define some key concepts to lay a
foundation for explaining the model.

2.1.1 Cortical Hierarchy

Cortical hierarchy is defined as a topological organization of laminar regularities

of interareal connectivity [18]. Early research found that rostral directed projections
are from supragranular layer to granular layer while caudal directed pathways are
from intragranular layer to granular layer. This relationship between projection di-
rectionality and interareal laminar connections between two pathways suggests that
there is a distinction between the two kinds of pathways defined by its directional-
ity. Furthermore, there is a topological organization of cortical areas according to
the strength of supragranular and intragranular inputs that each area receives. It is
therefore assumed that there is a cortical hierarchy, where higher areas receives more
input from supragranular layers of lower areas while lower areas receive more input
from intragranular layers of higher areas.

5
2.1.2 Feedforward/Feedback Projections

Following the idea of cortical hierarchy, the projections from lower to higher areas
is conceptualized as feedforward (FF), while the projections from higher to lower
areas is considered feedback (FB).

2.1.3 Spine Counts

Spine counts are the measurement of the number of apical dendritic spines.

2.1.4 FLN

In macaque model, the connectivity strengths are adopted from track tracing data,
which measure the fraction of neurons in the target area labeled by retrograde tracer
injected at the source area, or in short, the fraction of labeled neurons (FLN) [19].

2.1.5 SLN

In macaque model, the feedforward connection strengths were derived from the
measurement of the number of labeled neurons located on the supragranular layer
of a given source area in a tract tracing study. Dividing this number over the total
number of labeled neurons from a source area, we can define the supragranular layered
neurons (SLN) as the feedforward strength from that source area to the target area.

2.1.6 CIB

CIB stands for counter-stream inhibitory bias, which refers to gradual preferential
targeting onto inhibitory neurons by top-down (feedback) projections. In the macaque
model, CIB is expressed in 1 − 𝑆𝐿𝑁 .

2.1.7 Gradient

In the current project, gradient is defined as the heterogeneity of outgoing and

recurrent synaptic strengths of cortical areas along the hierarchy [19]. In macaque

6
model, the gradient is taken from direct measurement of spine counts, supplemented
by hierarchy value when spine counts data is not available.

2.1.8 Receptor Density

It refers to density of postsynaptic receptors on measured (target) area. In the

human model, NMDA receptor density across cortical areas can be used as an ap-
proximate of gradient.

2.2 anatomical data

2.2.1 connectivity strength

The overall connection strengths between areas are established upon white matter
connectome data measured from diffusion imaging. [7]. Since it is diffusion imaging
data, it is unidirectional, which is different from bidirectional macaque connectivity
data.

2.2.2 hierarchy

Unfortunately, there is no tract tracing data like SLN available for human brain.
Prior research shows an negative correlation between T1w/T2w MRI map values and
cortical hierarchy in macaque [3]. In the current study, we used the hierarchy values
derived from T1w/T2w map by Demirtas et al [7] as an estimation of hierarchy.

2.2.3 feedforward/feedback projections

As mentioned, there is no tract tracing data available to extract feedforward and

feedback projection strength for human brain. Since cortical hierarchy is defined
by organization of FF/FB projections, we decided to use T1w/T2w based hierarchy
values to derive FF strength.

7
 Specifically, we make the following assumptions: 1. when the target area are at
the same hierarchy as the source area, the FF strength from source to target is 0.5.
2. as the hierarchical distance between target and source areas increases, depend-
ing on whether source area is lower/higher than the target area, the FF strength
increases/decreases from 0.5.
Based on those two basic assumptions, we derived two versions of FF and FB
maps. The first one further assumes that for source areas at different hierarchies, the
relationship between hierarchical distance and FF strength is always linear. This leads
to a double convergence scheme as shown in . For the second version, we allow step-
wise nonlinear relationship between hierarchical distance and FF strength when source
area in not at 0.5 hierarchy. In this scheme, when source area moves up hierarchy
from 0.5, the FF strength drops/increases faster with increasing hierarchical distance,
leading to increasing areas with FF strengths of 0/1. This goes in the opposite way
when source area moves down hierarchy from 0.5. Intuitively, it means when an area
moves up hierarchy, it gives increasingly less FF inputs to target areas. As as shown
in , such scheme behaves in a double-divergent way.
To choose a more appropriate scheme, we made a comparison between the hier-
archy - FF strength relation defined here and that for macaque, shown in fig. The
idea is to find a scheme that approximates the relation used in macaque model. To
avoid the influence of hierarchy linearity, we made a non-linear transformation to the
human cortical hierarchy so that it resembles macaque cortical hierarchy. As can be
seen, the second scheme is more close to the relation for macaque, we therefore decide
to use it in the current model.

2.2.4 Gradient

For human brain, there is no spine counts data available for approximating the
gradient. Recently, Hansen et al. [14] measured NMDA receptor density in human
brain. Since glutamate binds to NMDA receptors to produce excitatory responses,
we can use NMDA receptor density as a proxy for gradient.
The NMDA receptor density map for parcellation with 34 areas [14] (Fig 2-3 B)

8
 A
FF

B
FF FF

C FF

Figure 2-1: FF ratio derivation process. A. macaque FF strength. The left panel
shows the ratio of FF projections for all macaque cortical areas, sorted by hierarchy
values. The right panel plots FF ratio against the hierarchy values for each target
areas. color coding indicating the hierarchy values of for different source areas (red-
green-purple: low to high) B. different FF derivation scheme from human T1:T2
data. The first pair shows the convergent scheme with a linear hierarchy. The second
pair shows the convergent scheme with a non-linear hierarchy. The third pair shows
the divergent scheme with a nonlinear hierarchy. C. final FF derivation, which is a
divergent scheme with linear hierarchy
.

is mapped to the 180 area parcellation for hierarchy values [7] (Fig 2-3 A).

2.3 Model overview

The model consists of multiple brain areas with same local circuit (although with
area-specific values for some parameters, to incorporate local heterogeneity due to
cortical gradients).

9
 The local circuit comes from a firing rate model is based on a firing rate model for
working memory and decision making [32], which includes two excitatory population
and one inhibitory population. During a task, the stimulus goes to the excitatory
populations and the one which received stronger input will win and reach high activity
while the other population will be silent, as shown in fig. The high firing activity of
the winning population will persist in the delay period after the input is gone, which
resembles working memory states. Three important ingredient for working memory in
the local circuit are: 1. Shared inhibition. The two excitatory populations competes
through the common inhibition, resulting in one winner population having sustained
firing activity which gives rise to working memory. 2. Slow recurrent dynamics. The
dynamics is slow enough so that evidence accumulation can be account for, meaning
if the input is too short, the working memory state is not reached. 3. Recurrent
excitation. Only if 𝐽𝑠 /𝐽𝑐 is large enough that the excitation is stronger than the
feedback inhibition. When 𝐽𝑠 exceeds a certain threshold, the winner-take-all decision
making and the follow-up working memory state emerge. This can be understood as
the system reaching bifurcation point, as show in fig. In the macaque model, the
bifurcation occurs once 𝐽𝑆 exceeds around 0.46.

In a large-scale model, we assume that distributed working memory can emerge

even if 𝐽𝑆 is below the bifurcation point for most of the areas, because of the interareal
connections, cortical hierarchy and gradient of synaptic strengths. For the current
model, areas are interconnected as shown in Fig 1 B, where excitatory populations
in one area project to both excitatory and inhibitory populations in another area.
Following cortical hierarchy, it is hypothesized here that the feedforward projections
are dominated by excitation (targeting excitatory populations) while feedback projec-
tions are predominantly inhibitory (targeting inhibitory populations). Gradients are
introduced to both local excitatory recurrent strength 𝐽𝑆 and inter-areal connection
strength.

10
 A Background input
C Input period Delay period
Js
Js
E I E

External Input Jc
D E
B

Figure 2-2: Model Schematics for human model and demonstration of neural dy-
namics from previous macaque model. A. Local Circuit with two excitatory and
one inhibitory population B. Interareal connections with excitation-dominated feed-
forward projections and inhibition-dominated feedback projections. C. Bifurcation
diagram for local circuit plotted against different 𝐽𝑆 values and the neural dynamics
of the circuit for a 𝐽𝑆 value above the bifurcation point. D. Firing rates temporal
dynamics across cortical areas (Macaque model). E. Macaque cortical areas.
.

2.4 Simulations and analysis

2.4.1 Sensory input

To adapt to the increasing number of areas compared to macaque model, sensory

input is sent to multiple early visual areas, including VM1, MT, V1, V3A, V6A, V4
and VMV3. Such choice is also motivated by the fact that thalamus also projects to
higher visual areas in animal models [2]. The input current is given to one excitatory

11
 A B

C D E

Figure 2-3: Anatomical Data for human model. A. T1:T2 ratio. B. NMDA receptor
density. left panel shows the map for parcellation in [14]. Right panel shows the
map we generated for the parcellation we are using in the current study [7]. C.
parcellation used for current study. D. Correlation between T1:T2 ratio and NMDA
receptor density. E. connectivity strength for 180 human cortical areas.
.

population of the local circuit in those areas with an intensity of 0.3 nA for 0.5 second.
See appendix Fig A-1

12
2.4.2 working memory tasks

A simple visual delayed response task task is used to test the model performance,
where a transient (0.5 second) excitation is given to a selective neural pool of the
chosen visual areas and observations are made on the delay period. To further eval-
uate the robustness of the model, we also simulated two more complicated working
memory tasks that are comparable to real experiments, shown in fig. For both tasks,
a participant is asked to remember the position (left or right) of a triangle shown on
a screen. After a fixation period, the stimulus is shown on the screen for a brief time
(0.5 second) and disappears. There is a delay period of 5 seconds, then the participant
is either cued by a brief (0.1 second) binaural auditory stimulus (auditory go-signal)
or distracted by a visual stimulus shown on the opposite side of the screen (distrac-
tor) before asked to report the position of the visual stimuli they originally see. For
the auditory-go-signal, we give inputs (c = 0.3 nA) to both excitatory populations
of seven auditory areas (‘A1’, ’RI’, ’TA2’, ’PBelt’, ’MBelt’, ’LBelt’, ’A4’). For the
visual distractor, we give inputs (c = 0.3 nA) to the opposing excitatory population
of the visual areas.

2.4.3 Hierarchy linearity

To investigate the effect of hierarchy linearity, we added a nonlinear transformation

to the hierarchy values with the following equation: 𝐻 = 𝑙𝑜𝑔10(ℎ + 𝑏) where 𝐻 is the
hierarchy value after the transformation, ℎ is the original hierarchy value and 𝑏 the
transformation coefficient. 𝑏 = 0.1 and 𝑏 = 0.05 were used for comparison.

2.4.4 functional connectivity

The functional connectivity is calculated from cross correlations between time se-
ries of BOLD signals derived from that of conductance for each pair of areas. See
2.5.4 for description of the balloon model for hemodynamic transformation from con-
ductance to BOLD signals.

13
2.4.5 macaque model

The macaque model is taken from the recent study by Mejias and Wang to compare
with the human model [19]. It includes 30 areas in macaque neocortex. The local
circuit and interareal connection rules are identical with the ones used in human.
Ratio of feedforward and feedback projection strength are determined by SLN 2.1.5.
The connectome is taken from tract-tracing data (see FLN 2.1.4). Hierarchy is based
on spine counts 2.1.3 data. For areas where spine counts data are not available,
hierarchy values derived from SLN are used. For simulations, inputs go to V1 with
same length and strength as in human model.

2.5 Model Specification

2.5.1 Local Circuit

The local circuit is based on the Wong and Wang’s decision making model [32].
The three variable version is adopted with one inhibitory population and two selec-
tive excitatory populations (2-2 A). The three populations are fully connected and
with self-recurrent connections. The interactions of the populations are described as
follows:

𝑑𝑆𝐴 𝑆𝐴
=− + 𝛾 (1 − 𝑆𝐴 ) 𝑟𝐴 (2.1)
𝑑𝑡 𝜏𝑁

𝑑𝑆𝐵 𝑆𝐵
=− + 𝛾 (1 − 𝑆𝐵 ) 𝑟𝐵 (2.2)
𝑑𝑡 𝜏𝑁

𝑑𝑆𝐶 𝑆𝐶
=− + 𝛾𝐼 𝑟𝐶 (2.3)
𝑑𝑡 𝜏𝐺
𝑆𝐴 , 𝑆𝐵 and 𝑆𝐶 are the NMDA conductance of selective excitatory populations and
GABA conductance of the inhibitory population respectively. The time constants are
𝜏𝑁 = 60𝑚𝑠, 𝜏𝐺 = 5𝑚𝑠, 𝛾 = 1.282 and 𝛾𝐼 = 2. 𝑟𝐴 , 𝑟𝐵 and 𝑟𝐶 are mean firing rates of
excitatory and inhibitory populations, given as follows:

14
 𝑑𝑟𝑖
𝜏𝑟 = −𝑟𝑖 + 𝜑𝑖 (𝐼𝑖 ) (2.4)
𝑑𝑡

The time contant 𝑡𝑎𝑢𝑟 = 2𝑚𝑠. The currents 𝐼𝑖 are given by:

𝐴
𝐼𝐴 = 𝐽𝑆 𝑆𝐴 + 𝐽𝐶 𝑆𝐵 + 𝐽𝐸𝐼 𝑆𝐶 + 𝐼0𝐴 + 𝐼𝑛𝑒𝑡 + 𝑥𝐴 (𝑡) (2.5)

𝐵
𝐼𝐵 = 𝐽𝐶 𝑆𝐴 + 𝐽𝑆 𝑆𝐵 + 𝐽𝐸𝐼 𝑆𝐶 + 𝐼0𝐵 + 𝐼𝑛𝑒𝑡 + 𝑥𝐵 (𝑡) (2.6)

𝐶
𝐼𝐶 = 𝐽𝐼𝐸 𝑆𝐴 + 𝐽𝐼𝐸 𝑆𝐵 + 𝐽𝐼𝐼 𝑆𝐶 + 𝐼0𝐶 + 𝐼𝑛𝑒𝑡 + 𝑥𝐶 (𝑡) (2.7)

The parameter values are 𝐽𝑠 = 0.3213𝑛𝐴, 𝐽𝑐 = 0.0107𝑛𝐴, 𝐽𝐼𝐸 = 0.15𝑛𝐴, 𝐽𝐸𝐼 =

−0.31𝑛𝐴, 𝐽𝐼𝐼 = −0.12𝑛𝐴, 𝐼0𝐴 = 𝐼0𝐵 = 0.3294𝑛𝐴 and 𝐼0𝐶 = 0.26𝑛𝐴. 𝐽𝑆 , 𝐽𝐶 and
𝐽𝐸𝐼 are the self- and cross-coupling between excitatory populations, and the coupling
between the inhibitory population to any of the excitatory ones. 𝐽𝐼𝐼 : the self cou-
pling strength of the inhibitory population. 𝐼0𝐴 , 𝐼0𝐵 and 𝐼0𝐶 are background inputs to
each population.𝐼𝑛𝑒𝑡 are the long-range input coming from other areas in the network,
which we will keep as zero for now but will be detailed later. Sensory stimulation
can be introduced here as extra pulse currents of strength 𝐼𝑝𝑢𝑙𝑠𝑒 = 0.3 and duration
𝑇 𝑝𝑢𝑙𝑠𝑒 = 0.5𝑠𝑒𝑐. 𝑥𝐴 , 𝑥𝐵 and 𝑥𝐶 are noises produced from Ornstein-Uhlenbeck pro-
cess, which introduces some level of stochasticity in the system. It is given by the
following equation:

𝑑𝑥𝑖 √
𝜏𝑛𝑜𝑖𝑠𝑒 = −𝑥𝑖 + 𝜏𝑛𝑜𝑖𝑠𝑒 𝜎𝑖 𝜉𝑖 (𝑡) (2.8)
𝑑𝑡

𝜎𝑖 , 𝜎𝐵 and 𝜎𝐶 are the noise strength for population A, B and C. Parameter values
are 𝜏𝑛𝑜𝑖𝑠𝑒 = 2𝑚𝑠, 𝜎𝐴,𝐵 = 0.005𝑛𝐴, 𝜎𝐶 = 0.

The transfer functions 𝜑𝑖 for the currents of different populations are given as
follows:
1 𝑎𝐼 − 𝑏
𝜑𝐴,𝐵 (𝐼) = (2.9)
2 1 − exp[−𝑑(𝑎𝐼 − 𝑏)]

15
 [︃ ]︃
1
𝜑𝐶 (𝐼) = (𝑐1 𝐼 − 𝑐0 ) + 𝑟0 (2.10)
𝑔𝐼 +

The parameter values for equation 2.9 are 𝑎 = 135𝐻𝑧/𝑛𝐴, 𝑏 = 54𝐻𝑧 and 𝑑 =
0.308𝑠. The parameter values for equation 2.10 are 𝑔𝐼 = 4, 𝑐1 = 615𝐻𝑧/𝑛𝐴, 𝑐0 =
177𝐻𝑧 and 𝑟0 = 5.5𝐻𝑧. The notation [𝑥]+ denotes rectification.

2.5.2 Gradient

The synaptic strength 𝐽𝑠 for different cortical areas are regulated by hierarchy
values (ℎ𝑖 ) as follows:
𝐽𝑆 (𝑖) = 𝐽min + (𝐽max − 𝐽min ) ℎ𝑖 (2.11)

Here, 𝐽𝑚𝑖𝑛 is set to 0.205. To maintain the excitatory input (and therefore the
spontaneous activity level S) constant while varying Js across areas, we have to keep
the quantity 𝐽𝑆 + 𝐽𝐶 + 2𝐽𝐸𝐼 𝐽𝐼𝐸 𝜁 ≡ 𝐽0 constant, from which we derived the following
relationship between the synaptic strength from excitatory population to inhibitory
population (𝐽𝐼𝐸 ) and the recurrent excitatory synaptic strength (J𝑆 ).

1
𝐽𝐼𝐸 = (𝐽0 − 𝐽𝑆 − 𝐽𝐶 ) (2.12)
2𝐽𝐸𝐼 𝜁

𝜏𝐺 𝛾𝐼 𝑐1
𝜁= (2.13)
𝑔𝐼 − 𝐽𝐼𝐼 𝜏𝐺 𝛾𝐼 𝑐1

2.5.3 Interareal connection

For interareal connections, excitatory populations (𝐴 or 𝐵) in a source area re-

ceives feedforward (𝐹 𝐹 ) input from the correspondent excitatory population (𝐴 or
𝐵) in each target area, while the inhibitory population (𝐶) in a source area receives
feedback (1 − 𝐹 𝐹 ) input from both excitatory populations (𝐴 and 𝐵) of each target
areas.
As specified in the following equations, excitatory population 𝐴/𝐵 in a target area
𝑥
receives the sum of feedforward current input 𝐼𝐴/𝐵,𝑛𝑒𝑡 from excitatory population 𝐴/𝐵

16
of each source area, which are the product of connectivity strength from source area 𝑦
to target area 𝑥 (𝑊 𝑥𝑦 ), ratio of feedforward strength from source area to target area
𝑦
(𝐹 𝐹 𝑥𝑦 ) and the conductance of the source area (𝑆𝐴/𝐵 ). This sum of inputs is then
multiplied by the global coupling strength 𝐺. For the inhibitory population in an
𝑥
target area, the current input 𝐼𝐶,𝑛𝑒𝑡 is a sum of the product of connectivity strength
between source area and each target area, the feedback ratio ((1 − 𝐹 𝐹 𝑥𝑦 ) between
source area and each target area, and the sum of the conductance of the excitatory
populations of each target area (𝑆𝐴𝑦 + 𝑆𝐵𝑦 ). This is regulated by global coupling
strength, the E-I balancing factor 𝑧 and the counterstream inhibitory strength, 𝛼.

𝑥
𝑊 𝑥𝑦 𝐹 𝐹 𝑥𝑦 𝑆𝐴𝑦
∑︁
𝐼𝐴,𝑛𝑒𝑡 =𝐺 (2.14)
𝑦

𝑥
𝑊 𝑥𝑦 𝐹 𝐹 𝑥𝑦 𝑆𝐵𝑦
∑︁
𝐼𝐵,𝑛𝑒𝑡 =𝐺 (2.15)
𝑦

𝐺 ∑︁ 𝑥𝑦
𝑥
𝐼𝐶,𝑛𝑒𝑡 =𝛼 𝑊 (1 − 𝐹 𝐹 𝑥𝑦 ) (𝑆𝐴𝑦 + 𝑆𝐵𝑦 ) (2.16)
𝑧 𝑦

The E-I balancing factor 𝑧 = 0.805 guarantees that, if populations 𝐴 and 𝐵

have the same activity level, their net effect on other areas will be zero, therefore
highlighting the selectivity aspect of the circuits. The calculation is listed below and
the detailed reasoning and derivation can be found in previous studies [19].

2𝑐1 𝜏𝐺 𝛾𝐼 𝐽𝐸𝐼
𝑍= (2.17)
𝑐1 𝜏𝐺 𝛾𝐼 𝐽𝐼𝐼 − 𝑔𝐼

For connectivity strength 𝑊 , we rescale the original connectivity strength 𝐷(measured

by diffusion imaging), by factor 𝑘1 = 1.2 and 𝑘2 = 0.3, specified as follows:

𝑊 𝑥𝑦 = 𝑘1 (𝐷𝑥𝑦 )𝑘2 (2.18)

17
2.5.4 Hemodynamic transformation

For hemodynamic transformation from conductance to blood-oxygen-level depen-

dent (BOLD) signal, we use Balloon-Windkessel model [11], a mechanistic model for
the hemodynamic response. The equations and parameters are taken from a previous
paper [7], listed as following:

𝑑𝑥(𝑡)
= 𝑆 𝐸 (𝑡) − 𝑘𝑥(𝑡) − 𝛾(𝑓 (𝑡) − 1) (2.19)
𝑑𝑡

𝑑𝑓 (𝑡)
= 𝑥(𝑡) (2.20)
𝑑𝑡

𝑑𝑣(𝑡) 1
𝜏 = 𝑓 (𝑡) − 𝑣 𝛼 (𝑡) (2.21)
𝑑𝑡

𝑑𝑞(𝑡) 𝑓 (𝑡)
[︂ (︂ )︂]︂
1 (︁ 1
)︁
𝜏 = 1 − (1 − 𝜌) 𝑓 (𝑡) − 𝑞 𝑣 𝛼 −1 (𝑡) (2.22)
𝑑𝑡 𝜌

[︃ (︃ )︃ ]︃
𝑞(𝑡)
𝑦(𝑡) = 𝑉0 𝑘1 (1 − 𝑞(𝑡)) + 𝑘2 1− + 𝑘3 (1 − 𝑣(𝑡)) (2.23)
𝑣(𝑡)

𝑆 𝐸 is excitatory synaptic gating variable, the conductance of excitatory population

which receives input. 𝑥 is the vasodililatory signal. 𝑓 is the blood inflow. 𝑣 is the
blood volume. 𝑞 is the deoxyhemoglobin content. 𝜌 = 0.34 is the resting oxygen
extraction fraction. 𝜏 = 0.98 is the hemodynamic transit time. 𝜅 = 0.65𝑠− 1 is the
rate of signal decay. 𝑘1 = 3.72, 2˛ = 0.53, 3˛ = 0.53 are three dimensionless magnetic
field strength-dependent parameter values. 𝛾 = 0.41𝑠− 1 is the rate of flow-dependent
elimination. 𝛼 = 0.32 is the Grubb’s exponent. 𝑉0 = 0.02 is the resting blood volume
fraction.

18
Chapter 3

Results

3.1 Distributed working memory in human brain

Our first goal is to simulate a visual stimulation task to determine whether per-
sistent firing rate activity emerges in our human brain model (see Methods), and
whether such activity is local or distributed. Our simulations show that, for certain
combinations of parameters, our simulation results show distributed working mem-
ory in the human brain. With 𝐺 = 0.72 and 𝛼 = 0.5, most of the brain areas reach
bistable states in the delay period of the visual response task when 𝐽𝑆 for all areas are
below bifurcation point (𝐽𝑚 𝑎𝑥 = 0.28). The resultant firing rates are quasi-linearly
distributed across a range from 0 to 57 spikes/second, which is consistent with the
cortical hierarchy and gradient we applied to the model. Fig 3-1 A shows the firing
rate patterns across cortical areas during input period (left) and delay period (right),
from which we can see that the areas higher in the hierarchy are able to maintain
a certain level of persistent activity while the early visual areas go silent during the
delay period. In fig, we can see that there are differences among the firing patterns in
different functional networks. Specifically, cinguo-opercular network (CON), dorsal
attention network (DAN) and frontoparietal network (FPN) show higher activities
than the other networks.

19
A B C

Figure 3-1: Distributed working memory in human brain. A. spatial activity map
during visual stimulation (left) and delay period (right). B. Firing rates across areas
sorted by hierarchy. C. Activity of selected areas during working memory tasks. D.
Activity of areas in eight different functional network, including three sensory (AUD,
auditory; VIS, visual; and SOM, somatomotor) and five association (DAN, dorsal
attention; FPN, frontoparietal; VAN, ventral attention; DMN, default mode; and
CON, cingulo-opercular) networks.
.

3.2 𝐺 − 𝛼 parameter space

To find a working regime of the model, we tried adjusting three parameters: the
global coupling strength (𝐺), the counterstream inhibition strength (𝛼), and the
maximum value of the gradient, 𝐽𝑚 𝑎𝑥. After finding a parameter sets that worked,
we explored the parameter space over 𝐺 and 𝛼 to find out how those two parameters
affect the firing rate patterns and in which regime the system behaviour most similar
to experimental evidence. Fig 3-2 A and B shows the firing rate patterns across the

20
parameter space of 𝐺 and 𝛼. In general, the firing rates and the number of areas
reaching working memory states increase with increasing 𝐺 and decreasing 𝛼, which
is consistent with what was observed in macaque model. There are two unexpected
observations. 1. when there is no counterstream inhibition (𝛼 = 0), there is a sudden
drop of FR activity, which might be a result of the gradient taking over control the
system. 2. there is narrow regime in the middle of the parameter space where there
is a sudden drop of FR activity. We suspect it might because of the deactivation
of some key hubs within the network caused by the anti-correlation between global
property (T1:T2 ratio) and cellular property (NMDA receptor density), which awaits
for further investigation.
Fig 3-2 C shows the experimental evidence map, with color coding representing
the number of evidence (normalized to 0-1) found for each area. We can see that there
are a lot of overlaps between the evidence map and our simulation results, especially
around temporal, frontal and prefrontal lobes. Details of literature can be found in
Appendix B-1.

A B
Alpha: strength of inhibition
G: global coupling

Figure 3-2: Activity spatial map across parameter space. A. activity map with dif-
ferent 𝛼 (0 - 1) and 𝐺 (0.5 - 1) values. B. number of areas in working memory states
during delay period across parameter space (𝐺0.6 = 1.2, 𝛼 = 0 − 1). C. map of
experimental evidence.
.

21
3.3 Robust working memory to distractors

While a piece of information is held in mind before being used for execution of
cognitive tasks, the brain has to maintain it against internal noises and distraction of
other irrelevant incoming information. To test the robustness of working memory in
our model, we designed two tasks, one with an auditory-go-signal and another with
a visual distractor. In a working memory task, an auditory-go-signal is usually a
short tone presented after the delay period to cue the participants to record what
they remembered. A visual distractor, on the other hand, gives a short presentation
of a different visual stimuli from the original one. We would like to make sure such
signals would not drag the system out of attractor state. Our simulations show that
higher cortical areas can sustain its working memory state even with the distractors.
As can be seen in Fig 3-3, after a small and brief perturbation, the neural activity of
the population receiving the original inputs went back to the same level before the
perturbation. The results confirmed the robustness of the model.

3.4 Stronger functional connectivity in higher areas

Functional connectivity measures temporal correlations of the activity among cor-

tical areas, which shows the interaction between areas on top of all the data and rules
we imposed to the model. Here we plotted functional connectivity matrix for fir-
ing rate activity during the resting period before stimuli and the delay period after
stimuli. Evident differences can be seen between the functional connectivity patterns
of resting state and working memory state. For resting state (Fig 3-4 B), no obvi-
ous structure can be observed in the functional connectivity matrix. During working
memory state (Fig 3-4 C), the functional connectivity between higher areas are strong
which can be explained by the fact that they are the areas in working memory states
and supporting each other to maintain the activity.

22
 A B

Go
L R L R L R L R
Visual cue Auditory go signal Visual cue Visual Distractor

Figure 3-3: Working memory tasks with distractors. A. working memory task with
an auditory-go-signal. B. working memory task with a visual distractor. The top
panel shows the scheme of the task. The middle panel shows the temporal dynamics
of the input areas (A1 and V1). The bottom panel shows the temporal dynamics of
the higher areas (PGS)
.

23
 A
Conductance BOLD signal

Ballon model

B C
Resting state Working memory state

Figure 3-4: Functional connectivity for resting and working memory states. A. scheme
of hemodynamic transformation. B. functional connectivity in resting state. C. func-
tional connectivity iin working memory state
.

3.5 Comparison to macaque model

By comparing macaque and human data (Fig 3-5), we can found following dif-
ferences. To begin with, there are much more areas included in the human model
(n = 180) than macaque (n = 30). For connectivity, interareal connections (density
= 0.99) are denser than that for macaque (density = 0.65). For hierarchy values,
those for human are more linear than for macaque. For FF/FB ratio, human data
are derived from hierarchy and do not contain noise compared to macaque data. We
also made a preliminary comparison between the firing patterns for the two models.
As can be seen in Fig 3-6, two simulations share similar patterns, with higher activity

24
around temporal lobe, frontal lobe and parietal lobe. In frontal and temporal lobe,
the firing activity is significantly lower for human than macaque model.

Figure 3-5: Comparison between macaque and human data. A. macaque anatomical
data. B. human anatomical data. From left to right are: 1) connectivity strength
matrix, 2) FF strength matrix, 3) FF strength plotted against hierarchy values of the
target areas, 4) Hierarchy values sorted from low to high
.

3.6 Effect of hierarchy linearity

By comparing the hierarchy values for macaque model and human model, we can
observe observe obvious difference in terms of linearity. As a preliminary investiga-
tion of the effect of the hierarchy linearity, we compared the simulation results of
human model with different degree of linearity, as shown in Figure . It turns out
as the hierarchy values become more nonlinear with the logarithmic transformation,
the less areas reaching working memory states. This may be explained by the non-
linear segregation between low and high areas and the increasing inhibition from the
increasing number of higher areas to the lower areas.

25
 A B C

0 40

Figure 3-6: Comparison between macaque and human simulation results. A. simula-
tion results for macaque model. B. simulation results for human model. C. box-plot
of firing rates in five cortical lobes from the simulations results of the two models.
PRE: prefrontal, FON: frontal, PAR: parietal, TEM: temporal, OCC: occipital
.

26
 A B C

Figure 3-7: Effect of hierarchy linearity on simulation results. A. hierarchy with-

out nonlinear transformation. B. hierarchy with non-linear transformation (𝐻 =
𝑙𝑜𝑔10(ℎ + 𝑏) for 𝑏 = 0.1). C. hierarchy with non-linear transformation (𝑏 = 0.05)
.

27
28
Chapter 4

Discussion

Recent evidence has shown activation of multiple cortical areas in working memory
tasks [17] [5] [28], which calls for computational models that promote understanding
of the underlying mechanism. Within the current study, we build a cortical net-
work model which can give arise to distributed working memory states in human
brain, where multiple cortical areas maintain sustained activity in the delay period of
working memory tasks. Constrained by human anatomical data, the activity map ap-
proximate the experimental evidence. We showed that the system is robust to noises
and distractors, and the firing patterns across areas is similar to that for macaque
model. Those results provide further evidence to the underlying mechanisms of dis-
tributed working memory we assumed, including interareal connections, cortical hier-
archy, gradient and feedback inhibition. Interareal connectivity provides a foundation
for the neural activity started in sensory areas to propagate across the brain. The
cortical hierarchy and gradient allow the firing activity to accumulate along the hi-
erarchy which lead to working memory sates emerging in higher areas. The feedback
inhibition suppresses firing activity in lower areas, to maintain a hierarchical activity
pattern, avoiding overflow of energy and information. The current model extends pre-
vious small cortical models [21] into a large-scale model which enables much stronger
explanatory power.
The model provides possibility of further investigation of working memory mecha-
nisms, the cognitive abilities that supported by working memory and facilitate clinical

29
interventions. By replicating a wide range of working memory tasks and finding the
discrepancies between simulation results and evidence, we can pose new hypothesis
to the model and further improve the understanding of neural mechanism of working
memory. As an executive function, working memory supports many other cognitive
functions, ranging from reasoning, learning and language comprehension [6]. With our
model, we may probe how working memory affects those functions by analysing activ-
ities of different functional networks. To study learning and cognitive development,
we may also integrate learning rules for the local connections and the developmen-
tal change of the global connectivity. There are neurological pathologies which are
linked to working memory. For example, studies have shown reduced working mem-
ory performance in schizophrenia. More importantly, such deficit are found to start
from a early course [16]. By using human working memory cortical model to find out
the underlying neural mechanism of working memory deficit, we can potentially make
predictions of the development of schizophrenia and help with targeting cortical areas
during treatment in the early phase.

4.1 Limitation

There are also a few limitations we need to consider. First is that there are certain
constraints to the anatomical data we are using. The connectome data is measured
from diffusion imaging, therefore do not indicate directionality. The hierarchy data
is not a direct measurement of FF and FB projection and may lack of accuracy. The
NMDA receptor density data used here can be a mixed measurement for receptors
that is on both inhibitory population and excitatory population, therefore may not
accurately approximate the gradient of local recurrent excitation strength. With more
informed and accurate data in the future, we can make further improvement to the
model to make it more biologically realistic.
We also need to consider that the model is not incorporating functional hetero-
geneity of different brain areas. In reality, some functional network may not function
to perform working memory tasks [29]. Furthermore, attention plays a role in encod-

30
ing, maintaining and manipulating information in working memory [1]. The effect of
attention is not incorporated in the current model, which may change how different
area respond to the tasks.
The model only include cortical areas. However, some sub-cortical areas such as
thalamus can play an important role in working memory [22]. This should be noted
when interpreting the simulation results.
Finally, the current theoretical framework proposes an essential role of attractor
states for maintaining working memory, which cannot explain activity-silent working
memory [25] [31], where brain activity during working memory varies with predictabil-
ity of the temporal structure of the tasks or attention. This poses challenges to the
importance of sustained activity from the attractor framework. We may alleviate this
in the future by incorporating short term plasticity [19].

4.2 Future Directions

Recent study of working memory in mice suggested that working memory rep-
resentations are embeded in high-dimensional population activity [28]. For future
research, we may start to investigate working memory states in a higher dimensional
space, by incorporating network of single neurons, such as spiking neurons network
[33], into the model. This would allow us to distinguish the role of neural populations
with different firing frequencies in coding working memory.
Studies show that dysfunction of network hubs can explain working memory deficit
[10]. Simulations from the macaque model [19] shows that by silencing key areas in
the network, the working memory activity for the whole network can be shut down.
To identify key areas/sub-networks for distributed working memory in human brain,
we can study the topology of cortical network and test hypothesis by applying lesions
to identified areas/sub-networks.
On top of that, we can study the information coding during working memory. Pre-
vious studies show that objects can be decoded from the late response enhancement
in early visual areas and it was hypothesized that inter-areal connections account for

31
it. We would like to use our model to test this hypothesis.

32
Appendix A

Appendix

33
 A B

Figure A-1: Input to V1 or multiple visual areas. A. Input to V1. B. Input to multiple
visual areas. From top to bottom: 1. location of the input areas, 2. hierarchy of the
input areas, 3. temporal dynamics across areas with inputs.
.

34
Appendix B

Appendix

35
Table-1
Brain area Description Literature
Prefrontal cortex-PFC Osaka et al. 2003, Chein et al.,
2011, Vartanian et al., 2013,
Buschman et al., 2011, Jacob et
al., 2014, Bourbon-Teles et al.,
2019, Mendoza-Halliday et al.,
2014, Vergara et al., 2016,
Cavanagh et al., 2018, Rigotti et
al., 2013, Parthasarathy et al.,
2017, Sarma et al., 2015
anterior PFC Vartanian et al., 2013, Jimura et
al., 2017
Dorsolateral prefrontal cortex- Jimura et al., 2017, Kim et al.,
DLPFC 2015, Moore et al., 2013, Murty
et al., 2011, Vartanian et al., 2013
Ventrolateral prefrontal cortex- Vartanian et al., 2013
VLPFC
lateral prefrontal cortex-LPFC Mendoza-Halliday et al., 2014
left dorsal frontal cortex-lDFC Bourbon-Teles et al., 2019
Posterior parietal cortex-PPC Chein et al., 2011, Jimura et al.,
2017, Christophel et al., 2012,
Sarma et al., 2015
Anterior cingulate cortex-ACC Osaka et al. 2003, Chein et al.,
2011, Kamiński et al., 2017
Dorsal Anterior cingulate Kamiński et al., 2017
cortex-dACC
Orbitofrontal cortex-OFC Vartanian et al., 2013, Cavanagh
et al., 2018
Frontopolar cortex-FPC Kim et al., 2015
frontal eye field-FEF Buschman et al., 2011,
Parthasarathy et al., 2017,
Armstrong et al., 2009; Clark et
al., 2012; Rezayat et al., 2021
Primary visual cortex-V1 Van Kerkoerle et al., 2017
Middle frontal gyrus-MFG Moore et al., 2013, Kim et al.,
2015
Inferior frontal gyrus-IFG Kim et al., 2015
Superior temporal gyrus-STG Osaka et al., 2003
Inferior frontal sulcus-IFS Moore et al., 2013
Inferior frontal junctions-IFJ Jimura et al., 2017
tempo-parietal junction-TPJ Jimura et al., 2017
Supplementary motor area- Vartanian et al., 2013
SMA
Pre Supplementary motor area- Jimura et al., 2017, Kamiński et
pSMA al., 2017, Vergara et al., 2016
Anterior insula-AI Jimura et al., 2017
Medial temporal lobe-MTL Chein et al., 2011, Kornblith et
al., 2017
medial superior temporal-MST Mendoza-Halliday et al., 2014
parietal and cingulate regions Moore et al., 2013, Murty et al.,
2011

Figure B-1: Literature collected for generating the evidence map

.

36
Bibliography

[1] Gazzaley Adam. The relationship between attention and working memory, vol-
ume 5. 2011.
[2] Antonin Blot, Morgane M. Roth, Ioana Gasler, Mitra Javadzadeh, Fabia Imhof,
and Sonja B. Hofer. Visual intracortical and transthalamic pathways carry dis-
tinct information to cortical areas. Neuron, 109(12):1996–2008.e6, 2021.
[3] Joshua B. Burt, Murat Demirtaş, William J. Eckner, Natasha M. Navejar,
Jie Lisa Ji, William J. Martin, Alberto Bernacchia, Alan Anticevic, and John D.
Murray. Hierarchy of transcriptomic specialization across human cortex cap-
tured by structural neuroimaging topography. Nature Neuroscience, 21:1251–
1259, 2018.
[4] Rishidev Chaudhuri, Kenneth Knoblauch, Marie Alice Gariel, Henry Kennedy,
and Xiao Jing Wang. A Large-Scale Circuit Mechanism for Hierarchical Dynam-
ical Processing in the Primate Cortex. Neuron, 88(2):419–431, 2015.
[5] Thomas B. Christophel, P. Christiaan Klink, Bernhard Spitzer, Pieter R. Roelf-
sema, and John Dylan Haynes. The distributed nature of working memory.
Trends in Cognitive Sciences, 21:111–124, 2017.
[6] Nelson Cowan. Working Memory Underpins Cognitive Development, Learning,
and Education. Educational psychology review, 26(2):197–223, jun 2014.
[7] Murat Demirtaş, Joshua B. Burt, Markus Helmer, Jie Lisa Ji, Brendan D. Ad-
kinson, Matthew F. Glasser, David C. Van Essen, Stamatios N. Sotiropoulos,
Alan Anticevic, and John D. Murray. Hierarchical heterogeneity across human
cortex shapes large-scale neural dynamics. Neuron, 101:1181–1194.e13, 2019.
[8] Sophie Denève and Christian K. Machens. Efficient codes and balanced networks.
Nature Neuroscience, 19:375–382, 2016.
[9] Albert Einstein. Zur Elektrodynamik bewegter Körper. (German) [On the elec-
trodynamics of moving bodies]. Annalen der Physik, 322(10):891–921, 1905.
[10] Hamdi Eryilmaz, Melissa Pax, Alexandra G. O’Neill, Mark Vangel, Ibai Diez,
Daphne J. Holt, Joan A. Camprodon, Jorge Sepulcre, and Joshua L. Roff-
man. Network hub centrality and working memory performance in schizophrenia.
Schizophrenia, 8(1), 2022.

37
[11] K. J. Friston, L. Harrison, and W. Penny. Dynamic causal modelling. NeuroIm-
age, 19(4):1273–1302, 2003.

[12] Joaquin M. Fuster and Garrett E. Alexander. Neuron activity related to short-
term memory. Science, 173(3997):652–654, 1971.

[13] P. S. Goldman-Rakic. Cellular basis of working memory. Neuron, 14(3):477–485,

1995.

[14] Justine Y Hansen, Golia Shafiei, Ross D Markello, Kelly Smart, Sylvia M. L.
Cox, Yanjun Wu, Jean-Dominique Gallezot, Étienne Aumont, Stijn Servaes,
Stephanie G Scala, Jonathan M. DuBois, Gabriel Wainstein, Gleb Bezgin,
Thomas Funck, Taylor W Schmitz, R. Nathan Spreng, Jean-Paul Soucy, Syl-
vain Baillet, Synthia Guimond, Jarmo Hietala, Marc-André Bédard, Marco Ley-
ton, Eliane Kobayashi, Pedro Rosa-Neto, Nicola Palomero-Gallagher, James M.
Shine, Richard E Carson, Lauri Tuominen, Alain Dagher, and Bratislav Misic.
Mapping neurotransmitter systems to the structural and functional organization
of the human neocortex. bioRxiv, page 2021.10.28.466336, 2021.

[15] James J. Jun, Nicholas A. Steinmetz, Joshua H. Siegle, Daniel J. Denman, Mar-
ius Bauza, Brian Barbarits, Albert K. Lee, Costas A. Anastassiou, Alexandru
Andrei, Çaǧatay Aydin, Mladen Barbic, Timothy J. Blanche, Vincent Bonin,
João Couto, Barundeb Dutta, Sergey L. Gratiy, Diego A. Gutnisky, Michael
Häusser, Bill Karsh, Peter Ledochowitsch, Carolina Mora Lopez, Catalin Mite-
lut, Silke Musa, Michael Okun, Marius Pachitariu, Jan Putzeys, P. Dylan Rich,
Cyrille Rossant, Wei Lung Sun, Karel Svoboda, Matteo Carandini, Kenneth D.
Harris, Christof Koch, John O’Keefe, and Timothy D. Harris. Fully integrated
silicon probes for high-density recording of neural activity. Nature, 551:232–236,
2017.

[16] O Kebir and K Tabbane. [Working memory in schizophrenia: a review].

L’Encephale, 34(3):289–298, jun 2008.

[17] Matthew L. Leavitt, Diego Mendoza-Halliday, and Julio C. Martinez-Trujillo.

Sustained activity encoding working memories: Not fully distributed. Trends in
Neurosciences, 40:328–346, 2017.

[18] Nikola T. Markov, Julien Vezoli, Pascal Chameau, Arnaud Falchier, René Quilo-
dran, Cyril Huissoud, Camille Lamy, Pierre Misery, Pascale Giroud, Shimon Ull-
man, Pascal Barone, Colette Dehay, Kenneth Knoblauch, and Henry Kennedy.
Anatomy of hierarchy: Feedforward and feedback pathways in macaque visual
cortex. Journal of Comparative Neurology, 522(1):225–259, 2014.

[19] Jorge F. Mejias and Xiao-Jing Wang. Mechanisms of distributed working memory
in a large-scale network of macaque neocortex. eLife, 2022.

[20] Kenji Morita. Dynamical foundations of the neural circuit for bayesian decision
making. Journal of Neurophysiology, 102:1–6, 2009.

38
[21] John D. Murray, Jorge Jaramillo, and Xiao Jing Wang. Working memory and
decision-making in a frontoparietal circuit model. Journal of Neuroscience,
37(50):12167–12186, 2017.

[22] Dheeraj S. Roy, Ying Zhang, Tomomi Aida, Chenjie Shen, Keith M. Skaggs,
Yuanyuan Hou, Morgan Fleishman, Olivia Mosto, Alyssa Weninger, and Guop-
ing Feng. Anterior thalamic circuits crucial for working memory. Proceedings of
the National Academy of Sciences, 119(20):e2118712119, 2022.

[23] Dheeraj S. Roy, Ying Zhang, Tomomi Aida, Chenjie Shen, Keith M. Skaggs,
Yuanyuan Hou, Morgan Fleishman, Olivia Mosto, Alyssa Weninger, and Guop-
ing Feng. Anterior thalamic circuits crucial for working memory. Proceedings of
the National Academy of Sciences, 119(20):e2118712119, 2022.

[24] Kartik K. Sreenivasan and Mark D’Esposito. The what, where and how of delay
activity. Nature Reviews Neuroscience, 20:466–481, 2019.

[25] Mark G. Stokes. ’Activity-silent’ working memory in prefrontal cortex: A dy-

namic coding framework. Trends in Cognitive Sciences, 19(7):394–405, 2015.

[26] Carsen Stringer, Marius Pachitariu, Nicholas Steinmetz, Charu Bai Reddy, Mat-
teo Carandini, and Kenneth D. Harris. Spontaneous behaviors drive multidi-
mensional, brainwide activity. Science, 364, 2019.

[27] H. Supèr, H. Spekreijse, and V. A.F. Lamme. A neural correlate of working

memory in the monkey primary visual cortex. Science, 293(5527):120–124, 2001.

[28] Ivan Voitov and Thomas D. Mrsic-Flogel. Cortical feedback loops bind dis-
tributed representations of working memory. Nature, 608(7922):381–389, 2022.

[29] Zhiwei Wang, Kristina Zeljic, Qinying Jiang, Yong Gu, Wei Wang, and Zheng
Wang. Dynamic network communication in the human functional connectome
predicts perceptual variability in visual illusion. Cerebral Cortex, 28(1):48–62,
2018.

[30] Kei Watanabe and Shintaro Funahashi. Neural mechanisms of dual-task in-
terference and cognitive capacity limitation in the prefrontal cortex. Nature
Neuroscience, 17(4):601–611, 2014.

[31] Michael J. Wolff, Janina Jochim, Elkan G. Akyürek, and Mark G. Stokes. Dy-
namic hidden states underlying working-memory-guided behavior. Nature Neu-
roscience, 20(6):864–871, 2017.

[32] Kong Fatt Wong and Xiao Jing Wang. A recurrent network mechanism of time
integration in perceptual decisions. Journal of Neuroscience, 26:1314–1328, 2006.

[33] Yi Zeng, Dongcheng Zhao, Feifei Zhao, Guobin Shen, Yiting Dong, Enmeng Lu,
Qian Zhang, Yinqian Sun, Qian Liang, Yuxuan Zhao, Zhuoya Zhao, Hongjian
Fang, Yuwei Wang, Yang Li, Xin Liu, Chengcheng Du, Qingqun Kong, Zizhe

39
Ruan, and Weida Bi. BrainCog: A Spiking Neural Network based Brain-inspired
Cognitive Intelligence Engine for Brain-inspired AI and Brain Simulation. pages
1–28, 2022.

40