Ann. N.Y. Acad. Sci. ISSN 0077-8923
ANNALS OF THE NEW YORK ACADEMY OF SCIENCES
Special Issue:
The Year in Cognitive Neuroscience
REVIEW
Between persistently active and activity-silent frameworks:
novel vistas on the cellular basis of working memory
Jan Kami
́
nski
1,2
and Ueli Rutishauser
1,3,4,2
1
Department of Neurosurgery, Cedars-Sinai Medical Center, Los Angeles, California.
2
Division of Biology and Biological
Engineering, California Institute of Technology, Pasadena, California.
3
Department of Neurology, Cedars-Sinai Medical
Center, Los Angeles, California.
4
Center for Neural Science and Medicine, Department of Biomedical Sciences, Cedars-Sinai
Medical Center, Los Angeles, California.
Address for correspondence: Jan Kami
́
nski, Department of Neurosurgery, Cedars-Sinai Medical Center, 127 S. S San Vicente
Blvd, Los Angeles, CA 90048—1804. jan.kaminski@cshs.org
Recent work has revealed important new discoveries on the cellular mechanisms of working memory (WM). These
findings have motivated several seemingly conflicting theories on the mechanisms of short-term memory main-
tenance. Here, we summarize the key insights gained from these new experiments and critically evaluate them in
light of three hypotheses: classical persistent activity, activity-silent, and dynamic coding. The experiments dis-
cussed include the first direct demonstration of persistently active neurons in the human medial temporal lobe
that form static attractors with relevance to WM, single-neuron recordings in the macaque prefrontal cortex that
show evidence for both persistent and more dynamic types of WM representations, and noninvasive neuroimag-
ing in humans that argues for activity-silent representations. A key insight that emerges from these new results
is that there are several neural mechanisms that support the maintenance of information in WM. Finally, based
on established cognitive theories of WM, we propose a coherent model that encompasses these seemingly contra-
dictory results. We propose that the three neuronal mechanisms of persistent activity, activity-silent, and dynamic
coding map well onto the cognitive levels of information processing (within focus of attention, activated long-term
memory, and central executive) that Cowan’s WM model proposes.
Keywords:
working memory; single-neuron recordings; persistent activity; dynamic coding; static coding; attractors
Introduction
Working memory (WM, see Table 1 for a list of
acronyms) is the capacity to hold and manipulate
information in mind.
1
WM is a fundamental cogni-
tive function that allows us to execute complex tasks
in a constantly changing environment.
1
Recent
years have brought substantial advancement in the
field of WM and this has driven an emergence of
new hypotheses regarding the neuronal mechanism
ofhowweholdandmanipulateitemsinmind.The
goal of this review is to describe these new discover-
ies and consider them in the context of established
cognitiveframeworksofWM.Ourfocushereison
new discoveries made at the single neuron level,
which provides an unprecedented opportunity to
directly observe the mechanisms that support WM.
Here, we summarize findings that together show
thatitemsinWMcanbestoredusingtwokinds
of mechanisms: one decodable from neural activ-
ity measured by electrophysiological or metabolic
means and one that is not (Fig. 1). Next, we examine
the newly emerging view that there are two forms of
active representations that maintain working mem-
ories: stable, persistently active neurons and dynam-
ically active neurons. We conclude by proposing
how these three different types of cellular mecha-
nisms for maintaining information fit into the cog-
nitive frameworks of WM.
Single-cell evidence for persistent activity
Almost 50 years ago, scientists for the first time
observed neurons that continue to fire during
doi: 10.1111/nyas.14213
64
Ann. N.Y. Acad. Sci. 1464 (2020) 64–75 © 2019 The Authors.
Annals of the New York Academy of Sciences
published by Wiley Periodicals, Inc. on behalf of New York Academy of Sciences.
This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and
distribution in any medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.
Kami
́
nski & Rutishauser
The cellular basis of working memory
Table 1.
Acronyms
ALM
Anterior lateral motor cortex
BOLD-fMRI
Blood-oxygen-level-dependent
functional magnetic resonance
imaging
EEG
Electroencephalogram
FEF
Frontal eye field
IPS
Intraparietal sulcus
LFP
Local field potentials
LTM
Long-term memory
MTL
Medial temporal lobe
PA
Persistent activity (of neurons)
PCA
Principal component analysis
PFC
Prefrontal cortex
STSP
Short-term synaptic plasticity
WM
Working memory
the maintenance period after the end of stimulus
presentation.
2
Such activity can last for many sec-
onds and is stimulus specific: cells continue to fire
only if their preferred stimulus (which typically
is a specific sensory input or the direction of an
instructed motor movement to be executed later)
is held in WM. This pattern of activity has become
known as
persistent activity
(PA, also called delay
activity) because it outlasts the time of stimulus pre-
sentation. Subsequently, PA has become the central
element in theories of the neuronal mechanism of
WM.
3
To date, signatures of PA have been observed
at the single-cell level in many brain areas, species,
and experimental paradigms.
4–16
In addition to much work in animal models,
recent work has revealed direct evidence for PA
anditsrelevancetoWMinhumans.
17,18
This work
was performed as part of invasive monitoring for
seizure localization, a clinical procedure for which
depth electrodes with embedded microwires are
implanted in human patients suffering from drug-
resistant epilepsy.
19
Using this method, the elec-
trical activity of individual neurons is recorded
while patients perform cognitive tasks.
20
The first
direct evidence for WM relevant PA in human
neurons
17,18
shows that highly selective “concept”
cellsinthehumanmedialtemporallobe(MTL)can
remain active for several seconds if the concept that
activates these neurons is held in WM (Fig. 2A).
Conceptcellsareatypeofcellwhoseproperties
have been studied extensively in humans and are
therefore relatively well understood.
21,22
However,
so far, concept cells have been viewed as represent-
ing aspects of declarative memories and not those of
WM.
23
Persistently active concept cells were found
in multiple areas of the MTL, including the hip-
pocampus, amygdala, parahippocampal cortex, and
entorhinal cortex. The strength of this activity pre-
dicted behavior and scaled with WM load.
17
In
addition, nonstimulus-specific PA in the MTL has
also been observed.
17,24
Together, this body of work
reveals evidence for persistently active cells in the
humanMTLwhoseactivityisrelatedtoWM.
Attractors as a framework to study PA
What are the mechanisms that give rise to sus-
tained neuronal activity? One possibility is cell
autonomous mechanisms, which exist in cer-
tain specialized cells,
25,26
including in humans.
27
Another possibility is at the network level, facil-
itated by recurrent synaptic connections. By and
large, current theories assume that the PA that
givesrisetoWMisduetonetwork-leveleffects.
26,28
However, in most experiments conducted so far (see
Ref.7foranotableexception),itremainsunknown
whether the observed PA is indeed dependent
on network-level interactions rather than cell-
autonomous mechanisms. The contribution of cell-
autonomous PA to WM thus remains an important
open question.
A useful theoretical framework to conceptualize
network-level PA is that of attractors, which are sta-
ble patterns of neural activity that are maintained
through recurrent excitation.
29,30
Each possible pat-
tern constitutes a different possible item held in
WM. There are two different classes of attractor
models that are typically considered in this con-
text: continuous attractor networks, which have a
continuum of stable states ideal for encoding ana-
log variables, and discrete attractors networks that
have a countable number of possible discrete states
that compete with each other. Recent single-neuron
studies in several different species (mice, rhesus,
and humans) have provided direct experimental
evidence for the presence of such attractors during
WM maintenance and their relation to behavior. We
will next summarize these key findings.
Recently, we used demixed PCA to assess the
population dynamics associated with the identity of
stimuli held in WM.
17,31
This revealed that during
WM maintenance, the speed of the neuronal tra-
jectories in the dimensions associated with stimulus
identity was low (comparable to the speed present at
65
Ann. N.Y. Acad. Sci. 1464 (2020) 64–75 © 2019 The Authors.
Annals of the New York Academy of Sciences
published by Wiley Periodicals, Inc. on behalf of New York Academy of Sciences.
The cellular basis of working memory
Kami
́
nski & Rutishauser
Item A
Item B
Item C
Item A
Item A
Item A
Item A
Item B
Item C
Persistent acvity
Dynamic acvity
Stable readout
Unstable readout
No readout
Acvity-silent
Time 1
2
Time 1
2
B
AC
Figure 1.
Summary of theories of the neuronal mechanisms supporting working memory. (A) The persistent activity framework
proposes that memoranda are maintained by
the sustained firing of stimulus-specifi
c groups of neurons. A decoder trained in
one period of time should be able to decode information at a different point of time (“1” versus “2” periods indicated). (B) The
activity-silent framework proposes that information held in WM is not visible by observing the activity of individual neurons.
(C) The dynamic coding framework proposes that the neurons carrying information about a specific item change as function of
time relative to the onset of the maintenance period. For example, some neurons encode the identity of an item only at a specific
period of time. In the figure, three neurons are shown, all of which represent item A, but during different periods of time, with
some neurons “ramping down” their activity (top), whereas othe
rs firing only during specific periods of time. A decoder trained
at one point of time will thus not generalize to a different point of time (“1” versus “2” periods indicated).
baseline).Atthesametime,wefoundthatthedis-
tance between the trajectories associated with dif-
ferent items was high. This suggests that neuronal
activity was pulled to a particular area in state space
and then remained there, which is the definition of a
discrete attractor (Fig. 2B).
29
These attractors were
behaviorally relevant: the distance of the neuronal
trajectory in a given trial to the center of the attrac-
tor was correlated with later accuracy and reaction
time for a test stimulus. This revealed that in trials
when activity drifted away from the center of the
attractor, the quality of the memory decreased.
Another study has shown evidence for contin-
uous “bump attractors” by revealing a direct rela-
tion between attractor dynamics and behavior.
32
Abumpattractorisatypeofcontinuousattrac-
tor, where activity forms a bell-shape like pat-
tern of activity that is centered on the value of
a currently maintained memorandum. Monkeys
performed an oculomotor spatial WM task that
required memorizing one of eight possible spatial
locations. During maintenance, the authors found
stimulus-specific PA in the prefrontal cortex. The
fluctuations in activity of these neurons correlated
trial-by-trial with inaccuracies of the behavioral
response (Fig. 2C). These fluctuations were such
that neurons that encoded positions adjacent to the
currently cued location increased their firing rate
proportionally to the behavioral bias toward the
position preferred by a given cell. This linear rela-
tionship could be explained by continuous, but not
by discrete attractor dynamics,
32
thereby revealing
experimental evidence for continuous attractors rel-
evant for WM at the single-cell level.
Attractordynamicswerealsocloselyexaminedin
a study in which mice needed to maintain informa-
tion about the position of a reward (left or right).
7
First, extracellular recordings in the anterior lateral
motor cortex (ALM) revealed evidence for strong
memory content–specific PA that formed attractors
in state space. Critically, the authors showed that
thePAinthiscasewasnotduetocell-autonomous
PA. To achieve this, they hyperpolarized persis-
tently active cells using whole-cell recordings. This
abolished spiking activity but not other signs of
persistent synaptic activation as measured by the
subthreshold membrane potential, thereby reveal-
ing a network-level origin of PA in ALM.
33
This
is, to our knowledge, the strongest
in-vivo
demon-
stration of the network origin of PA during WM
thus far. Second, strong optogenetic inhibition of
ALM cells abolished the ability of mice to maintain
66
Ann. N.Y. Acad. Sci. 1464 (2020) 64–75 © 2019 The Authors.
Annals of the New York Academy of Sciences
published by Wiley Periodicals, Inc. on behalf of New York Academy of Sciences.
Kami
́
nski & Rutishauser
The cellular basis of working memory
firing rate (Hz)
0
5
10
15
20
trial n
r (re-
ordered)
time form start of section (s)
01010101201
no preferred image in trial
preferred image shown
preferred image already encoded in the trial (preferred in the memory)
preferred image shown, when preferred in the memory
encoding 2
encoding 3
maintenance
retrieval
encoding 1
image presentation
dPC 3
dPC 2
dPC 5
-6
-4
-2
0
2
-2
0
2
4
-4
-2
0
2
img A encoding
img B encoding
img C encoding
img D encoding
maintenace
0.2 s
0.2 s
0.2 s
B
A
CDE
Figure 2.
Persistent activity represented by attractor dynamic. (A) Example of a persistently active concept cell recorded in the
human amygdala. Subjects memorized up to three images presented sequentially (encoding 1–3). Top: post stimulus time his-
togram. Middle: periods of significance (black) between the preferred versus nonpreferred stimuli of this cell. Bottom: raster plot
of trials reordered according to condition. During maintenance, the
activity of this cell is characterized by sustained activity only
when the preferred image of the cell is held in memory (blue) but not when other stimuli are held in memory (gray). Adapted
from Ref. 17. (B) Trajectories in neural state space formed by a population of persistently active concept cells in the human MTL.
Trajectories are projected into the 3D space formed by the three demixed principal components (dPCs) associated with picture
identity. Periods of time shown are encoding (thin line) and maintenance (thick line). Colors mark different images. Note how
during maintenance, activity settles at points in space that separate by memory content (attractors). Adapted from Ref. 17. (C)
Neuron whose activity is indicative of a continuous attractor during a delayed oculomotor task in rhesus monkey. Firing rate of
stimulus-selective neurons is sorted vertically according to preferred location. Note how activity drifts away from the initial posi-
tion (left) as time progresses. This drift predicts behavioral errors (right). Adapted from Ref. 50. (D) Persistent activity recorded
in mouse ALM during a task with variable delay durations. Blue color marks prefered location. (E) Population activity in mouse
ALM shows characteristic of a discrete attractor. Shown is a projection of the population activity onto the axis that maximally
distinguishing between the two possible conditions (left or right)
.Bluecolor(leftpanel)denotescorrectlick-righttrials,andred
denotes correct lick-left trials. Dark blue and dark red (right panel)
denote incorrect lick-right and li
ck-left trials, respectively.
Dashed lines denote trajectories of unperturbed correct trials, whereas solid lines denote perturbed trials. Light blue band on the
top shows time of photoinhibition. Note how the neural activity after offset of inhibition is pulled toward one of the two possible
trajectories, as expected from an attractor. Adapted from Ref. 7.
information in WM, showing the necessity of these
cells for information maintenance. Weak inhibi-
tion, on the other hand, revealed a remarkable phe-
nomenon: in some trials, such inhibition led to an
error, whereas in others it did not. In error tri-
als, the PA following offset of the inhibition resem-
bled that of the opposite direction (that was not
cued), whereas in correct trials, activity returned
tothecueddirection.Thus,thepatternofneu-
ronal activity was attracted toward one of two dis-
crete states (attractors, Fig. 2E) to which activity
returned after transient disruption. Together, this
study shows compelling evidence for network-level
PA that forms discrete attractors.
The study of Inagaki
et al.
7
revealed a critical
difference between when the maintenance duration
was of fixed versus randomized duration: stable PA
wasobservedonlyinthelatter(Fig.2D).Incontrast,
for fixed durations, stimulus selective delay-period
activity was characterized by slow, ramping activity.
While not classically expected from discrete attrac-
tor dynamics, slow ramping can be incorporated
67
Ann. N.Y. Acad. Sci. 1464 (2020) 64–75 © 2019 The Authors.
Annals of the New York Academy of Sciences
published by Wiley Periodicals, Inc. on behalf of New York Academy of Sciences.