Accepted Manuscript

Title: Top-down control: A unified principle of cortical

learning

Author: Hiroshi Makino

PII: S0168-0102(18)30170-6
DOI: https://doi.org/10.1016/j.neures.2018.08.004
Reference: NSR 4192

To appear in: Neuroscience Research

Received date: 24-3-2018

Revised date: 30-6-2018
Accepted date: 13-8-2018

Please cite this article as: Makino H, Top-down control: A unified principle of cortical
learning, Neuroscience Research (2018), https://doi.org/10.1016/j.neures.2018.08.004

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
 Makino

Top-down control: a unified principle of cortical learning

Hiroshi Makino*

Lee Kong Chian School of Medicine, Nanyang Technological University

11 Mandalay Road, Singapore, 308232

T
*
Correspondence:

IP
Hiroshi Makino
Lee Kong Chian School of Medicine

R
Nanyang Technological University

SC
11 Mandalay Road
Singapore, 308232
Email: [email protected]

U
N
Highlights:

A
Recruitment of top-down control is proposed as a unified principle of cortical learning.
 Microscopic circuit dissection revealed dynamic recruitment of top-down activity during sensory
M

learning.
 Macroscopic activity analysis uncovered top-down-mediated orchestration of cortical activity
D

during motor learning.

TE
EP
CC
A

1
 Makino

Abstract
Cognitive control of the brain flexibly maps incoming sensory information onto execution of actions
appropriate for the current goal. Learning is a process that enables the brain to estimate current states
of the world by extracting its spatiotemporal structure and generate goal-directed motor outputs
through selective association of events or movement refinement. Accumulating evidence suggests
that top-down control from higher-order brain areas modulates downstream neural activity and
changes local computations that are critical for the execution of learned behavior. Recent

T
technological advances in multi-site recordings and optogenetic approaches are beginning to reveal

IP
more direct evidence of top-down cognitive control by monitoring and perturbing activity of top-down
inputs and observing its causal consequences on behavior and downstream neural dynamics. Here I

R
highlight that learning-related changes in neural circuits in distinct domains of learning converge onto

SC
a unified principle; namely recruitment of top-down control whether it involves sensory, motor or
offline learning. Recruitment of top-down control may reflect experience-dependent adaptation and

U
integration of internal models for refined state estimation and goal-directed optimal behavior.
N
Keywords: Top-down control; Perceptual learning; Motor learning; Offline learning
A
M
D
TE
EP
CC
A

2
 Makino

1. Introduction
One of main themes in neuroscience is to decipher how structural and functional organization of the
brain is remodeled via perceptual and motor learning. Recent technological advances revealed
important principles underlying learning-related changes in the brain at multiple spatiotemporal
scales, including spine dynamics, activity dynamics of individual or populations of neurons, remodeling
of inhibitory circuits, neuromodulation and others (Makino et al., 2016).
It was classically considered that information processing is mediated through feedforward

T
pathways where hierarchically organized brain areas represent progressively more complex features

IP
of the environment (Felleman and Van Essen, 1991). Such an architecture has been successfully
employed in various neural network models of contemporary artificial intelligence, as exemplified in

R
deep artificial neural networks (Hassabis et al., 2017; LeCun et al., 2015). In biological neural networks,

SC
however, it is also evident that there exist feedback connections directing information in a reverse
order (Mignard and Malpeli, 1991), in addition to lateral connections projecting to the same

U
hierarchical level (Felleman and Van Essen, 1991). Such “top-down” signals from higher-order brain
areas exert cognitive influences and are considered to represent behavioral context, expectation or
N
attention, thereby modifying how information is processed in downstream areas (Gilbert and Li, 2013).
A
In this review, I provide a broad overview of the current status of neuroscience research in
M

neural circuit mechanisms of learning in the mammalian cortex by highlighting commonalities of

distinct forms of learning. The growing number of empirical evidence suggests that even different
D

forms of learning converge onto a unified principle of the brain operation, namely recruitment of top-
down control. Such a common feature of learning may reflect experience-dependent integration of
TE

prior knowledge for state estimations and goal representations to generate optimal behavior.
EP

2. Perceptual learning and recruitment of top-down control

Perceptual learning is an experience-dependent process to enhance detection or discrimination ability
CC

of the sensory system. An experienced sommelier, for example, can readily determine precise identity
of wine based on olfactory cues. One of strong evidence demonstrating the contribution of higher-
order brain areas during perceptual learning is transfer learning, which refers to the ability to
A

generalize learning to an untrained eye or to an untrained stimulus. For instance, sensory

discrimination of one feature (e.g. contrast) at one location in a visual field can be transferred to
another location if the subject is previously trained to discriminate another feature (e.g. orientation)
at the second location (Xiao et al., 2008) (Figure 1). This experiment challenged the traditional view

3
 Makino

that perceptual learning is limited in the retinotopic region of the brain and suggests the involvement
of recurrent interactions between the primary visual area and non-retinotopic higher areas.
In animals, more direct evidence has been presented to lend support to the significance of
top-down control during perceptual learning. Generally, experiments manipulate activity of top-down
sources and assess causal consequences in proposed downstream sites. For instance, a classical study
by Fuster et al. demonstrated that cooling the prefrontal cortex (PFC) of monkeys decreased neural
activity and stimulus-specificity in the downstream inferior temporal cortex (ITC) during a delayed-
match-to-sample task (Fuster et al., 1985). With a posterior-split-brain paradigm, Tomita et al. showed

T
that the PFC is sufficient to trigger reactivation of object-selective mnemonic representations in

IP
monkey ITC without bottom-up visual information (Tomita et al., 1999). Although top-down control
has been best described in the visual cortex of non-human primates (Cohen and Maunsell, 2009; Crist

R
et al., 2001; Hupe et al., 1998; Moore and Armstrong, 2003; Nienborg and Cumming, 2009; Reynolds

SC
et al., 2000), it is a universal phenomenon observed across a wide range of sensory modalities from
various species, including the mouse visual cortex (Poort et al., 2015; Zhang et al., 2014), mouse and

U
rat somatosensory cortex (Krupa et al., 2004; Kwon et al., 2016; Le Merre et al., 2018; Manita et al.,
2015; Xu et al., 2012) and mouse olfactory bulb (Otazu et al., 2015).
N
One potential function of top-down control may be prediction of inputs based on prior briefs
A
acquired through interactions with the environment. A theoretical notion known as predictive coding,
M

for example, postulates that the brain has an internal model of the environment that is continuously
optimized through incoming sensory inputs (Bastos et al., 2012; Friston, 2010). According to this view,
D

across different levels of cortical hierarchies, top-down predictions from a higher level “explain away”
or reduce bottom-up prediction errors in the level below through inhibitory action. If top-down
TE

expectation is violated, there is a strong error signal propagating upward in the hierarchy, and such a
reciprocal message passing between hierarchical layers continues until prediction errors are
EP

minimized. The most prominent empirical evidence for predictive coding is extra-classical receptive
field properties of the visual cortex (Vinje and Gallant, 2000). These properties become evident when
CC

stimulus attributes, such as orientation and direction of the visual stimulus, extend a neuron’s
receptive field to the surrounding region. Because of statistical regularities in space and time that are
inherent in nature, and as higher visual areas have larger receptive fields, spatially more widespread
A

top-down projections from these areas could in theory predict stimulus attributes of lower brain areas
representing features in smaller space (Rao and Ballard, 1999). As a result, neural responses in the
lower area, such as the monkey V1, can be greatly reduced as experimentally demonstrated, leading
to sparse neural representations (Vinje and Gallant, 2000). These experimental and theoretical studies

4
 Makino

indicate that learning the spatiotemporal structure of the environment may entail an increase in the
top-down prediction signal with a concurrent decrease in the bottom-up error signal.
Our recent study aimed at investigating whether such activity shifts occur in the mouse visual
cortex (Makino and Komiyama, 2015). When the mouse learned stimulus-action contingency, there
was a strong increase in activity of top-down inputs from the retrosplenial cortex to the visual cortex,
measured by two-photon calcium imaging of axonal projections in layer (L) 1. Concurrently, bottom-
up activity measured in L4 excitatory neurons of the visual cortex, a major thalamo-recipient layer,
was dramatically reduced (Figure 2). Furthermore, response profiles of excitatory neurons in L2/3, a

T
convergent layer of top-down, bottom-up and lateral inputs, changed gradually so that their temporal

IP
activity pattern became correlated with that of the top-down inputs. Importantly, pharmacological
inactivation of the retrosplenial cortex reversed the response profile of these L2/3 neurons to pre-

R
learning states, indicating that the post-learning response dynamics of these neurons were strongly

SC
driven by the top-down inputs from the retrosplenial cortex.
A similar notion of top-down inputs carrying prediction signals was also demonstrated in a

U
recent mouse visual flow experiment (Leinweber et al., 2017). In this experiment, the areas A24b (a
sub-region of anterior cingulate cortex) and secondary motor cortex (M2) were identified as top-down
N
input sources to V1. Their axonal projections to V1 were optogenetically inactivated in a virtual
A
environment where visual flow was coupled with locomotion in a closed-loop manner. Interestingly,
M

mismatch activity of L2/3 neurons in V1, which was prominent when the visual flow and locomotion
was experimentally decoupled, was reduced by the manipulation. Because such a mismatch signal
D

could represent an error signal under the predictive coding scheme, it was suggested that the top-
down inputs may carry the prediction signal.
TE

One of potential circuit mechanisms underlying predictive coding may involve intricate
interactions between long-range top-down inputs with local microcircuit operations. Top-down inputs
EP

from higher-order brain areas target mainly L1 and L6 of the sensory cortex. Neurons from the
retrosplenial cortex or A24b/M2, for example, form monosynaptic inputs to neurotensin receptor 1
CC

(NTSR1) neurons in L6 of the mouse visual cortex (Leinweber et al., 2017; Velez-Fort et al., 2014),
which in turn activate GABAergic neurons in the reticular thalamic nucleus. These inhibitory neurons
suppress activity of thalamo-cortical neurons from dorsal lateral geniculate nucleus projecting to L4.
A

In addition to the cortico-thalamo-cortical pathway, L6 NTSR1 neurons also suppress activity in upper
layers through disynaptic pathways within the cortex (Bortone et al., 2014; Olsen et al., 2012).
Therefore, the higher-order cortical areas could exert top-down modulation of the output gain of the
visual cortex through a reduction in the bottom-up activity. Consistent with the model, the importance
of top-down mediated L6 activation was suggested in a theoretical study (Rao and Ballard, 1999) based

5
 Makino

on the experimental observation on extra-classical receptive field properties in the visual system (Bolz
and Gilbert, 1986).

3. Motor learning and recruitment of top-down control

Recruitment of top-down control is not specific to perceptual learning but a general principle
underlying other domains of learning, such as motor learning. Motor learning refers to the acquisition
of a motor skill necessary to cope with on-going environmental demands. Whether playing the piano
or hitting a golf ball straight, motor skill learning involves a substantial reduction in movement

T
variability to generate precise actions at appropriate timing (Dhawale et al., 2017; Peters et al., 2017).

IP
Various circuit mechanisms underlying motor learning have been identified, such as reorganization of
dendritic spines and neural circuits, as well as increases in precision of activity patterns (Chen et al.,

R
2015; Fu et al., 2012; Huber et al., 2012; Masamizu et al., 2014; Peters et al., 2014; Xu et al., 2009).

SC
In humans, recruitment of top-down control during motor skill learning was suggested in
studies demonstrating intermanual transfer. Similar to the aforementioned transfer learning in the

U
visual domain, intermanual transfer is a form of generalization where motor skill learning with one
hand improves performance with the opposite hand (Figure 1). Perez and others, for example, showed
N
in humans intermanual transfer of motor skills with the serial reaction-time task, where subjects
A
executed a series of key presses with their fingers (Perez et al., 2007). Functional magnetic resonance
M

imaging (fMRI) revealed elevated activity in the supplementary motor area (SMA) when the
magnitude of intermanual transfer is high, and disruption of SMA activity with repetitive transcranial
D

magnetic stimulation interfered with intermanual transfer.

Functional roles of top-down control during motor behavior have been intensively scrutinized
TE

in animals. In a delayed-response task in rats, for example, neurons in both the dorsomedial PFC
(dmPFC) and motor cortex exhibit temporally correlated firing specifically during the delay period.
EP

Moreover, inactivation of dmPFC decreased delay-related activity but not movement-related activity
of neurons in the motor cortex (Narayanan and Laubach, 2006). These results support the notion that
CC

the motor cortex is under the influence of the frontal area, possibly conveying information about
abstract task structure (Miller and Cohen, 2001).
Our recent study also provided a macroscopic view of cortex-wide activity dynamics of the
A

mouse brain during motor learning (Makino et al., 2017). Wide-field calcium imaging was utilized to
monitor macro-scale changes in cortical activity over the course of a few weeks while the mouse
learned a lever-press task. When peri-movement activity was analyzed at the naïve stage, an activity
sequence spread from the retrosplenial cortex radially to other cortical areas. Over the course of
learning, however, additional stream of activity appeared, emanating from the premotor cortex to the

6
 Makino

rest of the cortical areas, and persisted thereafter (Figure 2). Moreover, the activity sequence was
compressed in time; as learning proceeded, the activity propagated faster from the first node of the
cortex (i.e. premotor cortex) to the last, even when similar movements were compared across
different stages of learning. Importantly, pharmacological blockade of the premotor area reversed
these learning-related changes to the pre-learning state. These results indicate that the premotor
cortex is recruited during motor learning to orchestrate activity in downstream areas through top-
down control. Similar dependency on the premotor cortex and flow of activity were also observed in
other goal-directed behavior tasks (Allen et al., 2017; Chen et al., 2017).

T
Motor learning allows the animal to optimally behave by minimizing movement costs while

IP
maximizing rewards. A major function of the top-down control may then be integration of prior
knowledge with statistical properties of an ongoing task. Theories posit that motor learning entails

R
acquisition of “forward” and “inverse” internal models (Wolpert et al., 2001). Forward models are

SC
essential for estimating the current state by predicting sensory consequences of motor commands,
and can be illustrated by our inability to tickle ourselves as sensations arising from self-motion are

U
presumably cancelled out by the forward prediction. Inverse models, on the other hand, are required
for motor planning to selectively generate motor outputs based on desired sensory consequences,
N
and are especially useful when sensory feedback is delayed to guide desired actions. It is possible that
A
internal representations of such state estimators and desired goals are residing in the higher-order
M

brain areas to influence downstream regions for generating optimal movements.

D

4. Offline learning and top-down control

It is evident that improvements in behavioral performance in the perceptual and motor domains often
TE

occur during sleep/rest in the absence of additional training, a phenomenon known as offline learning
(Stickgold et al., 2000; Walker et al., 2002). Offline learning can be sleep-related or due to the mere
EP

passage of time, depending on whether subjects have explicit knowledge of a task. For example, in
motor skill learning, benefits of sleep become evident when subjects are aware of a task rule, whereas
CC

the passage of time is sufficient to improve learning when they are not (Robertson et al., 2004). The
mechanisms of offline learning may involve a consolidation process where perceptual or motor
memories are slowly stabilized and become resistant to interference by competing memories, or a
A

reconsolidation process where transient reactivation of acquired memories reorganize memory traces
for more persistent representations (Dudai, 2004).
Currently, whether offline learning entails strengthening or weakening of synapses is a matter
of debate, but a theoretical notion known as synaptic homeostasis hypothesis (SHY) postulates that
learning-dependent synaptic potentiation during wakefulness is renormalized via net synaptic

7
 Makino

depression during sleep (Tononi and Cirelli, 2014). This enables reorganization of the neural circuit by
selectively retaining only behaviorally-relevant synapses while weak synapses are eliminated. In
support of SHY, growing evidence suggests that synaptic depression may be an important property
underlying sleep-related improvements in learning (de Vivo et al., 2017; Diering et al., 2017; Li et al.,
2017; Maret et al., 2011; Norimoto et al., 2018) (but see (Yang et al., 2014)).
Recently an elegant study demonstrated that activity of behaviorally-irrelevant neurons is
downscaled during sleep. Exploiting a neuroprosthetic-learning task with rats, Gulati et al. focused on
neurons that are causally linked to actuator movements, and showed that the firing rates of these

T
neurons are enhanced while the firing rates of the neurons that do not contribute to the actuator

IP
movements are reduced. Interestingly, optogenetic suppression of neural activity during non-rapid-
eye-movement (NREM) sleep disrupted the rescaling of task-relevant neural activity (Gulati et al.,

R
2017). Together with the evidence suggesting that firing rates of neurons are homeostatically

SC
regulated during sleep (Vyazovskiy et al., 2009), these results lend support to the notion that sleep
selectively up- or down-regulate neural activity depending upon task demands, which may be the key
mechanism underlying offline perceptual and motor skill learning.
U
How is offline learning related to the recruitment of top-down control? The observation that
N
intermanual transfer also occurs offline suggests that top-down control during sleep might play a role
A
(Witt et al., 2010). Interestingly, a recent study demonstrated in mice that top-down activity from the
M

secondary motor cortex to primary sensory area during NREM sleep is necessary for offline
improvement in perceptual task performance (Miyamoto et al., 2016). Although how such top-down
D

activity facilitates offline learning is yet unclear, this is an important step forward to understand how
offline learning modulates the neural structure and function through interactions with top-down input.
TE

Important future directions include investigation of how top-down inputs are activated during
different phases of sleep and whether top-down inputs are also necessary for offline improvements
EP

in motor skills. Further, it is critical to determine whether recruitment of top-down inputs is specifically
important for sleep-dependent explicit offline learning or also for sleep-independent implicit offline
CC

learning.

5. Conclusion
A

Although perception and action are distinct processes involving different brain areas, our recent
studies, together with others, have demonstrated that there is a unified principle underlying learning
in these domains. In this review, I highlighted recent studies focusing on learning-related recruitment
of top-down control from higher-order brain areas of the mammalian cortex. Depending on task
demands, these top-down control may uniquely change operation modes of downstream circuitries.

8
 Makino

The common function of the top-down control in different domains of learning may be “prior”
adaptation as discussed in a Bayesian framework (Kording and Wolpert, 2006). In the case of sensory
processing, experience-dependent development of prior briefs may strongly influence perception
especially when the state of the world is uncertain from sensory inputs alone. In the case of motor
behavior, optimal action selection depends on integration of current state estimates and prior
knowledge of potential costs and rewards for each action outcome. Recruitment of top-down control
may reflect such integration of priors for optimal motor commands.
There are many questions remaining to be answered in the context of top-down control

T
during learning. Recent findings that neurons sharing similar receptive field properties in the mouse

IP
primary visual cortex are more likely to be synaptically connected (Cossell et al., 2015; Ko et al., 2013;
Ko et al., 2011; Lee et al., 2016; Wertz et al., 2015) are elegant examples of how neural networks are

R
designed based on functional properties of individual units. Whether there are similar large-scale

SC
design principles across distant brain areas is not clear. Furthermore, how top-down control
adaptively changes operations of local circuits is not well established. For instance, although activity

U
of distinct inhibitory neurons change in a learning-dependent manner (Khan et al., 2018; Makino and
Komiyama, 2015), it is unknown whether and how top-down inputs drive these changes. Finally, it is
N
of broad interest to understand how offline learning remodels neural circuits to incorporate top-down
A
inputs for better behavioral performance. Understanding these basic operation principles of how top-
M

down control transforms cortical computations would provide new opportunities for better
understandings of the brain in health and disease.
D
TE
EP
CC
A

9
 Makino

Acknowledgements
I thank Ayumu Tashiro and Chi Ren for comments on the manuscript. This work was supported by the
NARSAD young investigator grant from the Brain & Behavior Research Foundation, Nanyang Assistant
Professorship, Nanyang Technological University Integrated Medicine Biology Environmental Life
Sciences (NIMBELS) grant and Ageing Research Institute for Society and Education (ARISE) grant from
Nanyang Technological University.

The author has no competing interests.

T
R IP
SC
U
N
A
M
D
TE
EP
CC
A

10
 Makino

References
Allen, W.E., Kauvar, I.V., Chen, M.Z., Richman, E.B., Yang, S.J., Chan, K., Gradinaru, V., Deverman, B.E.,
Luo, L., and Deisseroth, K. (2017). Global Representations of Goal-Directed Behavior in Distinct Cell
Types of Mouse Neocortex. Neuron 94, 891-907 e896.
Bastos, A.M., Usrey, W.M., Adams, R.A., Mangun, G.R., Fries, P., and Friston, K.J. (2012). Canonical
microcircuits for predictive coding. Neuron 76, 695-711.
Bolz, J., and Gilbert, C.D. (1986). Generation of end-inhibition in the visual cortex via interlaminar

T
connections. Nature 320, 362-365.

IP
Bortone, D.S., Olsen, S.R., and Scanziani, M. (2014). Translaminar inhibitory cells recruited by layer 6
corticothalamic neurons suppress visual cortex. Neuron 82, 474-485.

R
Chen, S.X., Kim, A.N., Peters, A.J., and Komiyama, T. (2015). Subtype-specific plasticity of inhibitory

SC
circuits in motor cortex during motor learning. Nat Neurosci 18, 1109-1115.
Chen, T.W., Li, N., Daie, K., and Svoboda, K. (2017). A Map of Anticipatory Activity in Mouse Motor
Cortex. Neuron 94, 866-879 e864.
U
Cohen, M.R., and Maunsell, J.H. (2009). Attention improves performance primarily by reducing
N
interneuronal correlations. Nat Neurosci 12, 1594-1600.
A
Cossell, L., Iacaruso, M.F., Muir, D.R., Houlton, R., Sader, E.N., Ko, H., Hofer, S.B., and Mrsic-Flogel, T.D.
M

(2015). Functional organization of excitatory synaptic strength in primary visual cortex. Nature 518,
399-403.
D

Crist, R.E., Li, W., and Gilbert, C.D. (2001). Learning to see: experience and attention in primary visual
cortex. Nat Neurosci 4, 519-525.
TE

de Vivo, L., Bellesi, M., Marshall, W., Bushong, E.A., Ellisman, M.H., Tononi, G., and Cirelli, C. (2017).
Ultrastructural evidence for synaptic scaling across the wake/sleep cycle. Science 355, 507-510.
EP

Dhawale, A.K., Smith, M.A., and Olveczky, B.P. (2017). The Role of Variability in Motor Learning. Annu
Rev Neurosci 40, 479-498.
CC

Diering, G.H., Nirujogi, R.S., Roth, R.H., Worley, P.F., Pandey, A., and Huganir, R.L. (2017). Homer1a
drives homeostatic scaling-down of excitatory synapses during sleep. Science 355, 511-515.
Dudai, Y. (2004). The neurobiology of consolidations, or, how stable is the engram? Annu Rev Psychol
A

55, 51-86.
Felleman, D.J., and Van Essen, D.C. (1991). Distributed hierarchical processing in the primate cerebral
cortex. Cereb Cortex 1, 1-47.
Friston, K. (2010). The free-energy principle: a unified brain theory? Nat Rev Neurosci 11, 127-138.

11
 Makino

Fu, M., Yu, X., Lu, J., and Zuo, Y. (2012). Repetitive motor learning induces coordinated formation of
clustered dendritic spines in vivo. Nature 483, 92-95.
Fuster, J.M., Bauer, R.H., and Jervey, J.P. (1985). Functional interactions between inferotemporal and
prefrontal cortex in a cognitive task. Brain Res 330, 299-307.
Gilbert, C.D., and Li, W. (2013). Top-down influences on visual processing. Nat Rev Neurosci 14, 350-
363.
Gulati, T., Guo, L., Ramanathan, D.S., Bodepudi, A., and Ganguly, K. (2017). Neural reactivations during
sleep determine network credit assignment. Nat Neurosci 20, 1277-1284.

T
Hassabis, D., Kumaran, D., Summerfield, C., and Botvinick, M. (2017). Neuroscience-Inspired Artificial

IP
Intelligence. Neuron 95, 245-258.
Huber, D., Gutnisky, D.A., Peron, S., O'Connor, D.H., Wiegert, J.S., Tian, L., Oertner, T.G., Looger, L.L.,

R
and Svoboda, K. (2012). Multiple dynamic representations in the motor cortex during sensorimotor

SC
learning. Nature 484, 473-478.
Hupe, J.M., James, A.C., Payne, B.R., Lomber, S.G., Girard, P., and Bullier, J. (1998). Cortical feedback

U
improves discrimination between figure and background by V1, V2 and V3 neurons. Nature 394, 784-
787.
N
Khan, A.G., Poort, J., Chadwick, A., Blot, A., Sahani, M., Mrsic-Flogel, T.D., and Hofer, S.B. (2018).
A
Distinct learning-induced changes in stimulus selectivity and interactions of GABAergic interneuron
M

classes in visual cortex. Nat Neurosci 21, 851-859.

Ko, H., Cossell, L., Baragli, C., Antolik, J., Clopath, C., Hofer, S.B., and Mrsic-Flogel, T.D. (2013). The
D

emergence of functional microcircuits in visual cortex. Nature 496, 96-100.

Ko, H., Hofer, S.B., Pichler, B., Buchanan, K.A., Sjostrom, P.J., and Mrsic-Flogel, T.D. (2011). Functional
TE

specificity of local synaptic connections in neocortical networks. Nature 473, 87-91.

Kording, K.P., and Wolpert, D.M. (2006). Bayesian decision theory in sensorimotor control. Trends
EP

Cogn Sci 10, 319-326.

Krupa, D.J., Wiest, M.C., Shuler, M.G., Laubach, M., and Nicolelis, M.A. (2004). Layer-specific
CC

somatosensory cortical activation during active tactile discrimination. Science 304, 1989-1992.
Kwon, S.E., Yang, H., Minamisawa, G., and O'Connor, D.H. (2016). Sensory and decision-related activity
propagate in a cortical feedback loop during touch perception. Nat Neurosci 19, 1243-1249.
A

Le Merre, P., Esmaeili, V., Charriere, E., Galan, K., Salin, P.A., Petersen, C.C.H., and Crochet, S. (2018).
Reward-Based Learning Drives Rapid Sensory Signals in Medial Prefrontal Cortex and Dorsal
Hippocampus Necessary for Goal-Directed Behavior. Neuron 97, 83-91 e85.
LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature 521, 436-444.

12
 Makino

Lee, W.C., Bonin, V., Reed, M., Graham, B.J., Hood, G., Glattfelder, K., and Reid, R.C. (2016). Anatomy
and function of an excitatory network in the visual cortex. Nature 532, 370-374.
Leinweber, M., Ward, D.R., Sobczak, J.M., Attinger, A., and Keller, G.B. (2017). A Sensorimotor Circuit
in Mouse Cortex for Visual Flow Predictions. Neuron 95, 1420-1432 e1425.
Li, W., Ma, L., Yang, G., and Gan, W.B. (2017). REM sleep selectively prunes and maintains new
synapses in development and learning. Nat Neurosci 20, 427-437.
Makino, H., Hwang, E.J., Hedrick, N.G., and Komiyama, T. (2016). Circuit Mechanisms of Sensorimotor
Learning. Neuron 92, 705-721.

T
Makino, H., and Komiyama, T. (2015). Learning enhances the relative impact of top-down processing

IP
in the visual cortex. Nat Neurosci 18, 1116-1122.
Makino, H., Ren, C., Liu, H., Kim, A.N., Kondapaneni, N., Liu, X., Kuzum, D., and Komiyama, T. (2017).

R
Transformation of Cortex-wide Emergent Properties during Motor Learning. Neuron 94, 880-890 e888.

SC
Manita, S., Suzuki, T., Homma, C., Matsumoto, T., Odagawa, M., Yamada, K., Ota, K., Matsubara, C.,
Inutsuka, A., Sato, M., et al. (2015). A Top-Down Cortical Circuit for Accurate Sensory Perception.
Neuron 86, 1304-1316.
U
Maret, S., Faraguna, U., Nelson, A.B., Cirelli, C., and Tononi, G. (2011). Sleep and waking modulate
N
spine turnover in the adolescent mouse cortex. Nat Neurosci 14, 1418-1420.
A
Masamizu, Y., Tanaka, Y.R., Tanaka, Y.H., Hira, R., Ohkubo, F., Kitamura, K., Isomura, Y., Okada, T., and
M

Matsuzaki, M. (2014). Two distinct layer-specific dynamics of cortical ensembles during learning of a
motor task. Nat Neurosci 17, 987-994.
D

Mignard, M., and Malpeli, J.G. (1991). Paths of information flow through visual cortex. Science 251,
1249-1251.
TE

Miller, E.K., and Cohen, J.D. (2001). An integrative theory of prefrontal cortex function. Annu Rev
Neurosci 24, 167-202.
EP

Miyamoto, D., Hirai, D., Fung, C.C., Inutsuka, A., Odagawa, M., Suzuki, T., Boehringer, R., Adaikkan, C.,
Matsubara, C., Matsuki, N., et al. (2016). Top-down cortical input during NREM sleep consolidates
CC

perceptual memory. Science 352, 1315-1318.

Moore, T., and Armstrong, K.M. (2003). Selective gating of visual signals by microstimulation of frontal
cortex. Nature 421, 370-373.
A

Narayanan, N.S., and Laubach, M. (2006). Top-down control of motor cortex ensembles by
dorsomedial prefrontal cortex. Neuron 52, 921-931.
Nienborg, H., and Cumming, B.G. (2009). Decision-related activity in sensory neurons reflects more
than a neuron's causal effect. Nature 459, 89-92.

13
 Makino

Norimoto, H., Makino, K., Gao, M., Shikano, Y., Okamoto, K., Ishikawa, T., Sasaki, T., Hioki, H., Fujisawa,
S., and Ikegaya, Y. (2018). Hippocampal ripples down-regulate synapses. Science.
Olsen, S.R., Bortone, D.S., Adesnik, H., and Scanziani, M. (2012). Gain control by layer six in cortical
circuits of vision. Nature 483, 47-52.
Otazu, G.H., Chae, H., Davis, M.B., and Albeanu, D.F. (2015). Cortical Feedback Decorrelates Olfactory
Bulb Output in Awake Mice. Neuron 86, 1461-1477.
Perez, M.A., Tanaka, S., Wise, S.P., Sadato, N., Tanabe, H.C., Willingham, D.T., and Cohen, L.G. (2007).
Neural substrates of intermanual transfer of a newly acquired motor skill. Curr Biol 17, 1896-1902.

T
Peters, A.J., Chen, S.X., and Komiyama, T. (2014). Emergence of reproducible spatiotemporal activity

IP
during motor learning. Nature 510, 263-267.
Peters, A.J., Liu, H., and Komiyama, T. (2017). Learning in the Rodent Motor Cortex. Annu Rev Neurosci.

R
Poort, J., Khan, A.G., Pachitariu, M., Nemri, A., Orsolic, I., Krupic, J., Bauza, M., Sahani, M., Keller, G.B.,

SC
Mrsic-Flogel, T.D., et al. (2015). Learning Enhances Sensory and Multiple Non-sensory Representations
in Primary Visual Cortex. Neuron 86, 1478-1490.

U
Rao, R.P., and Ballard, D.H. (1999). Predictive coding in the visual cortex: a functional interpretation
of some extra-classical receptive-field effects. Nat Neurosci 2, 79-87.
N
Reynolds, J.H., Pasternak, T., and Desimone, R. (2000). Attention increases sensitivity of V4 neurons.
A
Neuron 26, 703-714.
M

Robertson, E.M., Pascual-Leone, A., and Press, D.Z. (2004). Awareness modifies the skill-learning
benefits of sleep. Curr Biol 14, 208-212.
D

Stickgold, R., James, L., and Hobson, J.A. (2000). Visual discrimination learning requires sleep after
training. Nat Neurosci 3, 1237-1238.
TE

Tomita, H., Ohbayashi, M., Nakahara, K., Hasegawa, I., and Miyashita, Y. (1999). Top-down signal from
prefrontal cortex in executive control of memory retrieval. Nature 401, 699-703.
EP

Tononi, G., and Cirelli, C. (2014). Sleep and the price of plasticity: from synaptic and cellular
homeostasis to memory consolidation and integration. Neuron 81, 12-34.
CC

Velez-Fort, M., Rousseau, C.V., Niedworok, C.J., Wickersham, I.R., Rancz, E.A., Brown, A.P., Strom, M.,
and Margrie, T.W. (2014). The stimulus selectivity and connectivity of layer six principal cells reveals
cortical microcircuits underlying visual processing. Neuron 83, 1431-1443.
A

Vinje, W.E., and Gallant, J.L. (2000). Sparse coding and decorrelation in primary visual cortex during
natural vision. Science 287, 1273-1276.
Vyazovskiy, V.V., Olcese, U., Lazimy, Y.M., Faraguna, U., Esser, S.K., Williams, J.C., Cirelli, C., and Tononi,
G. (2009). Cortical firing and sleep homeostasis. Neuron 63, 865-878.

14
 Makino

Walker, M.P., Brakefield, T., Morgan, A., Hobson, J.A., and Stickgold, R. (2002). Practice with sleep
makes perfect: sleep-dependent motor skill learning. Neuron 35, 205-211.
Wertz, A., Trenholm, S., Yonehara, K., Hillier, D., Raics, Z., Leinweber, M., Szalay, G., Ghanem, A., Keller,
G., Rozsa, B., et al. (2015). PRESYNAPTIC NETWORKS. Single-cell-initiated monosynaptic tracing
reveals layer-specific cortical network modules. Science 349, 70-74.
Witt, K., Margraf, N., Bieber, C., Born, J., and Deuschl, G. (2010). Sleep consolidates the effector-
independent representation of a motor skill. Neuroscience 171, 227-234.
Wolpert, D.M., Ghahramani, Z., and Flanagan, J.R. (2001). Perspectives and problems in motor

T
learning. Trends Cogn Sci 5, 487-494.

IP
Xiao, L.Q., Zhang, J.Y., Wang, R., Klein, S.A., Levi, D.M., and Yu, C. (2008). Complete transfer of
perceptual learning across retinal locations enabled by double training. Curr Biol 18, 1922-1926.

R
Xu, N.L., Harnett, M.T., Williams, S.R., Huber, D., O'Connor, D.H., Svoboda, K., and Magee, J.C. (2012).

SC
Nonlinear dendritic integration of sensory and motor input during an active sensing task. Nature 492,
247-251.

U
Xu, T., Yu, X., Perlik, A.J., Tobin, W.F., Zweig, J.A., Tennant, K., Jones, T., and Zuo, Y. (2009). Rapid
formation and selective stabilization of synapses for enduring motor memories. Nature 462, 915-919.
N
Yang, G., Lai, C.S., Cichon, J., Ma, L., Li, W., and Gan, W.B. (2014). Sleep promotes branch-specific
A
formation of dendritic spines after learning. Science 344, 1173-1178.
M

Zhang, S., Xu, M., Kamigaki, T., Hoang Do, J.P., Chang, W.C., Jenvay, S., Miyamichi, K., Luo, L., and Dan,
Y. (2014). Selective attention. Long-range and local circuits for top-down modulation of visual cortex
D

processing. Science 345, 660-665.

TE
EP
CC
A

15
 Makino

Figures

T
R IP
SC
Figure 1. Evidence supporting the involvement of higher-order brain areas for perceptual and motor
learning in humans.
U
(A) In perceptual learning, an ability to visually discriminate different contrasts at one location of
N
a visual field can be transferred to another location if the subject is previously trained to
A
discriminate another feature such as orientation at the second location. Since receptive fields
M

of the primary visual cortical neurons are small, higher-order brain areas with larger receptive
fields should be involved in the transfer learning. (B) In motor learning, motor skills, such as
D

execution of a series of key presses with fingers with one hand, are transferred to the opposite
hand so that the performance is better than subjects without the training.
TE
EP
CC
A

16
 Makino

T
R IP
Figure 2. Learning-related recruitment of top-down control.

SC
(A) Left. Two-photon calcium imaging of distinct microcircuit components within the mouse visual
cortex uncovered that top-down axonal activity from the retrosplenial cortex (red) increases while
bottom-up L4 activity (blue) decreases during sensory learning (Makino and Komiyama, 2015). Line

U
thickness indicates activity strength of each circuit element but does not reflect its connection
N
strength. Green cells indicate L2/3 excitatory neurons. Right. Example GCaMP6-expressing cell bodies
A
(green) and axons (white). Red neurons are tdTomato-expressing inhibitory neurons. (B) Left. Wide-
field calcium imaging of the entire dorsal mouse cortex revealed macroscopic changes in peri-
M

movement activity dynamics during motor learning, namely emergence of a new activity flow from
the premotor cortex (Makino et al., 2017). Right. Changes in the activity flow during motor learning.
D

Activity propagates along the gradient of the response latency (from blue to red in the color bar). Note
TE

that after learning a secondary activity flow appears from the premotor area. Importantly, in both
studies inactivation of respective top-down sources reversed learning-related changes in neural
activity patterns. These studies using different imaging technologies in distinct domains of learning
EP

uncovered a unified principle of learning, namely, recruitment of top-down control.

CC
A

17