Cerebral Cortex, September 2021;31: 4220–4232
https://doi.org/10.1093/cercor/bhab080
Advance Access Publication Date: 12 April 2021
Original Article
ORIGINAL ARTICLE
The Brain Circuits and Dynamics of Curiosity-Driven
Behavior in Naturally Curious Marmosets
Xiaoguang Tian1,2 , Afonso C. Silva1,2 and Cirong Liu1,2,3,4
1 Department
of Neurobiology, University of Pittsburgh Brain Institute, University of Pittsburgh, Pittsburgh PA
Microcirculation Section, Laboratory of Functional and Molecular Imaging, National
Institute of Neurological Disorders and Stroke, National Institutes of Health, Bethesda MD 20892, USA,
3 Institute of Neuroscience, Center for Excellence in Brain Science and Intelligence Technology, Chinese
Academy of Sciences, Shanghai 200031, China and 4 Shanghai Center for Brain Science and Brain-Inspired
Intelligence Technology, Shanghai 201210, China
15261, USA, 2 Cerebral
Address correspondence to Dr Cirong Liu. Email: crliu@ion.ac.cn; Dr Afonso C. Silva. Email: afonso@pitt.edu.
Abstract
Curiosity is a fundamental nature of animals for adapting to changing environments, but its underlying brain circuits and
mechanisms remain poorly understood. One main barrier is that existing studies use rewards to train animals and motivate
their engagement in behavioral tasks. As such, the rewards become significant confounders in interpreting curiosity. Here,
we overcame this problem by studying research-naïve and naturally curious marmosets that can proactively and
persistently participate in a visual choice task without external rewards. When performing the task, the marmosets
manifested a strong innate preference towards acquiring new information, associated with faster behavioral responses.
Longitudinally functional magnetic resonance imaging revealed behavior-relevant brain states that ref lected choice
preferences and engaged several brain regions, including the cerebellum, the hippocampus, and cortical areas 19DI, 25, and
46D, with the cerebellum being the most prominent. These results unveil the essential brain circuits and dynamics
underlying curiosity-driven activity.
Key words: fMRI, functional connectivity, natural viewing, nonhuman primates, novelty seeking
Introduction
Living organisms, including humans, learn to use external
stimuli information to adapt to rapidly changing environments.
A decision-making strategy is reinforced when the choice
increases reward probability. However, organisms can also use
an alternative decision-making strategy of exploration, in which
the choice is not associated with any external rewards (Davis
et al. 1950; Harlow et al. 1950; Kidd and Hayden 2015). One innate
motivation for such a strategy is the brain’s information-seeking
drive that intrinsically urges the organisms to explore their
environments (Anselme 2010; Bromberg-Martin and Hikosaka
2011; Funamizu et al. 2012), similar to children playing and
adult hobbies. Although these intrinsically motivated behaviors
depend on internal factors that are difficult to characterize, one
particular intrinsic motivation, curiosity, provides opportunities
to investigate its underlying brain mechanisms.
Although curiosity engages multiple brain regions and
associated cognitive processes, it does not have a well-defined
psychological definition. Curiosity is a natural and active
exploratory process to gain new information. As one of our most
basic natures, curiosity generally induces many essential behaviors to explore activities or stimuli that are novel, surprising, or
intriguingly complex (Kidd and Hayden 2015). Yet, the neural
mechanisms underlying curiosity remain poorly understood.
Many studies about curiosity have been performed on human
subjects. However, their curiosity measures are highly subjective
(van Lieshout et al. 2018; Kobayashi et al. 2019; Schwartenbeck
et al. 2019). Other studies on curiosity used animal models,
for example, nonhuman primates (NHP) (Redgrave and Gurney
© The Author(s) 2021. Published by Oxford University Press. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com
Brain Circuits of Curiosity-Driven Behavior
2006; Bromberg-Martin and Hikosaka 2011). As animals cannot
directly report internal mental activities, these experiments
made use of external reward modulation (e.g., water, juice,
treats) to elicit exploratory behaviors and characterize curiosity
by measuring their willingness to sacrifice reward (Redgrave and
Gurney 2006; Bromberg-Martin and Hikosaka 2011). As such,
the rewards become significant confounders in interpreting
curiosity. However, the use of external rewards is not necessary
to elicit curiosity-driven behavior. For example, human infants
naturally explore new environments regardless of physical
rewards (Berlyne and Slater 1957; Kreitler et al. 1984). Macaques
also exhibit a robust motivation to solve mechanical puzzles
without extrinsic incentives (Davis et al. 1950; Harlow 1950).
Therefore, NHP can produce curiosity-driven behaviors without
external rewards.
Our study aimed to probe curiosity-related behaviors and
brain activities in NHP using functional magnetic resonance
imaging (fMRI). We chose to use the common marmoset, one
of the smallest but highly social NHP species. We purposely
recruited a large cohort of marmosets to identify naturally curious monkeys that can actively engage in the experiments without reward-based training. Furthermore, marmosets are easy
to be handled, constrained, and acclimatized for awake magnetic resonance imaging (MRI) scanning (Silva et al. 2011; Liu
et al. 2019), which provides critical practical advantages to track
their behavioral changes and fMRI responses from a researchnaïve state. To elicit curiosity-driven behaviors, we used a classic delayed free-choice task without the influence of external rewards (Fig. 1A), so that the choice depends entirely on
the animals’ innate preferences. In this task, animals choose
freely between two target images (novel and familiar) after
watching a short movie clip. To minimize the influence of any
reward-driven experience, we recruited research-naïve animals
for our study and monitored their behavioral changes and fMRI
responses longitudinally.
Materials and Methods
Animals
All procedures were approved by the Animal Care and Use Committee of the National Institute of Neurological Disorders and
Stroke, National Institutes of Health. A total of 15 healthy adult
marmosets (nine males and six females, 3–8 years old) were
recruited and screened for the study (the detailed information
is shown in Supplementary Table S1, sheet 1). All marmosets
had received a 3-week acclimatization training for MRI scans so
that they could be familiar with MRI environments and endured
long-term MRI scanning with minimum stress. The MRI acclimatization training involved no rewards, and the detailed protocol
was described in our previous paper (Silva et al. 2011; Liu et al.
2019). After the adaptation period, marmosets underwent longitudinal MRI scanning experiments with calibration and delayed
free-choice tasks. We did not implement any restriction of food
or fluid intake for the marmosets before the experiments. Note
that we attempted to deliver random liquid rewards to these
marmosets on the first day after each successful trial regardless
of their choices. However, since the marmosets never received
any previous training, they seldomly consumed the rewards
delivered. Therefore, we discontinued offering rewards after the
first day. We excluded any animals that failed to get involved in
the tasks from our experiments. Among the 15 marmosets, only
three individuals were actively and persistently engaged in the
Tian et al.
4221
task (see Supplementary Table S1, sheet 1 for detail). The three
marmosets included one male (named “A”) and two females (“K”
and “B”).
Behavioral Tasks
Marmosets wore customized jackets and helmets to stabilize
their heads inside the MRI (Silva et al. 2011). A computer monitor
was set in front of the animal with a resolution of 1920 × 1080.
The maximum viewing angles were restricted within 10◦ on
the horizontal axis and 8◦ on the vertical axis, and the eye
movements were sampled at 1 kHz by the infrared eye tracker
(Arrington Research, Inc.). Before the delayed free-choice task,
marmosets underwent an eye-tracking calibration task to obtain
accurate eye positions. In the calibration task, marmosets
directed their gaze to the center of the screen when a center
white dot (diameter 1◦ of visual angle) appeared. After the
fixation within a 2◦ radius of visual angle for 100–200 ms, the
central dot disappeared. A face image randomly selected from
a prepared dataset (54 pieces in total) appeared on the screen
in 21 random positions (each lasting 1 s). The image attracted
animals to make a saccade to them, and the landing positions
were manually verified and extracted for the offline calibration
(an example in Supplementary Fig. S1). The number of daily
trials varied between 105 and 147, depending on the animal’s
performance. In general, the percentage of valid trials for the
calibration was >90%.
The delayed free-choice task involved four periods: the
movie-watching, the delay, the decision-making, and the resting
periods (Fig. 1A). Marmosets fixated on a central white circle
within a 2◦ radius of visual angles for a random period of 1–1.6 s
to initiate playing a 20 s video clip. There were 440 clips in our
video database. Videos clips were chosen to minimize content
repetition. Supplementary Table S1 (sheets 2–4) describes the
number of video clips played in a random sequence for each
marmoset. After the video clip ended, the central white circle
reappeared for a randomly variable period of 1.5–2.5 s to allow
marmosets to regain fixation. The marmoset had to maintain
its gaze within a 2◦ visual angle radius around the fixation dot,
during which (500 ms right after the circle appeared) two target
images were randomly shown at the horizontal ±5◦ within the
size of 6◦ × 5◦ . One of the targets was from the played video, and
the other was a novel image unrelated to the video and never
seen by the marmosets before. The whole delay process lasted
1.5–2.5 s. The choice consisted of the first response saccade after
the central white fixation circle disappeared (fixation offset).
After the decision period (6 s), a resting period started with a
blank screen. An example trial of the delayed free-choice task is
shown in Supplementary Movie S1.
fMRI Data Acquisition
MRI scans were conducted in a 7 T/300 mm magnet (Bruker
Biospin, Inc.) with a 150 mm gradient set (450 mT/m with
150 μs; Resonance Research Inc.), a 16-rung high-pass birdcage
radiofrequency transmission coil, a custom-built eight-channel
phased-array receiver coil, and a ParaVision 6.0.1 system. The
fMRI data were collected with a 2D gradient-echo Echo-planar
imaging (EPI) sequence [repetition time (TR) = 2 s, echo time
(TE) = 22.2 ms, flip angle = 70.4◦ , field of view (FOV) = 28 × 36 mm,
matrix size = 56 × 72, 38 axis slices with a 0.5 mm slice thickness,
voxel resolution = 0.5 mm isotropic, 180 time points and 6 min
per run]. For each session, half of the runs were collected with
4222
Cerebral Cortex, 2021, Vol. 31, No. 9
Figure 1. The longitudinal performance in the delayed free-choice task. (A) Task design: Marmosets foveated to a central fixation dot for 1–1.6 s to initiate the playing
of a 20 s video clip. After the video, the fixation dot reappeared for 1.5–2.5 s to allow marmosets to return fixation to the center, during which two target images were
shown (one was from the played video, and the other was a novel image). The animals’ choice consisted of the first saccade made after the fixation dot disappeared.
After the decision period (6 s), the marmosets rested for 16 s before the next trial. (B) The behavioral performances for each marmoset: The top-left panel shows the
proportion of choices to the novel target (the black dashed line highlights the 50% probability). The top-right panel shows the percentage difference in viewing time
between two targets (novel–familiar) during the decision period. The black dashed line highlights the same viewing time in two targets. The bottom-left panel depicts
the average reaction time of the choice saccades (ms). The bottom-right panel shows the average peak velocities of the choice saccades (deg/s). (C) Comparison of the
behavioral responses across days. Behavioral data were split into two halves by the median of days for each marmoset (Day 1–5 and 6–11 for the marmoset A; Day 1–6
and 7–13 for the marmoset K; Day 1–3 and 4–6 for marmoset B). The opaque colors highlight the pooled data of later days, and transparent colors represent early day’s
data. The first column depicts the relationship between the reaction time (x-axis) and the amplitude (y-axis) of the choice saccades, and the second column shows
the relationship between the peak velocities (x-axis) and the amplitude (y-axis). The later-training days show significant differences in these measures between the
novel and the familiar target choices (Wilcoxon rank-sum test, see detailed statistics in Supplementary Table S4). All data are presented as mean ± standard error of
the mean and the statistic test ∗P < 0.05, ∗∗P < 0.01, ∗∗∗P < 0.001, corrected.
the L-to-R (blip-up) phase-encoding direction and the other
half with the R-to-L (blip-down) phase-encoding direction to
compensate for different EPI distortion and signal drop-out.
Also, two sets of spin-echo EPI with opposite phase-encoding
directions (LR and RL) were collected for the EPI-distortion correction (TR = 3 s, TE = 0.44 ms, flip angle = 90◦ , FOV = 28 × 36 mm,
matrix size = 56 × 72, 38 axis slices with a 0.5 mm slice
thickness, 8 volumes per set). In each session, a T2 -weighted
Brain Circuits of Curiosity-Driven Behavior
structural image was scanned for the co-registration purpose
(TR = 6000 ms, TE = 9 ms, flip angle = 90◦ , FOV = 28 × 36 mm,
matrix size = 112 × 144, 38 axis slices, resolution = 0.25 × 0.25 ×
0.5 mm3 , number of averages = 8).
A total of 11 MRI sessions (one session per day) were collected for marmoset A, 13 sessions for marmoset K, and six
sessions for marmoset B. Note that the marmoset B was not
available after 6-day experiments due to an accident in the
marmoset housing facility. There were 6–14 runs in each session, varying daily depending on the animal’s performance (see
Supplementary Table S1, sheets 2–4 for detail). Each run lasted
for 6 min and contained seven trials.
Data Analysis
Preprocessing of Behavioral Data
For the eye position calibration, eye landing positions for the 21
screen locations were manually extracted and then fitted with a
polynomial method described in our previous studies (Tian et al.
2016, 2018). Supplementary Figure S1 shows an example of calibration. The delayed free-choice task’s eye movements were also
verified in post hoc analyses to correct misses or false detections
based on the eye movement velocity and accelerations. The
updated eye movement data were used to characterize choice
properties (the choices to novel or familiar images, reaction
time (RT) times, peak velocities, and amplitudes) and to exclude
invalid trials, in which the marmoset quitted the task, failed
to maintain fixation during the delay period, or made saccadic
choices too early (<50 ms) or too late (>700 ms).
Preprocessing of fMRI Data
Data preprocessing was performed using the Analysis of Functional NeuroImages (AFNI) (Cox 2012) and the FMRIB Software
Library (FSL) (Jenkinson et al. 2012) software packages. For each
run, the first four time points were removed for the magnetization to reach a steady state. The remaining time series
was despiked by the “3dDespike”, slice timing corrected by the
“3dTshift,” motion corrected by the “3dvolreg” of the AFNI, and
corrected for EPI distortions by the “topup” of the FSL. Motion
censoring, nuisance signals removal, and band passing were
performed in one regression model by the “3dDeconvolve” of
the AFNI. In detail, any TRs and the previous TRs were censored if the motion was >0.2 mm to minimize the influence of
head movements. Nuisance signals were regressed, including
demeaned and derivatives of motion parameters, the mean
white matter signal, and the mean cerebrospinal fluid signal.
No extra spatial smoothing was performed as our analyses
were region of interest (ROI) based. To examine the effect of
frequencies, we tested different band-passing filters, including
0.004–0.05 Hz (extremely low frequency), 0.01–0.1 Hz (low frequency), and 0.1–0.2 Hz (high frequency). By repeating the same
analysis (see below “Quantification of behavior-most-correlated
states”), the data with low-frequency filtering provided the most
significant contributions [Supplementary Fig. S2, tested by oneway analysis of variance (ANOVA), see detailed statistics in
Supplementary Table S4]. The results and conclusions from the
data with low-frequency filtering were similar to the data without band-passing filtering. In the main figures, we only reported
the results from the unfiltered data.
Dynamic Connectivity Analysis
We extracted the time series of 130 brain regions defined by
the Paxinos parcellation of the Marmoset Brain Mapping Atlas
Tian et al.
4223
(Liu et al. 2018, 2020) and used a sliding window method to
calculate the dynamic connectivity (Allen et al. 2014). The sliding
window was defined as a length of 3 TRs (6 s) and a step of 1 TR
(2 s), which generated 130 × 130 covariance matrices Dr,s,w for
each run r, each session s, and each time point w (Choe et al.
2017; Preti et al. 2017). In previous studies (Hutchison et al. 2013;
Barttfeld et al. 2015), the covariance was estimated from the regularized inverse covariance matrix by the graphical LASSO (least
absolute shrinkage and selection operator) method with an
additional L1 norm constraint on the inverse covariance matrix
to enforce sparsity. Here, we adopted l = 0.1 as most previous
studies (Friedman et al. 2008; Barttfeld et al. 2015; Cai et al. 2018).
The covariance matrix of each time window was Fisher’s ztransformed (Zr,s,w = tan−1 Dr,s,w ) to z-score for further analysis.
In addition, the sliding window size did not affect results.
Unsupervised Clustering and Brain States
To identify the reoccurring connectivity patterns (brain states),
we applied the K-means clustering algorithm (with K from 3 to
10) on the covariance matrices Zr,s,w using the cosine distance
metric implemented in MATLAB 2018b. As the choice preference
was used to select the K (see the next section), we excluded
Zr,s,w matrices of the decision period (6 s) to avoid potential
biases, except for the special analyses in Figures 4 and 5 and
Supplementary Figure S5 (pooling all time periods). The resulting centroids (the median of clusters) of the K-mean algorithm
are referred to as the brain state BSn (as shown in Fig. 2B) and
each matrix Zr,s,w is assigned a BSn (we also tested other metrics
for the clustering, the Pearson’s correlation, and the Euclidean
distance, which resulted in similar results and conclusions).
To select the number (K) of brain states (Fig. 2A), we calculated the occurrence rate of each state n along with the time
dimension of each session s:
BSs
pns = n n s .
i=1 BSn
Then, we computed the Pearson’s correlation coefficient
between the longitudinal change of the occurrence rate and
the probability of choosing novel targets. The selected number
of K and the behavior-most-correlated state were the ones with
the highest correlation, respectively.
Normalized Occurrence Rates in Different Periods
We applied a running window of 1 s with a step of 100 ms to
compute averaged occurrences rate (over all sessions) in the
periods of movie-watching, resting, and decision (Figs 3–5). Due
to different total lengths of movies, resting, delay, and decision
periods, the occurrence rates were normalized by their length
ratio in each session (we also tested different sizes of running
windows, e.g., 2 s, and obtained similar patterns). Based on the
same analysis, we further quantified the normalized occurrence
rates under different choice conditions (choices to novel or
familiar images) and the rate difference between two choice
conditions (novel–familiar). In addition, the sliding window size
did not affect results.
The Similarity of Normalized Occurrence Rates (Coupling Periods)
across Monkeys
We repeated the above analysis for each marmoset to obtain
the differences in occurrence rates between two choice conditions of different periods (novel–familiar). We then estimated
the autocorrelation of such rate-difference curves across three
4224
Cerebral Cortex, 2021, Vol. 31, No. 9
Figure 2. Dynamic functional connectivity reveals behavior-most-correlated brain states. (A) The number of clusters (brain states) in K-mean clustering. We explored
different numbers of clusters by the K-means algorithm (K = 3–10) to find brain states with a high correlation with choice preferences. Each gray dot represents
a correlation value between each brain state’s occurrence rates across days and the longitudinal changes of the choice preference. Each colored star represents the
average Pearson’s correlation value of all states for each K, and red arrowheads highlight the clusters with the highest mean correlation values. The data was presented
as mean ± standard error of the mean. (B) Brain states for the selected K-mean cluster ranked by their Fisher’s z-transformed correlation values with the behavior data.
For each animal, we selected the K-mean cluster with the highest mean correlation from above result A (K = 7 for marmoset A; K = 5 for marmosets K and B). Black
outlines highlight the brain states with the highest correlation in each marmoset. (C) The behavior-most-correlated brain states across different marmosets share
common components, indicated by their highest similarities. Black outlines highlight the correlation values of one marmoset’s behavior-most-correlated brain state
to all brain states of another marmoset.
Brain Circuits of Curiosity-Driven Behavior
Tian et al.
4225
Figure 3. Occurrence rates of brain states during movie-watching and resting periods. For the behavior-most-correlated states (A) and the behavior-least-correlated
states (B), the left panels present the normalized occurrence rates with choices to novel targets (red lines) and familiar targets (blue lines), as well as the difference
between these two occurrence rates (novel–familiar, black lines), in the different periods (movie and resting). The right panels depict across-marmoset patterns
of their occurrence rate differences. The first row shows the pooling results of the occurrence rate difference (novel–familiar) from the three marmosets for
each period. The second row shows the autocorrelation spectrums of the occurrence rate differences across marmosets at time lag 0. The spectrums during
the movie-watching uncover the coupling periods across marmosets (highlighted by orange arrowheads), including 2–7 s, 13–18 s after the movie onset (the first
column), but not in the resting period (the second column) or any periods of the behavior-least-correlated state. Data are presented as mean ± standard error of the
mean.
monkeys at the lag 0 (implemented by the MATLAB 2018b). Given
the rate-difference curves of three monkeys, Y = Y1 , Y2 , Y3 , at
time T1 , T2 , . . . , TN , the spectrum of lag 0 autocorrelation function
at time r1 , r2 , . . . , rN is
3 3
Yj=i − Y
i=1
j=1 Yi − Y
.
r=
2
3
Y
−
Y
i
i=1
The resulting spectrum was smoothed with a small window
(0.1 × 0.1) for visualization. Higher values in the spectrum mean
higher similarity of the temporal pattern of rate-difference
curves across monkeys.
Quantification of Behavior-Most-Correlated States
For each brain connection, we counted how many other
states were significantly different from the behavior-mostcorrelated state in their correlation values (repeated oneway ANOVA measures, P < 0.05, see detailed statistics in
4226
Cerebral Cortex, 2021, Vol. 31, No. 9
Figure 4. Occurrence rates of brain states around the decision period. For the behavior-most-correlated states (A) and the behavior-least-correlated states (B), the left
panels present the normalized occurrence rates with the choices to the novel targets (red lines) and the familiar targets (blue lines), as well as the difference between
these two occurrence rates (novel–familiar, black lines), around the decision period (aligned on the fixation offset). The light gray areas are the range of response
times of all trials. The right panels present the analysis of the occurrence rate difference across marmosets. The first row shows the pooling results of the occurrence
rate difference (novel–familiar) from the three marmosets. The second row shows the autocorrelation spectrums of the occurrence rate difference at the time lag
0. The spectrums reveal a prominent coupling time across marmosets, which is 1 s around fixation offset (highlighted with the orange arrowheads), but not in the
behavior-least-correlated states. Data are presented as mean ± standard error of the mean (SEM). (C) The longitudinal change of the occurrence rate pattern of the
behavior-most-correlated states at the time of delay and decision periods. The left panels show the daily changes of the normalized occurrence rate of the states during
the 1 s before the fixation offset. The right panels show the pooled results of the normalized occurrence rate and across-subject coupling during the early training
days and late training days, respectively. (D) The longitudinal change of the occurrence rate pattern of the behavior-most-correlated states around 1 s of fixation offset.
The left panels show the daily changes of the normalized occurrence rate of behavior-most-correlated states during the 1 s after the fixation offset. The right-top
panel shows the pooled results of the normalized occurrence rate and across-subject coupling during the early training days and late training days, respectively. The
bottom-right panels show the normalized probability histogram of the corresponding saccades for target selection. Data are presented as mean ± SEM.
Brain Circuits of Curiosity-Driven Behavior
Tian et al.
4227
Figure 5. The key brain regions in the behavior-most-correlated brain state. (A) The right histograms present the ranking indices of all brain regions in each marmoset
(normalized number of significant functional connections; Supplementary Table S2 provides the detailed ranking information). The arrows mark the top 10 regions.
The left pie charts summarize the proportion of the normalized values in the top 10 ranked regions in each marmoset, highlighting five overlapped brain regions across
marmosets. (B) The left panel displays the five top-ranked regions across all marmosets; the pie charts show the normalized ranking indices for the five regions (from
inner to outside represent the data of marmosets A, K, and B, respectively). The matrices on the right show the normalized number of significant states among these
five brain regions, demonstrating the cerebellum’s core role (tested by repeated-measures one-way ANOVA; see detailed statistics in Supplementary Table S4). For
verification, Supplementary Figure S5A displays extended results for the top 10 regions across all marmosets, and Supplementary Figure S5B shows the top five areas
for each marmoset. (C) The most significant areas in the cerebellum. The first panel shows the lobule parcellations of the cerebellum. The second panels show the 48
ROIs defined in the cerebellum (24 per hemisphere). The red nodes represent the ROIs that are more significant than the entire cerebellum in all three marmosets (see
detailed statistics in Supplementary Table S4). The orange nodes show the ROIs that are more significant in any two marmosets. The gray nodes show less significant
ROIs than the entire cerebellum in at least two marmosets.
Supplementary Table S4). The numbers were then normalized
to a range of 0–1 (ranking index) by scaling normalization.
To identify the contributions of brain areas to the behaviormost-correlated states, we summed these numbers of total
connections in each brain region and sorted them in descending
order. The most important brain area had the highest value
[Fig. 5A and Supplementary Tables S2 (sheets 2–4) and S3]. We
also summed the numbers from three monkeys and sorted them
in descending order to examine the contribution of each brain
area at the population level (Fig. 5B and Supplementary Table S2,
sheet 1).
Seed-Based Analysis of the Cerebellum
A total of 24 ROIs (sphere, diameter = 2 mm) were evenly
distributed across each cerebellum hemisphere (48 in total).
The mean time series of each ROI was extracted from the
preprocessed fMRI data and replaced the whole cerebellum’s
mean time series to calculate dynamic functional connectivity
(FC) (see “Dynamic connectivity analysis”). The resulting
dynamic FC then replaced the corresponding data of the
cerebellum in each brain state. With the replaced data, we
repeated the analysis in “Quantification of behavior-mostcorrelated states” and obtained the contributions of each ROI
4228
Cerebral Cortex, 2021, Vol. 31, No. 9
to the behavior-most-correlated states. The number of total
significant states of each ROI was then compared with the whole
cerebellum’s previous corresponding results to identify the ROIs
that contributed more significantly than the whole cerebellum
(tested by repeated-measures ANOVA, see detailed statistics in
Supplementary Table S4).
Results
The Behavioral Performance Reveals an Innate
Preference Towards the Novel Information
We monitored the behavioral performance of research-naïve
marmosets in a delayed free-choice task. The task required
marmosets to watch a 20 s video clip, which had minimal repeats of the videos played each day across monkeys
(Supplementary Table S1, sheets 2–3). The marmosets then
freely choose between two target images: One selected from
the watched video or a novel image they had never seen before
(Fig. 1A). We recruited 15 research-naïve marmosets (details
are in Supplementary Table S1, sheet 1). The only training
these marmosets took was the necessary MRI acclimatization procedures before the experiment (Liu et al. 2019). As
these marmosets had not experienced any prior training for
behavioral experiments, nine animals showed no interest in
engaging in any visual tasks, even on the first day. Three of
them participated in the eye-calibration session and a few trials
of the delayed free-choice task. However, their engagement only
lasted 1 or 2 days, after which they quit most of the trials due
to a lack of motivation. The three remaining marmosets (A, K,
and B) showed active and persistent involvement in the task.
We recorded their behavioral performance and fMRI responses
longitudinally (11 days for the marmoset A, 13 days for K, and
6 days for B).
As the experiments included no external rewards, the
longitudinal changes of choice saccades and viewing time
reflected the marmosets’ innate preference. In the beginning,
their decision strategy formed under uncertainty, and their
choices were random (Fig. 1B, top left). As they became more
familiar with the experiments, they reinforced the strategy
of selecting novel images, which provided more information
than familiar images. The choice preference was not affected
by the stimuli spatial location because the novel images were
shown randomly on either the left or the right side of the central
fixation dot (Supplementary Fig. S3B, Wilcoxon rank-sum test,
P > 0.05, see detailed statistics in Supplementary Table S4). More
importantly, the animals’ viewing time of unknown images
increased, reflecting their clear preference for novel information
(Fig. 1B, top right, and Supplementary Fig. S3A).
The three marmosets also showed other consistent longitudinal behavioral changes, including a significant decrease
in faster responses (Fig. 1B, bottom panels) and invalid trials
(Supplementary Fig. S3C). In the beginning, about half of the
trials were invalid, in which the marmosets either quit the
trials, failed to maintain fixation during the delay period, or
made the saccadic choice responses too early (<50 ms) or too
late (>700 ms). As the marmosets became more familiar with
the task, the number of invalid trials dropped dramatically
(Supplementary Fig. S3C). As marmosets became more familiar
with the task, they also responded faster during the decisionmaking period, quantified by the average RT (Fig. 1B, bottom
left) and the average peak velocity (Fig. 1B, bottom right) of the
response saccades. For marmosets A, K, and B, the average RT
dropped significantly from 355, 403, and 381 ms (first day) to
239, 207, and 214 ms (last day), respectively. Correspondingly, the
average peak velocity increased significantly from 214, 212, and
212.3 deg/s to 328, 298.5, and 309.2 deg/s, respectively.
As previous studies suggested, we further examined whether
the novelty modulated the saccadic metrics (the response
speed) (Reppert et al. 2015; Ebitz et al. 2018). By comparing
the average RT and peak velocities of the two conditions
across days directly, we found no significant differences
between the choices of familiar targets and those of novel ones
(Supplementary Fig. S3D, Wilcoxon rank-sum test, P > 0.05; see
detailed statistics in Supplementary Table S4). As the saccadic
peak velocity and RT depend on the amplitude, the considerable
amplitude variation of the two conditions in each day may
account for the insignificance. To minimize the influences
of amplitude variation, we pooled the longitudinal dataset
into two halves (early training days and late training days)
by the median recording day for each marmoset (Day 1–
5 and 6–11 for the marmoset A; Day 1–6 and 7–13 for the
marmoset K; Day 1–3 and 4–6 for marmoset B). We measured
the amplitude–RT relationship (Fig. 1C, left column) and the
amplitude–peak–velocity relationship (Fig. 1C, right column).
We found a significant difference for each marmoset in the RT
and the average peak velocities between the two conditions
in the late training days (Fig. 1C, opaque color, Wilcoxon
rank-sum test: P < 0.01, corrected, see detailed statistics in
Supplementary Table S4), but not in the early training days
(Fig. 1C, transparent color, Wilcoxon rank-sum test: P > 0.05,
corrected, see detailed statistics in Supplementary Table S4).
Thus, when the animals became familiar with the task (late
training days), they preferred the novel images with faster
response times and higher peak velocities. The longitudinal
improvement also supports the idea of the novel image selection
and preference.
Behavior-Most-Correlated Brain States Reveal the
Mechanism of Choice Preference
After finding novel image selection and preference performance,
we then examined the temporal dynamics of brain connectivity
underlying innate preferences. From the fMRI data of each trial,
we extracted the time series of 130 brain regions defined by the
Paxinos parcellation of the Marmoset Brain Mapping Atlas (Liu
et al. 2018, 2020). We then applied a sliding window approach to
calculate dynamic FC patterns associated with each time point
in 130 × 130 correlation matrices. After clustering these matrices
into several brain states (K), we examined the Pearson’s correlation between the longitudinal changes of the choice preference
(shown in Fig. 1B, top left) and the occurrence rate changes
in brain states. The analysis revealed brain states in multiple
K solutions that were strongly correlated with the behavior
(Fig. 2A). These behavior highly correlated brain states provided
similar results in our following analysis. Here, we present brain
states from K = 7, 5, and 5 for marmosets A, K, and B as examples,
since these K solutions showed the highest mean correlations
(Fig. 2A). We ranked these selected brain states according to
their correlation levels with behavioral changes to identify the
behavior-most-correlated brain states (Fig. 2B). The behaviormost-correlated brain states shared the highest cross-subject
similarities, suggesting that these states were not random but
had consistent temporal and spatial features across marmosets
(Fig. 2C).
Brain Circuits of Curiosity-Driven Behavior
To characterize the behavior-most-correlated states’ temporal features, we pooled all trials of the two choice conditions. We
then plotted the normalized occurrence rates of the behaviormost-correlated state during the periods (movie-watching
before the decision period and the resting period after the
decision period). The temporal variations of the normalized
occurrence rates presented a similar pattern for the three
marmosets during the movie-watching periods, which was
prominent in the curve of occurrence rate difference (novel–
familiar) between the two choice conditions (Fig. 3A, left panel
first and second rows). The autocorrelation spectrum of the
curves at time lag 0 further highlighted a similar pattern by
revealing prominent coupling periods across marmosets at 2–
7 s and 13–18 s after the movie’s start (Fig 3A, right panel, column
1). The pattern only appeared during the movie-watching
period. We did not observe any consistent patterns across
marmosets during the resting period (Fig. 3A, left panel, 3–4
rows, and right panel, column 2) or in the other states with
low correlations, for example, the behavior-least-correlated
states (Fig. 3B). Therefore, the temporal pattern was specific for
the behavior highly correlated states and the movie-watching
period.
The coupling pattern of the behavior-most-correlated states
was unrelated to either eye movement patterns or movieshowing sequence, as saccade occurrence showed no significant
difference between the two conditions (Supplementary Fig. S4A,
Wilcoxon rank-sum test, P > 0.05, see detailed statistics in
Supplementary Table S4) and the movie-showing sequences
are random for different marmosets (Supplementary Fig. S4B;
see detailed statistics in Supplementary Table S4). The movie
contents did not affect the brain states, as the normalized
occurrences of different movie contents showed no significant
bias (Supplementary Fig. S4D). By splitting the brain states’
occurrence rate data into two halves by the median of days,
we found that the coupling pattern did not appear at the
early training days but became prominent at the later period
(Supplementary Fig. S4C). These results suggest that the brain
states’ temporal pattern may reflect information accumulation
during the movie-watching period.
Next, we examined the temporal pattern of the behaviormost-correlated brain states during the decision period, when
the animals expressed their innate preference. We first reclustered the brain states by including all time points of
delay and decision periods and identified highly similar
patterns of the behavior-most-correlated states as above
(Supplementary Fig. S5, Pearson’s correlation, ranging from
r = 0.70 to r = 0.94 value, P < 0.05, see detailed statistics in
Supplementary Table S4). We found similar temporal dynamics
for the three marmosets by plotting their occurrence rates
around the decision period (Fig. 4A, left panel). The autocorrelation spectrums revealed a prominent time coupling
at 1 s before the decision period onset (Fig. 4A, right panel).
We did not find any similar patterns in the other states
with low correlations, such as the behavior-least-correlated
states (Fig. 4B). The pattern was also robust in the behaviormost-correlated states of other clustering K solutions. Still, it
became weaker when the correlation with behaviors decreased,
suggesting that the pattern does not depend on the choice of
the clustering K (Supplementary Fig. S6). Before the decision, the
coupling pattern also presented a longitudinal training effect,
as it was weaker in the early training days but became stronger
in the later training days (Fig. 4C). Therefore, the longitudinal
change of the behavior-most-correlated brain states reveal
Tian et al.
4229
an enhancement of target identification and facilitate choice
making in favor of the novel target (Fig. 1B,C).
Besides the prominent effect before the choice making, the
coupling patterns also occurred at the time of choice making
(1 s after fixation offset), although they were relatively weak
(Fig. 4A). Repeating the same analysis as for Figure 4C, we found
an opposite effect that was more prominent in early training
days than in later training days (Fig. 4D). The coupling patterns at
this time reflected the motivation of information-seeking driven
by the uncertainty of the selected target. When the animals were
unfamiliar with the task on early training days, the uncertainty
of choice outcomes drives the information seeking, resulting in
a more prominent coupling. On later training days, the animals
were familiar with the task, and the information seeking for
novel targets became goal directed. Thus, the coupling patterns
of the behavior-most-correlated brain states became weaker.
In summary, the temporal patterns of the behavior-mostcorrelated brain states reveal numerous components of curiosity: the information accumulation during the movie-watching
period (Fig. 3 and Supplementary Fig. S4C), the identification of
novel target before fixation offset (Fig. 4A–C), and the motivation
of information seeking for the selected target during the decision period (Fig. 4A,B,D). The longitudinal change of such temporal patterns is also strongly associated with behavior improvement (Fig. 1B,C). Thus, the behavior-most-correlated brain states
indeed reflect an intact, curiosity-driven information-seeking
process.
The Cerebellum is a Central Region in the
Behavior-Most-Correlated Brain States
After characterizing the temporal dynamics of the behaviormost-correlated brain states, we investigated which brain
areas contributed most to it. For each brain connection,
we counted how many other states were significantly different from the behavior-most-correlated state (tested by
repeated-measures one-way ANOVA, see detailed statistics in
Supplementary Table S4), and then summed these numbers
to associate with each brain region. We normalized the
added numbers to a range from 0 to 1 and ranked all
brain regions according to their numbers. Since the ranking results per brain hemisphere (Supplementary Table S3)
were similar to the pooled results of the two hemispheres
(Supplementary Table S2), we presented the pooled results.
Figure 5A shows brain region rankings (right histograms)
and the top 10 ranked regions of each marmoset (left pie
charts), highlighting five overlapped brain regions. These five
brain regions are top ranked and share high similarity in
all marmosets (Fig. 5A,B), including the cerebellum (ranked
first), the hippocampus (second), the peristriate visual cortex
(A19DI, third), the subgenual cingulate cortex (A25, fourth),
and the dorsolateral prefrontal cortex (A46D, fifth). Among
these top-ranked regions, connections involving the cerebellum
contributed significantly to the other brain regions (Fig. 5B,
repeated one-way ANOVA measures, see detailed statistics in
Supplementary Table S4).
In contrast, connections involving other brain regions did
not show the same level of contributions (Fig. 5B). We also
observed similar results in the top 10 brain regions’ contribution matrices across marmosets (Supplementary Fig. S7A,
repeated one-way ANOVA measures, see detailed statistics
in Supplementary Table S4). In other words, the cerebellum is
not only the most prominent region but also contributes most
4230
Cerebral Cortex, 2021, Vol. 31, No. 9
to the other top-ranked regions, indicating its central role in
the behavior-most-correlated states. For the top-ranked brain
regions of each marmoset, since the cerebellum is one of them,
we also found that it has relatively high correlation with other
brain areas (Supplementary Fig. S7B).
The cerebellum is a large brain region consisting of multiple
functionally distinct areas (Fig. 5C). As there was no digital atlas
of the marmoset cerebellum, we defined 48 (24 per hemisphere)
ROIs to examine which parts of the cerebellum contributed the
most to the behavior-most-correlated states (Fig. 5C). Repeating
the same brain state analysis for each ROI, we found that multiple areas of the cerebellum contributed more significantly to
the behavior-most-correlated states than the whole cerebellum
(by repeated-measures one-way ANOVA, see detailed statistics
in Supplementary Table S4). There is no clear lateralization, as
the significant ROIs are distributed closely within the two hemispheres. For example, ROI 20 on the left and ROI 22 on the
right belong to the same area (Crus I/II). Although significant
ROIs were widespread, a few ROIs were more prominent and
consistent across all marmosets, including the Crus I/II (ROIs 20,
22), lobule IX (ROI 10), and lobule IV (ROI 2). These regions may
have unique roles in task learning and innate choice preference.
Discussion
Curiosity is an intrinsic desire that motivates information seeking and maximizes knowledge or opportunities (Loewenstein
1994). However, its accurate psychological concept is still not
well defined, and therefore its underlying brain circuits and
mechanisms remain poorly understood. Here, we successfully
induced the curiosity-related information-seeking behavior
in the research-naïve marmosets. The marmosets showed
an innate preference for novel images over familiar ones
without any influences of extrinsic rewards, reflected by both
saccadic metrics of choices and the viewing time (Fig. 1 and
Supplementary Fig. S3). This preference is consistent with
the description of curiosity as a typical information-seeking
behavior (Berlyne 1978; Baranes et al. 2015; Oudeyer and Smith
2016) that is intrinsically rewarding (Loewenstein 1994; Kang
et al. 2009).
Using longitudinal fMRI and dynamic FC analysis, we further revealed the brain states that were closely correlated with
behavioral performance and potential curiosity functions. During the movie-playing period, the temporal patterns of the brain
states reflected the process of information accumulation, which
is consistent with the description of curiosity functions, the
improvement of knowledge (Kidd and Hayden 2015; Gruber and
Ranganath 2019; Kobayashi and Hsu 2019). Before the decision,
the temporal patterns indicated an enhancement for the novel
target selection, consistent with previous findings that curiosity
needs executive control (Gottlieb et al. 2014; Foley et al. 2017;
Lau et al. 2020). At the time of target selection, the uncertainty
of selected target information decreases with training, reflected
by the weak coupling patterns. Since curiosity improves learning
and optimizes learning experience (Berlyne 1954, 1960, 1978;
Berlyne and Slater 1957; Loewenstein 1994; Kidd and Hayden
2015; Oudeyer and Smith 2016), it is reasonable to consider that
the novel target and its related information seeking became
goal directed when animals were familiar with the task. Such
possibility is also supported by many studies, which might benefit the learner by enhancing the coding and retention of new
information (Kinney and Kagan 1976; Kang et al. 2009; Kidd and
Hayden 2015; Oudeyer and Smith 2016).
By examining the brain state, we identified several key brain
regions supporting the potential cognitive processes underlying
curiosity. The hippocampus has been regarded as an indispensable component of curiosity in previous human and theoretical studies (Kang et al. 2009; Gruber et al. 2014; Gruber
and Ranganath 2019). Curiosity also recruited the dorsolateral
prefrontal cortex (A46D) and the peristriate visual cortex (A19DI),
two central regions for attention (Gottlieb and Oudeyer 2018;
Kobayashi et al. 2019) that ensure efficient information seeking
(Kinney and Kagan 1976). Curiosity is also linked with the emotional state (Loewenstein 1994; Kang et al. 2009). Consistently,
we found that the subgenual cingulate cortex (A25), an area
involving autonomic and emotional control (Joyce and Barbas
2018), contributed significantly to the brain state. Finally, the
most prominent area associated with curiosity-driven behavior
is the cerebellum. By employing ROI-based analysis, we further
revealed the subregions of cerebellum that contribute more to
the behavior-most-correlated states, especially the Crus I/II and
lobules IX and IV. Since there are few studies on the marmoset
cerebellum, we explored the subregions functions based on
comparative evidence from humans and macaques. In these
species, Crus I, Crus II, and the lobule IX correspond to the
first, second, and third nonmotor representative areas of the
cerebellum, respectively (Guell et al. 2018; King et al. 2019), and
anatomically communicate with forebrain structures, including
the areas we found here (Strick 1985; Dum and Strick 2003; Strick
et al. 2009; Bernard et al. 2014; Proville et al. 2014). In addition,
recent studies even revealed that the inactivation of the Crus I
only disrupted the proportion of correct choices without disrupting the ability to make a choice (Deverett et al. 2018), as well as
the loss of prior knowledge (Deverett et al. 2019). Besides these
high cognitive functions, the lobule IV of the anterior cerebellum
is mainly associated with sensorimotor control and learning
(Kelly and Strick 2003). The involvement of these subregions
indicates that multiple nonmotor functions in the cerebellum
are involved in curiosity-related behaviors.
Although we have identified temporal and spatial patterns of
curiosity-related brain states, caution should be exercised when
making direct inferences about how curiosity works. First, similar to the information-seeking neuronal circuit (White et al. 2019;
Cervera et al. 2020), curiosity might also involve the regions for
the information evaluation and the demand control (effector) to
accomplish its functions. The lobule IV of the cerebellum is more
likely related to the effector regions due to its associated sensorimotor functions (Kelly and Strick 2003), whereas the other
subregions of the cerebellum (Crus I, Crus II, and the lobule IX),
the hippocampus, and the cortical regions identified here might
relate to information evaluation because of their involvement
in multiple high-cognitive functions. Second, the improvement
of saccadic metrics suggests that sensorimotor learning may
also be involved in our study, as curiosity optimizes the learning
experience. However, how curiosity influences the sensorimotor
learning is unclear and worthy of future investigations. Finally,
we did not detect several other curiosity-related brain areas
reported in previous studies, including the subcortical dopamine
system, the lateral intraparietal cortex, the orbitofrontal cortex, and the anterior insular regions (Jepma et al. 2012; Gruber
et al. 2014; Blanchard et al. 2015; van Lieshout et al. 2018).
This discrepancy may be explained by the experimental design,
in which we removed the influence of external rewards so
that some brain areas may not be strongly activated. Also, our
FC analysis only detected the most significant brain regions
based on FC. As curiosity enrolls many cognitive processes, our
Brain Circuits of Curiosity-Driven Behavior
analysis might miss some curiosity-related brain areas working
independently.
In conclusion, we induced curiosity-relevant behaviors in
research-naïve marmosets and examined the associated brain
areas by analyzing noninvasive fMRI data and dynamic FC brain
states. The brain states’ temporal dynamics and their related key
regions unveil an important evidence on the role of curiosity.
Although facing several limitations, the current study provides
new insights on the brain regions involved in curiosity-driven
behavior and warrants further investigations combining neuroimaging, electrophysiology, and pharmacological inactivation
approaches to further elucidate the underpinning mechanisms
of curiosity.
Supplementary Material
Supplementary material can be found at Cerebral Cortex online.
Funding
This research was supported, in part, by the Intramural Research
Program of the National Institutes of Health, National Institute
of Neurological Disorders and Stroke (ZIA NS003041); the PA
Department of Health (SAP #4100083102 to A.C.S.); and Shanghai Municipal Science and Technology Major Project (Grant No.
2018SHZDZX05 to C.L.).
Authors’ Contributions
X.T. and C.L. designed the research, performed the experiments,
analyzed the data, and interpreted the results. X.T., C.L., and
A.C.S. wrote and revised the paper. C.L. and A.C.S. supervised
the project.
Notes
The authors thank D. Szczupak and X. (Lisa) Zhang for their
assistance in animal handling; C. C. Yen for his assistance in
MRI scanning; R. Krauzlis and D. Leopold for the experimental
equipment support; Z. Hafed, H. Jiang, L. Wang, and C. Xue for
the scientific discussions and suggestions for draft preparation
and revision; and the National Institutes of Health Fellows Editorial Board for the editorial assistance. Conflict of Interest: None
declared.
References
Allen EA, Damaraju E, Plis SM, Erhardt EB, Eichele T, Calhoun
VD. 2014. Tracking whole-brain connectivity dynamics in the
resting state. Cereb Cortex. 24:663–676.
Anselme P. 2010. The uncertainty processing theory of motivation. Behav Brain Res. 208:291–310.
Baranes A, Oudeyer PY, Gottlieb J. 2015. Eye movements reveal
epistemic curiosity in human observers. Vision Res. 117:81–90.
Barttfeld P, Uhrig L, Sitt JD, Sigman M, Jarraya B, Dehaene S. 2015.
Signature of consciousness in the dynamics of resting-state
brain activity. Proc Natl Acad Sci U S A. 112:887–892.
Berlyne DE. 1954. A theory of human curiosity. Br J Psychol.
45:180–191.
Berlyne DE. 1960. Conflict, arousal, and curiosity. New York:
McGraw-Hill Book Company.
Berlyne DE. 1978. Curiosity and learning. Motiv Emot. 2:97–175.
Tian et al.
4231
Berlyne DE, Slater J. 1957. Perceptual curiosity exploratory
behavior, and maze learning. J Comp Physiol Psychol. 50:
228–232.
Bernard JA, Peltier SJ, Benson BL, Wiggins JL, Jaeggi SM,
Buschkuehl M, Jonides J, Monk CS, Seidler RD. 2014. Dissociable functional networks of the human dentate nucleus. Cereb
Cortex. 24:2151–2159.
Blanchard TC, Hayden BY, Bromberg-Martin ES. 2015.
Orbitofrontal cortex uses distinct codes for different choice
attributes in decisions motivated by curiosity. Neuron.
85:602–614.
Bromberg-Martin ES, Hikosaka O. 2011. Lateral habenula neurons signal errors in the prediction of reward information.
Nat Neurosci. 14:1209–1216.
Cai B, Zhang G, Zhang A, Stephen JM, Wilson TW, Calhoun VD,
Wang Y. 2018. Capturing dynamic connectivity from resting
state fMRI using time-varying graphical Lasso. IEEE Trans
Biomed Eng. 66(7):1852–1862.
Cervera RL, Wang MZ, Hayden BY. 2020. Systems neuroscience
of curiosity. Curr Opin Behav Sci. 35:48–55.
Choe AS, Nebel MB, Barber AD, Cohen JR, Xu Y, Pekar JJ,
Caffo B, Lindquist MA. 2017. Comparing test-retest reliability of dynamic functional connectivity methods. Neuroimage.
158:155–175.
Cox RW. 2012. AFNI: what a long strange trip it’s been. Neuroimage. 62:743–747.
Davis RT, Settlage PH, Harlow HF. 1950. Performance of normal
and brain-operated monkeys on mechanical puzzles with
and without food incentive. J Genet Psychol. 77:305–311.
Deverett B, Kislin M, Tank DW, Wang SS. 2019. Cerebellar disruption impairs working memory during evidence accumulation. Nat Commun. 10:3128.
Deverett B, Koay SA, Oostland M, Wang SS. 2018. Cerebellar
involvement in an evidence-accumulation decision-making
task. Elife. 7:e36781.
Dum RP, Strick PL. 2003. An unfolded map of the cerebellar
dentate nucleus and its projections to the cerebral cortex. J
Neurophysiol. 89:634–639.
Ebitz RB, Albarran E, Moore T. 2018. Exploration disrupts choicepredictive signals and alters dynamics in prefrontal cortex.
Neuron. 97:475.
Foley NC, Kelly SP, Mhatre H, Lopes M, Gottlieb J. 2017. Parietal
neurons encode expected gains in instrumental information.
Proc Natl Acad Sci U S A. 114:E3315–E3323.
Friedman J, Hastie T, Tibshirani R. 2008. Sparse inverse covariance estimation with the graphical Lasso. Biostatistics.
9:432–441.
Funamizu A, Ito M, Doya K, Kanzaki R, Takahashi H. 2012. Uncertainty in action-value estimation affects both action choice
and learning rate of the choice behaviors of rats. Eur J Neurosci.
35:1180–1189.
Gottlieb J, Hayhoe M, Hikosaka O, Rangel A. 2014. Attention, reward, and information seeking. J Neurosci. 34:
15497–15504.
Gottlieb J, Oudeyer PY. 2018. Towards a neuroscience of active
sampling and curiosity. Nat Rev Neurosci. 19:758–770.
Gruber MJ, Gelman BD, Ranganath C. 2014. States of curiosity modulate hippocampus-dependent learning via the
dopaminergic circuit. Neuron. 84:486–496.
Gruber MJ, Ranganath C. 2019. How curiosity enhances
hippocampus-dependent memory: the Prediction, Appraisal,
Curiosity, and Exploration (PACE) framework. Trends Cogn Sci.
23:1014–1025.
4232
Cerebral Cortex, 2021, Vol. 31, No. 9
Guell X, Schmahmann JD, Gabrieli J, Ghosh SS. 2018. Functional
gradients of the cerebellum. Elife. 7:e36652.
Harlow HF. 1950. Learning and satiation of response in intrinsically motivated complex puzzle performance by monkeys. J
Comput Physiol Psychol. 43:289–294.
Harlow HF, Harlow MK, Meyer DR. 1950. Learning motivated by a
manipulation drive. J Exp Psychol. 40:228–234.
Hutchison RM, Womelsdorf T, Allen EA, Bandettini PA, Calhoun VD, Corbetta M, Della Penna S, Duyn JH, Glover GH,
Gonzalez-Castillo J et al. 2013. Dynamic functional connectivity: promise, issues, and interpretations. Neuroimage.
80:360–378.
Jenkinson M, Beckmann CF, van Behrens TE, Wollrich MW, Smith
SM. 2012. FSL. Neuroimage. 62:782–790.
Jepma M, Verdonschot RG, van Steenbergen H, Rombouts SA,
Nieuwenhuis S. 2012. Neural mechanisms underlying the
induction and relief of perceptual curiosity. Front Behav Neurosci. 6:5.
Joyce MKP, Barbas H. 2018. Cortical connections position primate area 25 as a keystone for interoception, emotion, and
memory. J Neurosci. 38:1677–1698.
Kang MJ, Hsu M, Krajbich IM, Loewenstein G, McClure SM, Wang
JT, Camerer CF. 2009. The wick in the candle of learning:
epistemic curiosity activates reward circuitry and enhances
memory. Psychol Sci. 20:963–973.
Kelly RM, Strick PL. 2003. Cerebellar loops with motor cortex
and prefrontal cortex of a nonhuman primate. J Neurosci.
23:8432–8444.
Kidd C, Hayden BY. 2015. The psychology and neuroscience of
curiosity. Neuron. 88:449–460.
King M, Hernandez-Castillo CR, Poldrack RA, Ivry RB, Diedrichsen J. 2019. Functional boundaries in the human cerebellum revealed by a multi-domain task battery. Nat Neurosci.
22:1371–1378.
Kinney DK, Kagan J. 1976. Infant attention to auditory discrepancy. Child Dev. 47:155–164.
Kobayashi K, Hsu M. 2019. Common neural code for reward and
information value. Proc Natl Acad Sci U S A. 116:13061–13066.
Kobayashi K, Ravaioli S, Baranes A, Woodford M, Gottlieb J.
2019. Diverse motives for human curiosity. Nat Hum Behav.
3:587–595.
Kreitler S, Zigler E, Kreitler H. 1984. Curiosity and demographic
factors as determinants of children’s probability-learning
strategies. J Genet Psychol. 145:61–75.
Lau JKL, Ozono H, Kuratomi K, Komiya A, Murayama K. 2020.
Shared striatal activity in decisions to satisfy curiosity
and hunger at the risk of electric shocks. Nat Hum Behav.
4:531–543.
Liu C, Ye FQ, Newman JD, Szczupak D, Tian X, Yen CC, Majka
P, Glen D, Rosa MGP, Leopold DA et al. 2020. A resource for
the detailed 3D mapping of white matter pathways in the
marmoset brain. Nat Neurosci. 23:271–280.
Liu C, Ye FQ, Yen CC, Newman JD, Glen D, Leopold DA, Silva AC.
2018. A digital 3D atlas of the marmoset brain based on multimodal MRI. Neuroimage. 169:106–116.
Liu C, Yen CC, Szczupak D, Ye FQ, Leopold DA, Silva AC. 2019.
Anatomical and functional investigation of the marmoset
default mode network. Nat Commun. 10:1975.
Loewenstein G. 1994. The psychology of curiosity: a review and
reinterpretation. Psychol Bull. 116:75–98.
Oudeyer PY, Smith LB. 2016. How evolution may work through
curiosity-driven developmental process. Top Cogn Sci.
8:492–502.
Preti MG, Bolton TA, Van De Ville D. 2017. The dynamic functional
connectome: state-of-the-art and perspectives. Neuroimage.
160:41–54.
Proville RD, Spolidoro M, Guyon N, Dugue GP, Selimi F, Isope P,
Popa D, Lena C. 2014. Cerebellum involvement in cortical sensorimotor circuits for the control of voluntary movements.
Nat Neurosci. 17:1233–1239.
Redgrave P, Gurney K. 2006. The short-latency dopamine signal: a role in discovering novel actions? Nat Rev Neurosci.
7:967–975.
Reppert TR, Lempert KM, Glimcher PW, Shadmehr R. 2015. Modulation of saccade vigor during value-based decision making.
J Neurosci. 35:15369–15378.
Schwartenbeck P, Passecker J, Hauser TU, FitzGerald TH, Kronbichler M, Friston KJ. 2019. Computational mechanisms of
curiosity and goal-directed exploration. Elife. 8.
Silva AC, Liu JV, Hirano Y, Leoni RF, Merkle H, Mackel JB,
Zhang XF, Nascimento GC, Stefanovic B. 2011. Longitudinal
functional magnetic resonance imaging in animal models.
Methods Mol Biol. 711:281–302.
Strick PL. 1985. The cerebellum: the cerebellum and neural
control. Science. 229:547.
Strick PL, Dum RP, Fiez JA. 2009. Cerebellum and nonmotor
function. Annu Rev Neurosci. 32:413–434.
Tian X, Yoshida M, Hafed ZM. 2016. A microsaccadic account of
attentional capture and inhibition of return in Posner cueing.
Front Syst Neurosci. 10:23.
Tian X, Yoshida M, Hafed ZM. 2018. Dynamics of fixational
eye position and microsaccades during spatial cueing:
the case of express microsaccades. J Neurophysiol. 119:
1962–1980.
van Lieshout LLF, Vandenbroucke ARE, Muller NCJ, Cools R, de
Lange FP. 2018. Induction and relief of curiosity elicit parietal
and frontal activity. J Neurosci. 38:2579–2588.
White JK, Bromberg-Martin ES, Heilbronner SR, Zhang K, Pai J,
Haber SN, Monosov IE. 2019. A neural network for information seeking. Nat Commun. 10:5168.