Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
1
The predictive power of intrinsic timescale during the perceptual decision-
making process across the mouse brain
Elaheh Imani
1, Alireza Hashemi
2, Setayesh Radkani
3, Seth W. Egger4, Morteza Moazami Goudarzi*3
1Department of Computer Engineering, Ferdowsi University of Mashhad, Mashhad, Iran
2 The Decision Lab, Montreal, Canada
3 Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
4 Department of Neurobiology, Duke University School of Medicine, Durham, NC 27710, USA.
*: Corresponding author: moazami@mit.edu
: These authors contributed equally to this work
Abstract
Across the cortical hierarchy, single neurons are characterized by differences in the extent to which they
can sustain their firing rate over time (i.e., their "intrinsic timescale"). Previous studies have demonstrated
that neurons in a given brain region mostly exhibit either short or long intrinsic timescales. In this study,
we sought to identify populations of neurons that accumulate information over different timescales in the
mouse brain and to characterize their functions in the context of a visual discrimination task. Thus, we
separately examined the neural population dynamics of neurons with long or short intrinsic timescales
across different brain regions. More specifically, we looked at the decoding performance of these neural
populations aligned to different task variables (stimulus onset, movement). Taken together, our
population-level findings support the hypothesis that long intrinsic timescale neurons encode abstract
variables related to decision formation.
Furthermore, we investigated whether there was a relationship between how well a single neuron
represents the animal’s choice or stimuli and their intrinsic timescale. We did not observe any significant
relationship between the decoding of these task variables and a single neuron’s intrinsic timescale. In
summary, our findings support the idea that the long intrinsic timescale population of neurons, which
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
2
appear at different levels of the cortical hierarchy, are primarily more involved in representing the
decision variable.
Keywords: intrinsic timescale, decision-making process, multiple timescales of integration
Introduction
Living in natural environments entails performing a variety of tasks, each of which requires processing
information over multiple timescales. For example, organisms need to extract information from visual
scenes that change on the order of milliseconds and then hold and manipulate information in working
memory on the order of minutes, if not hours. The heterogeneity of local microcircuits and their long-
range connectivity with other circuits allow the brain to function on multiple timescales (Chaudhuri,
Knoblauch, Gariel, Kennedy, & Wang, 2015). A key question is whether and how the intrinsic dynamic
properties of neurons are systematically linked to their functional specializations with respect to task
demands that vary in the temporal dimension.
Several studies in humans and non-human primates have demonstrated that different brain regions exhibit
distinct properties in terms of the timescales of information integration. These timescales are organized
hierarchically across the brain, and this hierarchical organization reflects the anatomical hierarchy of
these regions (Cavanagh, Hunt, & Kennerley, 2020; Chen, Hasson, & Honey, 2015; Honey et al., 2012;
Imani et al., 2022; Murray et al., 2014; Pinto, Tank, & Brody, 2022). In macaques, the timescale of
intrinsic fluctuations in spiking activity has been shown to increase across the cortical hierarchy (Murray
et al., 2014). A similar hierarchical organization is shown to exist across the somatosensory network
(Rossi-Pool et al., 2021), and association brain areas (Gao, van den Brink, Pfeffer, & Voytek, 2020) in
macaques, as well as visual areas in the mouse brain (Siegle et al., 2021). Consistent with these animal
studies, studies using electrocorticographic (ECOG) signals in humans have shown that the neural
timescale is consistent with the anatomical hierarchy and increases from sensory to association regions.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
3
In addition to the overall organization of timescales across multiple brain regions that correspond to the
anatomical hierarchy of those regions, there is heterogeneity in the timescale of neuronal dynamics within
each functionally relevant brain region. Recent studies have revealed that the intrinsic timescale of
neurons at rest, within each brain region, is associated with their response selectivity during the task
(Cavanagh et al., 2020; Cavanagh, Wallis, Kennerley, & Hunt, 2016; Fascianelli, Tsujimoto, Marcos, &
Genovesio, 2019; Fontanier, Sarazin, Stoll, Delord, & Procyk, 2021).
Neurons with longer timescales show higher response selectivity for temporally extended computations
such as decision making and reward integration, whereas shorter timescale neurons are implicated in
sensory processing. The variability of intrinsic timescales within the prefrontal cortex (PFC) in non-
human primates predicts the involvement of these neurons in decision making, with longer timescale
neurons exhibiting stronger correlations with the animal’s choice (Cavanagh et al., 2016). Populations of
longer time-scale neurons were also found to represent working memory in ventrolateral PFC (Cavanagh,
Towers, Wallis, Hunt, & Kennerley, 2018). Finally, longer timescale neurons in dorsolateral PFC show
the stronger and more prolonged encoding of decision-related information (Fascianelli et al., 2019).
Altogether, these findings suggest a potential connection between intrinsic temporal properties of neurons
and their functional specialization, both across and within different brain regions.
However, the investigation of the organizational principles of brain computation across the whole brain
has been limited due to the lack of physiological data with high spatiotemporal resolution from multiple
brain regions. Indeed, the organizational hierarchy of neural timescales and their functional relevance has
been studied only in a limited set of brain regions. Moreover, the link between the temporal properties of
neurons and their functional specialization has been studied for a limited set of cognitive processes. For
instance, the majority of existing work on intrinsic timescales did not investigate the temporal properties
of neurons involved in perceptual decision-making.
Here, we address these gaps in the literature by investigating how the intrinsic timescale of baseline
neural activity relates to the functional specialization (i.e., stimulus and choice selectivity in a perceptual
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
4
decision-making task) of both single neurons and populations of neurons across the mouse brain. For this,
we use two large-scale datasets recorded by Neuropixels and widefield Calcium imaging in the same
visual discrimination task (Steinmetz, Zatka-Haas, Carandini, & Harris, 2019; Zatka-Haas, Steinmetz,
Carandini, & Harris, 2021). Using these datasets allows us further to examine the consistency of the two
recording modalities. Our population-level findings support the hypothesis that neurons with longer
intrinsic timescales within and across different brain regions are more strongly implicated in the
representation of the decision and stimulus.
Results
Our analyses were performed on the activity of neurons within and across multiple brain regions of 10
mice performing a visual discrimination task (Steinmetz et al., 2019) (see Methods). Visual stimuli of
varying contrast could appear on the left, right, both, or neither side of the screen on a given trial. Mice
earned a water reward by turning a wheel and indicating which side had the greatest contrast or by not
turning the wheel if no stimulus was presented (Figure 1 - a).
While mice performed this task, Neuropixel probes were used to collect data from approximately 30,000
neurons in 42 brain regions (Figure 1 - b & Table 1, see Methods for selection criteria). We also analyzed
widefield calcium imaging data that measured the activity of the dorsal cortical regions during the same
task but recorded in a separate experiment (Zatka-Haas et al., 2021). Electrophysiological and calcium
imaging data formed the basis for all of our analyses.
The "intrinsic timescale" of a neuron is defined as the decay time constant of the autocorrelation structure
for that neuron. We computed this metric over a 1s baseline period (Murray et al., 2014) (see Methods).
Previously, most studies of the functional significance of intrinsic timescales assigned a single timescale
to an entire brain region. In contrast, we were more interested in the diversity of timescales within
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
5
regions, and whether neurons with different intrinsic timescales (within particular brain regions) served
distinct functional roles.
Figure 1 - Large scale neural recording during the task and the spike count autocorrelations of the neurons within the baseline
period was used to estimate the neural timescale. a) Task setup. Mice earned water rewards by turning a wheel to indicate which
of two visual gratings had higher contrast or by not turning the wheel if no stimulus was presented, adapted from (Steinmetz et
al., 2019). b) Schematic of brain regions. Recordings were made from each of the seven-colored brain regions adapted from
(Steinmetz et al., 2019). c) Autocorrelation functions for a representative neuron from each brain region which was used to
measure intrinsic timescales. d) The first row is the variance explained (R-squared) by the exponential fit to each neuron,
averaged across neurons in each brain region. Error bars indicate confidence interval (CI). The second row is the number of
neurons that survive all pre-selection criteria per region.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
6
This interest stems from the fact that, in cognitive processes such as perceptual decision-making and
memory retrieval, the intrinsic timescale of the neurons has been shown to partially determine their
functional role (Cavanagh et al., 2018; Cavanagh et al., 2016; Fascianelli et al., 2019). However, the
relationship between an individual neuron's encoding properties and its intrinsic timescale within the
brain region is poorly understood and can be broadly assessed in two ways.
Table 1 - Brain regions within each group of areas in the physiological dataset according to the Allen CCF
Group name Regions within each group
Hippocampus POST, SUB, DG, CA1, CA3
Thalamus LP, LD, RT, MD, MG, LGd, VPM, VPL, PO, POL
Visual VISp, VISrl, VISam, VISpm, VISl, VISa
Striatum CP, GPe, ACB, LS
Frontal MOs, ACA, PL, ILA, ORB
MopSSp MOp, SSp
Midbrain MRN, SCm, SCs, APN, PAG, SNr
Table 2 - Brain regions within each group in the calcium imaging dataset according to the Allen CCF
Group name Regions within each group
Visual VISp, VISpm, VISam, VISrl, VISal, VISpl
Frontal MOs, ACA, PL
MopSSp MOp, SSp
First, we can determine how well two pseudo-populations of neurons with different intrinsic timescales
encode a particular task variable, such as the animal's choice or the stimuli it was exposed to (see
Methods). It is worth noting that this distinction does not imply that a single neuron with a short or long
intrinsic timescale encodes the stimulus or decision variable independently, but rather that populations of
neurons with different intrinsic timescales contribute to the representation of a task variable to varying
degrees.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
7
Alternatively, each individual neuron may be responsible for encoding the animal's choice or stimulus on
its own. In such a case, we would expect a strong correlation between the intrinsic timescale of individual
neurons within brain regions and how accurately they represent the stimulus or the animal's choice. This
paper will thoroughly examine both of these possibilities, beginning with the first.
The heterogeneity of intrinsic timescales across brain regions
We restrict all of our subsequent analyses in the physiological dataset to neurons within the seven broad
brain regions defined in Table 1. Rather than pooling the autocorrelograms of all neurons within these
seven regions (which is the most common procedure), we first determined the intrinsic timescale for each
neuron separately (Figure 1 - c). Despite the fact that this resulted in a more noisy fitting procedure
because some neurons were inevitably poorly represented by a simple exponential decay function (Figure
1 - d, top panel), a significant proportion (~15% per area) of neurons showed a decay in autocorrelation
structure that could be reliably quantified using an exponential function and were therefore included in
subsequent analyses (Figure 1 - d, bottom panel, See Methods).
In the widefield calcium imaging data, we also grouped the areas in the dorsal cortex into three regions (
Table 2). To be consistent with the physiological dataset, we restricted our analysis to the left hemisphere
of the calcium imaging data. We then computed the timescale of the pixels by fitting the autocorrelation
of the neural activity by an exponential decay function (see Methods). Similar to the physiological data,
we excluded the poorly fitting pixels from further analysis (Supplemental Figure 4- a).
We observed a high degree of within-region diversity in intrinsic timescales both in the physiological
(Figure 2 - a) and calcium imaging data (Supplemental Figure 4- c). The presence of single neurons with
a diverse range of timescales across different brain regions suggests that this observed heterogeneity has a
functional relevance. This finding prompted us to examine the functional properties of neurons with
different intrinsic timescales within brain regions.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
8
Figure 2 - Heterogeneity of single-neuron timescales within each brain region and the hierarchical ordering of the timescales
across regions. a) Distribution of intrinsic timescales for each brain region. Vertical lines represent the quartiles. b) Intrinsic
timescales averaged across neurons within each brain region and sorted in an ascending manner. Errors indicate CI. c) Pairwise
differences in intrinsic timescales across brain regions. Marker ‘*’ symbolizes significant differences between region’s timescale
in a permutation test corrected for multiple comparisons.
As a first step, we computed the average of intrinsic timescales within each brain region (Figure 2 - b
& c, Supplemental Figure 4- b). We arranged brain regions in ascending order based on their average
intrinsic timescale. We will hereafter refer to this ordering as the “intrinsic timescale hierarchy”, in
short, ITH. As can be observed from these figures, the timescale hierarchy of the regions in the
dorsal cortex is consistent in both datasets. In the physiological data, this hierarchy begins with
hippocampal neurons and ends with midbrain neurons. Moreover, it is loosely correlated with the
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
9
general functions of these regions. For example, brain regions implicated in motor and decision-
making activities had intrinsic timescales that were longer on average than that of visual neurons.
Figure 3 - Population decoding of movement for total, short, and long-timescale neurons. a) Timecourse of population decoding
of movement, aligned to the onset of the wheel turn. Shaded areas indicate CI. b) Maximum population decoding of movement in
different brain regions sorted according to ITH. c) Maximum population decoding of movement in neurons with short and long-
time scales (median split) sorted according to ITH. Across all areas (except for visual), longer timescale neurons (neurons that
fell into the Top 50th percentile) represented information about the animal’s movement with higher accuracy than shorter
timescale neurons. The signs ‘***’ and ‘*’ symbolize p-value < 0.001 and p-value < 0.05 produced by a Wilcoxon rank-sum test
corrected for multiple comparisons.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
10
Relationship between the Neural Intrinsic Timescale and Decision-making
Given that the integration of sensory evidence is a temporally prolonged activity, we hypothesize that
single neurons with longer intrinsic timescales within and across different brain regions are more strongly
implicated in the representation of the animal’s decision. As a first attempt, we addressed this hypothesis
by splitting the neural population within each brain region based on its intrinsic timescale and performing
a population decoding analysis of both the stimulus and the animal's choice for each of these splits.
The direction of wheel movement during correct trials was the first task variable we decoded. Across the
ITH, we observed a broadly improved representation of this measure (Figure 3 - a & b). In other words,
brain regions that had a higher average intrinsic timescale represented the animal’s movement more
accurately than those with a lower average intrinsic timescale (Figure 3 - b).
Furthermore, when we separated neurons within these brain regions based on their intrinsic timescale, we
found that long intrinsic timescale neurons were better at decoding wheel movement, specifically within
brain regions that are heavily implicated in decision-making and learning (Figure 3 - c, e.g., striatum and
frontal).
Relationship between the Neural Intrinsic Timescale and Stimulus Decoding
We also hypothesized whether or not the neural population with a longer timescale within each brain
region can predict the stimulus contrast level of the contralateral stimulus better than shorter ones. One
significant limitation of this decoding measure is that it conflates choice decoding with stimulus
decoding. In other words, a neuron or population of neurons may respond preferentially to the stimulus or
decision variable. However, because there is a strong correlation between the strength of the stimulus and
the animal's choice, it is not easy to distinguish whether the population of neurons more strongly
represents either of these variables. Therefore, we devised a method to separate a neuron or a neural
population’s response to the stimulus from its response to the choice (see Methods). We will evaluate this
metric in subsequent sections.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
11
Accordingly, we divided the trials into 12 groups based on three different animal choices and four
contrast levels. We then compared the activity of these neural populations between trials where the right
stimulus contrast was zero, and the right stimulus contrast was non-zero within each group. The combined
accuracies across 12 groups were used for computing final stimulus decoding. We found that we could
decipher visual information at above-chance levels by decoding the activity of both short and long
intrinsic timescale neurons in these locations (Figure 4 - a).
Furthermore, for the majority of the brain regions we evaluated (5 out of 7), long intrinsic timescale
neurons decoded the stimulus with greater accuracy than short ones (Figure 4 - c). And aside from visual
areas, the midbrain and frontal areas exhibited strong stimulus decoding, which is consistent with their
well-documented roles in perceptual decision-making (Coen, Sit, Wells, Carandini, & Harris, 2021; Scott
et al., 2017; Steinmetz et al., 2019; L. Wang, McAlonan, Goldstein, Gerfen, & Krauzlis, 2020; Zatka-
Haas et al., 2021). Consistent with these findings, our calcium imaging results also support that longer
intrinsic timescale neurons contribute more to stimulus decoding across brain regions (Supplemental
Figure 5- e).
Relationship between the Neural Intrinsic Timescale and Decision (Detect Probability) Decoding
We sought additional evidence to support our hypothesis that long intrinsic timescale neurons encode
abstract variables associated with decision formation. As a result, we used a metric to decode the animal’s
choice independent of stimulus strength. We will refer to this decision-making metric as “population
detect probability” (population DP). As mentioned before, there is a strong correlation between stimulus
and animal choice. So similar to the stimulus decoding, we used the combined condition decoding
approach by splitting the trials into 12 groups according to the different combinations of right and left
stimulus contrast levels. We then separated the trials within each group into Hit (correctly turning the
wheel) and Miss classes.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
12
Figure 4 - Population decoding of stimulus and decision for total, short, and long-timescale neurons. a) Timecourse of stimulus
population decoding. Shaded areas indicate CI. b) Maximum decoding accuracy of the stimulus for different brain regions sorted
according to ITH. c) Median split of the maximum population for stimulus decoding. d) Maximum decoding accuracy of decision
across brain regions. e) Median split of the population for the decision decoding. The signs ‘***’, ‘**’ and ‘*’ symbolize p-value
< 0.001, p-value < 0.01 and p-value < 0.05 in a Wilcoxon rank-sum test corrected for multiple comparisons.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
13
We generally expect higher-level regions to be involved in representing the correct perceptual decision by
reading out noisy information from lower levels of the hierarchy to form a choice. Consistent with this,
we found that decision-making and learning areas had a higher average population DP than sensory areas.
Furthermore, we then divided the population into long and short intrinsic timescales; we discovered that
neurons with longer intrinsic timescales (Top 50th percentile) could predict the animal's decision more
accurately than neurons with shorter intrinsic timescales (Figure 4 - e, Bottom 50th percentile). Consistent
with these findings, our calcium imaging results also lend support to the idea that longer intrinsic
timescale neurons contribute more to decision decoding across brain regions (Supplemental Figure 5- f).
Intrinsic neural timescales and single neuron decoding of the animal’s choice and stimuli
Our initial hypothesis was that neurons with longer intrinsic timescales would more strongly represent the
decision-making process than neurons with shorter intrinsic timescales. This follows from the fact that
sensory evidence integration is a temporally prolonged activity. In the previously reported analyses, we
addressed and largely confirmed this hypothesis by splitting the neural population based on its intrinsic
timescale and performing a population decoding analysis. Here we present a secondary approach; namely,
we also wanted to find out if single neurons with long or short intrinsic timescales represented
information regarding the animal’s decision or the stimuli presented to the animal.
The first step was to determine which subset of neurons more strongly encoded the stimulus and which
neurons more strongly encoded the choice (i.e., the animal's decision to move or not move the wheel).
This was a significant challenge because, during the course of a given task, individual neurons in the
cortex may encode a variety of distinct task variables.
This simultaneous representation of diverse components of a task is referred to as a 'mixed'
representation, and it is difficult to decipher. Kobak, Brendel et al. (2016) developed a data analysis tool
called demixed PCA (dPCA) to address this problem (Kobak et al., 2016). Similar to other dimensionality
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
14
reduction approaches, dPCA breaks down the activity of a population of neurons into a limited number of
components. Unlike other approaches, each dPCA component relates only to a particular task-relevant
variable, making it easy to interpret. For example, the component of the neural population representing
the animal’s choice can be differentiated from the one that carries information about the stimuli it was
exposed (Figure 5 – a & Supplemental Figure 1).
Figure 5 - Demixing neural population into task-related components. a) Each subplot shows a demixed principal component
(corresponding to the stimulus, the decision, and interaction). Each subplot has four lines corresponding to four conditions (see
legend). b) Explained variance of the individual demixed principal components for each brain region. c) Clustering of three
neuron subtypes: stimulus (blue), decision (pink), and other (grey) based on demixed principal components (left). auROC
validation of the three neuron subtypes (right). Shaded areas indicate CI. Stimulus neurons (blue) decode the stimulus much
better than the other neuron subtypes. Decision neurons (pink) encode the decision (or “detect probability) better than the other
neuron subtypes. d) Proportion of different subtypes for this example region (visual).
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
15
We clustered neurons using the stimulus and choice components of dPCA (see Methods) based on
whether they responded more strongly to the stimulus or the choice. Accordingly, the activity of the
neurons was projected to the stimulus and choice components using the decoder matrices of dPCA. We
then reconstructed the neural activity using the choice and stimulus encoder matrices. Using the neurons'
explained variance (R-squared) and fuzzy C-Means algorithm, we indicated whether they responded more
strongly to the stimulus or the choice. We discovered that neurons across brain regions generally fell into
one of three clusters: those best represented by the stimulus component, the choice component, or their
interaction. As a result, we used clustering to determine whether a neuron was stimulus-selective, choice-
selective, or neither (Figure 5 – b & Supplemental Figure 2- a).
The measures we used to evaluate single neuron decoding were based on an auROC metric (Figure 5 – c
& Supplemental Figure 2- b & c). This method is commonly used to calculate the difference between
spike count distributions across trials (see Methods). It can be used to calculate a neuron's sensitivity to
parametrically varying stimuli (stim decoding) or to quantify the relationship between a neuron's activity
and the animal’s choice (detect probability, DP). Due to differences in firing rate associated with each
stimulus strength, DP was independently calculated for each stimulus value.
Previous research has found that DPs typically increase in value as one approaches the site of decision
making and that DP values in early sensory areas are significantly lower than those in sensory areas.
Given that intrinsic timescales vary according to brain region, we investigated whether there was also a
relationship between these auROC-based measures of choice and stimulus selectivity and individual
neuron intrinsic timescales.
We discovered that a significant proportion of the neurons we examined were either stimulus or choice
selective. The proportion of each type of neuron within each brain region varied greatly (Figure 6 - a). We
hypothesized that individual neurons with longer timescales ought to represent decision variables. Put
another way; there should be a positive relationship between neural timescales and single neuron
decoding of the animal's decision (detect probability). However, no statistically significant relationship
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
16
was discovered between either the decision or stimulus decoding properties of individual neurons (within
brain regions) and their intrinsic timescale (Figure 6 - a, middle and right panel, Table 3). Moreover, we
did not observe a significant difference between how short and long intrinsic timescale neurons
represented the stimulus or choice signal at any point after stimulus onset (Supplemental Figure 3).
Furthermore, we discovered that there was no evidence of a stimulus or decision decoding hierarchy
across the brain regions we studied either (Figure 6 - b & c).
Table 3 - Pearson’s correlation between intrinsic timescale and measures of single neuron decoding
Stimulus decoding Detect probability
Brain region correlation p-value correlation p-value
Hippocampus -0.062 0.3 -0.053 0.52
Thalamus 0.095 0.044 -0.11 0.12
Visual 0.15 0.062 0.16 0.061
Striatum -0.03 0.68 0.16 0.18
Frontal 0.088 0.13 -0.085 0.28
MopSSp 0.27 0.0042 0.16 0.25
Midbrain 0.0086 0.92 0.098 0.44
In conclusion, the relationship between neural timescales and their role in decision-making was observed
at the population level but not at the single neuron level. This relationship was especially strong in brain
regions involved in decision-making and learning. Neurons in these brain regions play a critical role in
dynamically tracking value and volatility in the environment (Massi, Donahue, & Lee, 2018).
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
17
Figure 6 - Decoding accuracy of stimulus and choice selective neurons. a) the first column indicates the distribution of timescale
for different types of neurons. The second and third columns demonstrate the scatter plots for individual neuron stimulus
decoding and DP, respectively. b) Maximum stimulus decoding for different brain regions sorted according to ITH. c) Detect
probability for different brain regions according to the ITH ordering. The signs ‘***’ and ‘**’ symbolize p-value < 0.001 and p-
value < 0.01 produced by the Wilcoxon rank-sum test corrected for multiple comparisons.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
18
Discussion
We have demonstrated that intrinsic timescales have high predictive power when it comes to
characterizing the decision-making and stimulus encoding properties of populations of neurons. The brain
regions we examined contained single neurons with a wide range of timescales. Nonetheless, these brain
regions varied in their average intrinsic timescales, giving rise to a hierarchical organization based on
temporal dynamics. However, we did not find strong evidence for a systematic relationship between
individual neurons’ intrinsic timescales and their functional specialization.
Within brain areas, we also identified functional variations between neurons with varying intrinsic
timescales. We found that neurons with more sustained activity (longer intrinsic timescales) were more
likely to be involved in perceptual decision-making at the population level but not at the single neuron
level. Taken together, our findings at the population level lend support to computational theories that
attribute evidence integration to highly recurrent attractor networks (X.-J. Wang, 2002).
There are two potential biophysical mechanisms that could give rise to longer intrinsic timescales
(Murray et al., 2014). One potential mechanism is stronger synaptic connections mediating recurrent
excitement, which slow intrinsic dynamics by partially negating leak (Goldman, Compte, & Wang, 2010).
Another mechanism is an increase in the number and density of excitatory synapses onto pyramidal cells
across the cortical hierarchy (Elston, 2003). Modeling studies have revealed that strong recurrent
connections enable neurons to exhibit persistent activity in working memory (Kim & Sejnowski, 2021;
Zylberberg & Strowbridge, 2017), as well as the ability to accumulate information slowly during
decision-making (X.-J. Wang, 2002).
A hierarchy of timescales may offer a distinct computational advantage for the mammalian nervous
system. Previous studies have shown that when basic recurrent neural networks (RNNs) are task-
optimized, a hierarchy of neural timelines (within brain regions) automatically emerges. In particular,
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
19
RNNs with a hierarchy of timescales have been shown to be effective at learning tasks that are a re-
composition of previously learned sub-goals (Han, Doya, & Tani, 2020).
There is one important caveat with our study: while decoding can reveal how much information a
neuronal population has about a function, high decoding accuracy does not necessarily imply that a brain
area is directly involved in processing that function. Furthermore, while our calcium imaging results lend
support to the idea that longer intrinsic timescale neurons contribute more to sensory processing and
decision-making across brain regions, we did not observe a similar functional hierarchy across brain areas
to what we found in single units. This disparity could be explained by the fact that Neuropixels record
neural activity from multiple layers of the cortex, whereas calcium imaging only records neural activity
from the first layer of the cortex. This discrepancy may possibly be explained by the fact that the number
of neurons included in MOpSSp after preprocessing was much smaller than the number of calcium
imaging channels.
To summarize, in this paper, we identified a functionally relevant heterogeneity in intrinsic timescales,
which enables the brain to perform tasks that necessitate longer-term representations of choice and
stimulus information. We found that this result holds true across brain regions and modalities. We made
several novel contributions to the existing body of work. We devised a method to decompose the
contribution of intrinsic timescales to decision variables from their representation of the stimulus. The
results presented in this publication are more reliable than prior research on intrinsic timescales, which
did not always adequately separate stimulus and choice contributions to population brain activity.
Materials and Methods
Behavioral Task
We analyzed a dataset made freely available by (Steinmetz et al., 2019). This dataset contains behavioral
and physiological data collected from ten mice over the course of 39 sessions on a visual discrimination
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
20
task (two-alternative unforced choices). Mice rested on a plastic apparatus with their forepaws on a
rotating wheel, surrounded by three computer monitors. To begin the trial, the animal briefly held the
wheel. Then, using 16 various conditions, visual stimuli (Gabor patch with sigma 9 and 45° direction)
with four grading levels might be exhibited on the right, left, both, or neither screen.
The animal did not need to move its head to perceive the stimuli because they were presented in the
mouse's central monocular zones. Mice were rewarded by moving the wheel such that the stimulus with
the highest contrast moved to the center of the screen or by not turning the wheel at all if no discernable
stimuli were shown. Otherwise, they received a white noise signal to indicate an improper wheel
movement. So, three types of outcomes (turn right, turn left, no turn) led to the reward according to the
stimulus presentation.
After the stimulus presentation, there was a random delay interval of 0.5–1.2 s during which the mouse
could freely turn the wheel, but visual stimuli were locked in place, and incentives were unavailable. At
the conclusion of the delay interval, an auditory tone cue (8 kHz pure tone for 0.2 s) was supplied, at
which point the visual stimulus position became coupled to wheel movement.
Electrophysiological data
Recordings
Neuropixel electrode arrays were utilized to record approximately 30,000 neurons in 42 in the left
hemisphere brain areas during the task. Due to the capacity of a single probe to record from numerous
brain regions and the usage of many probes concurrently, each session produced data recorded from many
areas simultaneously. We divided the regions into seven main groups according to the Allen Mouse Brain
Common Coordinate Framework (CCF) atlas (Q. Wang et al., 2020) (Figure 1 - b). All the analyses were
performed on these groups of regions.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
21
Data preprocessing
We excluded cells from our decoding analysis based on the neuron's timescale goodness of fit. Spikes
were binned at 0.005s and smoothed using a half-gaussian kernel with a standard deviation of 0.02s.
Across all trials and time points, data from all neurons were z-scored by subtracting the mean and
dividing by the standard deviation calculated during the baseline period (-0.9s to -0.1s, stimulus aligned).
All analyses were carried out using customized MATLAB code. Statistical analyses, SVM classification,
and fuzzy C-means clustering were performed using MATLAB toolboxes. We also used the open-source
dPCA toolbox (Kobak et al., 2016) to decompose neural activity into different task-related variables.
Intrinsic timescale
Our measure of intrinsic timescale was based on the spike count autocorrelation function proposed by
(Murray et al., 2014). The spike count of each neuron was first binned at
0.05
using non-
overlapping windows during a 1s baseline period over Go trials. Next, the across trial Pearson’s
correlation was computed between each pair of time bins
and
(
) separated by time lag
.
The autocorrelation values follow an exponential decay from lower to the higher time lags. So, the
intrinsic timescales were estimated by fitting an exponential decay function to the autocorrelation values
of each neuron, as follows:
(1)
Where
is the time lag,
is the intrinsic timescale, A is the amplitude, and B denotes the contribution
of timescales longer than the observation window. Some neurons had low correlation values at shorter
time lags which may reflect negative adaptation (Murray et al., 2014). Fitting began with the lag with the
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
22
greatest autocorrelation reduction to overcome this feature. Using five different initial parameter values,
the Levenberg-Marquardt method was used to fit equation (1) to the autocorrelation values. The model
with the lowest mean square value (MSE) was chosen as the best model for describing the decay of a
particular neuron's autocorrelation function.
The temporal autocorrelation of some neurons could not be described by an exponential decay. Therefore,
we automatically excluded neurons that can not satisfy the following criterias: (1): minimum average
firing rate of 1HZ during 1s baseline period across Go trials, (2): minimum R-squared value of 0.4 for the
exponential decay fitting, (3):
0.01 1
. Accordingly, about half of the neurons satisfied these
requirements and were further analyzed by the visual inspection process. During this phase, neurons were
visually inspected and eliminated if their temporal autocorrelation function was not well-captured by an
exponential decay function. A significant proportion of neurons (~15% per region) passed both the
automated and visual inspection procedure and were thus selected for further analyses.
Population Decoding
We used a binary support vector machine (SVM) classifier with a linear kernel function to decode task
variables from neural activity. We measured how neural activity predicted the direction of wheel
movement, whether or not the animal turned the wheel, and the stimulus that the animal was exposed to
using this classifier.
Decoding the direction of wheel movement
In order to decode the direction of wheel movement, neurons within each brain region were combined to
form a pseudo-population by randomly selecting correct trials for each direction (without replacement,
sample size = 10). Sessions with less than ten correct trials for each direction were excluded from
subsequent decoding analyses. We measured the accuracy of the classifier using a 3-fold cross-validation
procedure over 50 iterations of trial sub-sampling at each time point. We used the activity of neurons in a
window ranging from -0.3s to 0.1s (aligned with wheel movement) for this analysis.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
23
Decoding the contralateral stimulus
By design, the encoding of the stimulus is highly correlated with the animal's choice. As a result, the
decoding of wheel movement direction reflects an interaction between both the stimulus and choice
variables. To address this limitation, we performed our decoding analyses within different choices and
stimulus conditions and combined their accuracies. Because the data was recorded from the left
hemisphere, we only decoded the stimulus on the right screen. The trials for contralateral stimulus
decoding were divided into 12 groups based on the animal's choice alternatives (right, left, NoGo) and
four left stimulus contrast levels (100, 50, 25, 0) (Zatka-Haas et al., 2021). We used the activity of
neurons in a window ranging from -0.1s to 0.3s (aligned with stimulus onset) for this analysis.
Following that, the trials within each group were divided into two categories: (1) trials with a right
stimulus contrast level of zero; (2) trials with a right stimulus contrast level greater than zero. We then
generated a pseudo-population for each condition by selecting ten trials from each class at random across
sessions. If a session did not meet the trial cutoff criteria for a specific condition, it was removed from the
pseudo-population for that condition. We used a 3-fold cross-validation procedure to assess the
performance of 12 classifiers at each time point. The final accuracy for each time point was then
calculated by averaging the accuracies of these 12 classifiers. To compute confidence intervals, this
sampling procedure was repeated 50 times.
Detect probability
We next measured how the neural activity encodes whether or not the animal turned the wheel correctly
and referred to this measurement as ‘Detect probability’ (Hashemi, Golzar, Smith, & Cook, 2018). We
similarly divided the trials into 12 groups according to the different combinations of right and left
stimulus contrast levels (0, 25, 50, 100), ignoring equal contrast pairs. The trials within each group were
separated into Hit (correctly turning the wheel) and Miss classes. We then sub-sampled the trials within
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
24
each class randomly without replacement with sample size five to prepare 12 pseudo-populations across
sessions. The sessions with a trial number lower than five within each condition were excluded from the
condition’s pseudo-population. We then measured the performance of the decoder by combining the
accuracy of 12 classifiers. Classifier accuracy was assessed using a 3-fold cross-Validation procedure
applied to each time point during the stimulus epoch (-0.1s to 0.3s stimulus alignment). The sampling
process was repeated 50 times to measure the classifier’s confidence interval.
We also measured how the neural activity encodes task variables as a function of the neuron’s intrinsic
timescale. Accordingly, we split the neurons within each brain region into short and long timescales
according to their median timescale values. For each group of neurons, we repeated the above procedures
for measuring the performance of each decoder.
Single neuron decoding
The single neuron decoding analysis was performed using the area under the receiver operating
characteristic (auROC) analysis (Britten, Newsome, Shadlen, Celebrini, & Movshon, 1996). This method
is commonly used to calculate the difference between spike count distributions across two conditions. As
mentioned before, the encoding of the stimulus and the animal’s choice are highly correlated. Therefore,
similar to the population decoding, we utilized the combined condition auROC analysis to resolve this
limitation for decoding the stimulus and detect probability during the stimulus epoch (-0.1s to 0.3s,
aligned to stimulus onset).
We only decoded the contralateral stimulus by dividing the trials into 12 groups according to the three
animal choice alternatives and four left stimulus contrast levels. We then measured the neuron’s stimulus
selectivity by applying the auROC analysis per condition. The final neuron’s stimulus selectivity was
measured by the weighted average across 12 auROC values per time point.
We also assessed how a single neuron could encode whether or not the animal’s turned the wheel (Detect
probability). Similar to the population decoding, we first divided the trials into 12 groups based on
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
25
different combinations of right and left stimulus levels ignoring the equal contrast pairs. We then
evaluated the neuron’s detect probability by taking the weighted average of auROC analysis across
conditions.
Task-related neural data decomposition
Most neurons encode different types of task information and exhibit a mixed selectivity. The complexity
of single-neuron responses can conceal the type of information expressed by them and how it is
represented. To circumvent this constraint, we used demixed principal component analysis (dPCA)
(Kobak et al., 2016). Briefly, dPCA decomposes the population neural activity into a few latent
components, with each capturing a specific aspect of the task. The compressed subspace derived from
dPCA captures the bulk of the variation in the data while also decoupling different task-related
components, such that each component captures variance primarily associated with a single task variable.
We first prepared a matrix containing the marginalized population's neural activity for different stimulus
and choice conditions to construct the latent subspace. Accordingly, we divided the trials into four groups
based on the contralateral contrast levels. Within each contrast level, we computed the average firing rate
of neurons based on whether or not the animal turned the wheel. The neural population matrix
ேൈସൈଶൈ்
contains the activity of N neurons over four different stimulus contrast levels and two alternative animal
wheel movement statuses with T time points during the stimulus epoch (-0.1s to 0.3s). We then applied
dPCA to the neural population to construct a latent subspace containing 20 components characterizing the
stimulus (
௦௧
), decision (
ௗ௧
), stimulus-decision interaction (
௦ௗ௧
), and condition-independent
information (
௧
) by minimizing the following loss function (Kobak et al., 2016):
ௗ
ఈ
ఈ
ఈ
ଶ
ఈ
, ! " #$, $, %$, %$&
(2)
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
26
Where F and D refer to the decoder and encoder for each task variable
!
. Using the estimated decoder
and encoder for each task variable, we can compute the explained variance (R-squared) of the neurons
using the following equations:
1
4
1
2
1
)
௦ௗ௧
்
௧ୀଵ
ଶ
ௗୀଵ
ସ
௦ୀଵ
ఈ
ଶ
1 ∑∑ ∑
ఈ
ఈ
௦ௗ௧
௦ௗ௧
ଶ்
௧ୀଵ
ଶ
ௗୀଵ
ସ
௦ୀଵ
∑∑∑
+
௦ௗ௧
,
ଶ
்
௧
ୀଵ
ଶ
ௗୀଵ
ସ
௦ୀଵ
(3)
Where
ఈ
ଶ
is the explained variance of each task variable
!"#$, $, %$, %$&
. The neurons within each
brain area were then clustered (number of clusters = 3) based on their stimulus and choice-related R-
squared values (
௦௧
ଶ
,
ௗ௧
ଶ
), using the fuzzy C-means clustering algorithm (Bezdek, 2013).
Widefield calcium imaging data
Image acquisition and preprocessing
We analyzed the calcium image data made public by (Zatka-Haas et al., 2021) with the same task
protocol and animal subjects as the electrophysiological dataset. Details on the data acquisition and
preprocessing steps are described in (Zatka-Haas et al., 2021). The widefield calcium images were
acquired through a fluorescence macroscope (Scimedia THT-FLSP) containing an sCMOS camera (PCO
Edge 5.5) with a frame rate of 70HZ and pixel size 21.7µm. The authors mapped each mouse's average
cortical fluorescence activity to the Allen Common Coordinate Framework (CCF) using the grid centered
on bregma. They also applied different preprocessing steps for denoising and normalizing the images.
We created a grid box at the CCF space covering the whole dorsal cortex with a grid size of 10µm. We
then mapped the mouse’s image to this grid box by spatially binning pixels falling within each grid.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
27
Across all trials and time points, the activity of pixels was z-scored by subtracting the mean and dividing
by the standard deviation calculated during the baseline period (-1s to 0s, stimulus aligned).
Timescale analysis
We measured the intrinsic timescale using the autocorrelation structure of fluorescence activity during the
baseline period. The activity of each pixel was linearly interpolated at 20 time-bins separated by 0.05s
time lag during a 1s baseline period over Go trials. Next, the across trial Pearson’s correlation was
computed between each pair of time bins
and
(
) separated by time lag (
-
). The
autocorrelation values follow an exponential decay from lower to higher time lags. So, similar to the
physiological data, the intrinsic timescales were estimated by fitting an exponential decay function to the
autocorrelation values of each pixel.
The temporal autocorrelation of some pixels could not be described by an exponential decay. Therefore,
we automatically excluded pixels having a timescale lower than 0.01s or higher than 1s. The remaining
pixels were further analyzed by the visual inspection process. During this phase, pixels were visually
inspected and eliminated if their temporal autocorrelation function was not well-captured by an
exponential decay function. A significant proportion of pixels (~50%) passed both the automated and
visual inspection procedure and were thus selected for further analyses.
Decoding analysis
Similar to the single neuron decoding analysis, we used combined condition auROC analysis to decode
the stimulus and detect the probability of each pixel during the stimulus epoch (-0.1s to 0.3s, aligned to
stimulus onset). Consistent with the previous analysis, we only considered pixels at the left hemisphere.
We performed the auROC analysis per session, and final decoding accuracy was computed by taking the
average of the auROC values of each pixel and time point across sessions
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
28
Supplemental Figures
Supplemental Figure 1 - Demixed principal component analysis projections on task-related components. The figure shows the
projection of population activity on stimulus (top row), decision (middle row), and stimulus-decision interaction (bottom row)
components averaged over different components.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
29
Supplemental Figure 2 - Separating neurons according to their response selectivity. a) Clustering of three neuron subtypes:
stimulus (blue), decision (pink), and other (gray) based on demixed principal components. Panels b and c show the auROC
validation of the three neuron subtypes. Stimulus neurons (blue) decode the stimulus much better than the other neuron subtypes.
Decision neurons (pink) decode the decision (or detect probability) better than the other neuron subtypes. Shaded areas
demonstrate CI.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
30
Supplemental Figure 3 - auROC timecourses for stimulus (panel a) and detect probability (panel b) for the short and long-
timescale neurons. The curves represent the average of auROC values across neurons, and the shaded areas indicate the CI.
Supplemental Figure 4 - Timescale of the widefield calcium imaging data. a) Intrinsic timescale heatmap for the left hemisphere
excluding pixels with poor timescale fits. b) The hierarchy of average timescales over pixels within each group area of the left
hemisphere. c) Distribution of timescale within each brain region.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
31
Supplemental Figure 5 -
S
timulus decoding and detect probability on the fluorescence data. a,c) Heatmap of contra stimu
lus
d
ecoding accuracy and detect probability. b,d) Time course of the average decoding accuracy across pixels having va
lid
timescale fitting for contra s
t
imulus and detect probability. Shaded areas show CI. e) Max accuracy of contra stimulus for to
tal
valid pixels (left column) and median split of pixels based on the intrinsic timescale (right column). f) Max accuracy of det
ect
probability for total valid pixels (left column) and median split of pixels based on the intrinsic timescale (right column).
T
he si
gn
‘***’ symbolizes p-value < 0.001 in the Wilcoxon rank-sum test corrected for multiple comparisons.
lus
lid
tal
ect
gn
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
32
References
Bezdek, J. C. (2013). Pattern recognition with fuzzy objective function algorithms: Springer
Science & Business Media.
Britten, K. H., Newsome, W. T., Shadlen, M. N., Celebrini, S., & Movshon, J. A. (1996). A
relationship between behavioral choice and the visual responses of neurons in macaque
MT. Vis Neurosci, 13(1), 87-100. doi:10.1017/s095252380000715x
Cavanagh, S. E., Hunt, L. T., & Kennerley, S. W. (2020). A diversity of intrinsic timescales
underlie neural computations. Frontiers in Neural Circuits, 14, 81.
doi:10.3389/fncir.2020.615626
Cavanagh, S. E., Towers, J. P., Wallis, J. D., Hunt, L. T., & Kennerley, S. W. (2018).
Reconciling persistent and dynamic hypotheses of working memory coding in prefrontal
cortex. Nature communications, 9(1), 1-16. doi:10.1038/s41467-018-05873-3
Cavanagh, S. E., Wallis, J. D., Kennerley, S. W., & Hunt, L. T. (2016). Autocorrelation structure
at rest predicts value correlates of single neurons during reward-guided choice. Elife, 5,
e18937. doi:10.7554/eLife.18937
Chaudhuri, R., Knoblauch, K., Gariel, M.-A., Kennedy, H., & Wang, X.-J. (2015). A Large-Scale
Circuit Mechanism for Hierarchical Dynamical Processing in the Primate Cortex. Neuron,
88(2), 419-431. doi:10.1016/j.neuron.2015.09.008
Chen, J., Hasson, U., & Honey, C. J. (2015). Processing timescales as an organizing principle
for primate cortex. Neuron, 88(2), 244-246. doi:10.1016/j.neuron.2015.10.010
Coen, P., Sit, T. P., Wells, M. J., Carandini, M., & Harris, K. D. (2021). Mouse frontal cortex
mediates additive multisensory decisions. bioRxiv.
Elston, G. N. (2003). Cortex, cognition and the cell: new insights into the pyramidal neuron and
prefrontal function. Cereb Cortex, 13(11), 1124-1138. doi:10.1093/cercor/bhg093
Fascianelli, V., Tsujimoto, S., Marcos, E., & Genovesio, A. (2019). Autocorrelation structure in
the macaque dorsolateral, but not orbital or polar, prefrontal cortex predicts response-
coding strength in a visually cued strategy task. Cerebral cortex, 29(1), 230-241.
doi:10.1093/cercor/bhx321
Fontanier, V., Sarazin, M., Stoll, F., Delord, B., & Procyk, E. (2021). Local inhibitory control of
frontal network metastability underlies the temporal signature of cognitive states.
bioRxiv.
Gao, R., van den Brink, R. L., Pfeffer, T., & Voytek, B. (2020). Neuronal timescales are
functionally dynamic and shaped by cortical microarchitecture. Elife, 9, e61277.
doi:10.7554/eLife.61277
Goldman, M. S., Compte, A., & Wang, X.-J. (2010). Neural integrator models. Encyclopedia of
neuroscience, 165-178. doi:B978-008045046-9.01434-0
Han, D., Doya, K., & Tani, J. (2020). Self-organization of action hierarchy and compositionality
by reinforcement learning with recurrent neural networks. Neural Networks, 129, 149-
162. doi:10.1016/j.neunet.2020.06.002
Hashemi, A., Golzar, A., Smith, J. E. T., & Cook, E. P. (2018). The Magnitude, But Not the Sign,
of MT Single-Trial Spike-Time Correlations Predicts Motion Detection Performance. The
Journal of Neuroscience, 38(18), 4399-4417. doi:10.1523/jneurosci.1182-17.2018
Honey, C. J., Thesen, T., Donner, T. H., Silbert, L. J., Carlson, C. E., Devinsky, O., . . . Hasson,
U. (2012). Slow cortical dynamics and the accumulation of information over long
timescales. Neuron, 76(2), 423-434. doi:10.1016/j.neuron.2012.08.011
Imani, E., Radkani, S., Hashemi, A., Harati, A., Pourreza, H., & Goudarzi, M. M. (2022).
Distributed coding of evidence accumulation across the mouse brain using microcircuits
with a diversity of timescales. bioRxiv.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
33
Kim, R., & Sejnowski, T. J. (2021). Strong inhibitory signaling underlies stable temporal
dynamics and working memory in spiking neural networks. Nature neuroscience, 24(1),
129-139. doi:10.1038/s41593-020-00753-w
Kobak, D., Brendel, W., Constantinidis, C., Feierstein, C. E., Kepecs, A., Mainen, Z. F., . . .
Machens, C. K. (2016). Demixed principal component analysis of neural population data.
Elife, 5, e10989. doi:10.7554/eLife.10989
Massi, B., Donahue, C. H., & Lee, D. (2018). Volatility facilitates value updating in the prefrontal
cortex. Neuron, 99(3), 598-608. e594. doi:10.1016/j.neuron.2018.06.033
Murray, J. D., Bernacchia, A., Freedman, D. J., Romo, R., Wallis, J. D., Cai, X., . . . Lee, D.
(2014). A hierarchy of intrinsic timescales across primate cortex. Nature neuroscience,
17(12), 1661-1663. doi:10.1038/nn.3862
Pinto, L., Tank, D. W., & Brody, C. D. (2022). Multiple timescales of sensory-evidence
accumulation across the dorsal cortex. Elife, 11, e70263. doi:10.7554/eLife.70263
Rossi-Pool, R., Zainos, A., Alvarez, M., Parra, S., Zizumbo, J., & Romo, R. (2021). Invariant
timescale hierarchy across the cortical somatosensory network. Proceedings of the
National Academy of Sciences, 118(3). doi:10.1073/pnas.2021843118
Scott, B. B., Constantinople, C. M., Akrami, A., Hanks, T. D., Brody, C. D., & Tank, D. W.
(2017). Fronto-parietal cortical circuits encode accumulated evidence with a diversity of
timescales. Neuron, 95(2), 385-398. e385. doi:10.1016/j.neuron.2017.06.013
Siegle, J. H., Jia, X., Durand, S., Gale, S., Bennett, C., Graddis, N., . . . Luviano, J. A. (2021).
Survey of spiking in the mouse visual system reveals functional hierarchy. Nature,
592(7852), 86-92. doi:10.1038/s41586-020-03171-x
Steinmetz, N. A., Zatka-Haas, P., Carandini, M., & Harris, K. D. (2019). Distributed coding of
choice, action and engagement across the mouse brain. Nature, 576(7786), 266-273.
doi:10.1038/s41586-019-1787-x
Wang, L., McAlonan, K., Goldstein, S., Gerfen, C. R., & Krauzlis, R. J. (2020). A causal role for
mouse superior colliculus in visual perceptual decision-making. Journal of Neuroscience,
40(19), 3768-3782. doi:10.1523/JNEUROSCI.2642-19.2020
Wang, Q., Ding, S.-L., Li, Y., Royall, J., Feng, D., Lesnar, P., . . . Ng, L. (2020). The Allen
Mouse Brain Common Coordinate Framework: A 3D Reference Atlas. Cell, 181(4), 936-
953.e920. doi:10.1016/j.cell.2020.04.007
Wang, X.-J. (2002). Probabilistic decision making by slow reverberation in cortical circuits.
Neuron, 36(5), 955-968. doi:10.1016/s0896-6273(02)01092-9
Zatka-Haas, P., Steinmetz, N. A., Carandini, M., & Harris, K. D. (2021). Sensory coding and the
causal impact of mouse cortex in a visual decision. Elife, 10, e63163.
doi:10.7554/eLife.63163
Zylberberg, J., & Strowbridge, B. W. (2017). Mechanisms of Persistent Activity in Cortical
Circuits: Possible Neural Substrates for Working Memory. Annu Rev Neurosci, 40, 603-
627. doi:10.1146/annurev-neuro-070815-014006
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
34
Data availability
All neural and behavioral data analyzed in this study are available at https://figshare.com/articles/
steinmetz/9598406.
Author contribution
M.M.G. conceptualization. E.I., A.H, and M.M.G. performed the analyses. E.I., A.H., and M.M.G.
contributed to the methodology. E.I., A.H, and MMG carried out visualization. E.I., A.H., S.R. S.W.E.,
and M.M.G interpreted the results. E.I., A.H., S.R., and M.M.G. wrote the paper. S.R. and S.W.E review
and editing, M.M.G. supervised the project.
Competing interests
The authors declare no competing interests.
.CC-BY-NC-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted January 3, 2023. ; https://doi.org/10.1101/2023.01.01.522410doi: bioRxiv preprint
