For many decades, neuroscientists have studied task-dependent response properties of individual neurons, and this has laid the groundwork for our current understanding of brain function. More recently, a major shift from looking at individual neurons to studying population responses has begun to shed light on larger-scale neural dynamics, thus providing insight into previously hidden circuit mechanisms (Cunningham et al., 2014; Yuste, 2015). Usually, in this approach, the major axes of variation of a neural population are defined and the population activity is projected into this neural state space. This new way of looking at neural firing rates has begun to revolutionize our understanding of various neural processes, including the evolution of choice in parietal and frontal cortices (Harvey et al., 2012; Mante et al., 2013; Morcos and Harvey, 2016; Raposo et al., 2014), the critical stages of movement preparation and reaching in premotor and motor cortices (Churchland et al., 2012; Kaufman et al., 2014; Li et al., 2016), and the mechanisms underlying working memory in prefrontal cortex (PFC) (Murray et al., 2017).

Yet deriving neural population trajectories requires invasive recordings because the population vector is constructed from the firing rate of individual neurons. As a consequence, to date, no comparable non-invasive techniques for use in healthy human participants have been established. Here we asked whether we can recover ‘population-like’ trajectories in humans and thus provide insight into the unfolding of choice on a millisecond basis, by selectively suppressing neural populations to different features using repetition suppression in magnetoencephalography (MEG) recordings.

Repetition suppression (or ‘adaptation’) takes advantage of a feature first observed over 50 years ago (Gross et al., 1967), showing that the activity of a neuron will be suppressed when it is repeatedly exposed to features it is sensitive to. This phenomenon has since been widely replicated and shown to be a robust property of neurons in single-unit recordings (for review, see Barron et al., 2016a). Importantly, the bulk signal measured from thousands or millions of neurons using techniques such as electroencephalography (EEG), MEG, or functional magnetic resonance imaging (fMRI) will also show suppression if a subset of these neurons is sensitive to a repeated feature. Thus, repetition suppression provides insight into the activity of specific subpopulations of neurons in these non-invasive recordings available for use in humans and, in that way, allows us to examine the underlying neural mechanisms.

In this study, we focus on choice processes unfolding in dorsal premotor cortex (PMd). PMd is the key region for choosing hand digit responses (Dum and Strick, 2002; Rushworth et al., 2003), the response modality in our choice task. Thus, the neural representations of interest are located within one brain region. Given this spatial focus, repetition suppression provides the best resolution achievable using non-invasive MEG: a single sensor or voxel is sufficient to reveal feature-processing using repetition suppression. By contrast, multivariate approaches rely on spatial patterns detected across sensors, which would not offer the required spatial scale.

Here, we extend the repetition suppression framework in one crucial way: we suppress the MEG signal to multiple different features within the same experiment. Adaptation along each feature can be conceptualised as ‘squashing’ the neural response along one task dimension, or task-axis. This assimilates task-axes in multi-dimensional state space derived from recording many neurons, but using an experimental manipulation of a univariate, rather than multivariate, signal. Thus, we ask whether repetition suppression along multiple features can mimic projections onto multiple task-axes. If so, this would be the closest we can get to measuring multiple cellular-level representations within a single brain region in humans using MEG, with temporal resolution in the order of milliseconds, thanks to the temporal precision of MEG.

Our first key result shows that it is possible, using repetition suppression, to simultaneously induce adaptation along multiple task-axes in non-invasive human recordings. The recovered human choice traces show a progression from a processing of choice inputs to a processing of the motor response, just like in invasive recordings from non-human primates (NHPs) (Mante et al., 2013).

Once the feasibility of our non-invasive adaptation approach was established, we asked whether the mechanisms of choice uncovered in NHPs also generalize to humans. More precisely, we examined whether the mechanisms for input selection during choice were comparable between the two species. A large body of evidence in human and non-human primates shows that selection of relevant sensory inputs occurs through top-down modulations from prefrontal and parietal regions onto early sensory regions (Buschman and Miller, 2007; de Lange et al., 2010; Desimone and Duncan, 1995; Egner and Hirsch, 2005; Everling et al., 2002; Friston, 2005; Friston and Kiebel, 2009; Gazzaley et al., 2007; Kok et al., 2016; Michalareas et al., 2016; Moore and Zirnsak, 2017; Noudoost et al., 2010; Pezzulo and Cisek, 2016; Rao and Ballard, 1999; Richter et al., 2017; Squire et al., 2013; van Wassenhove et al., 2005; Wyart et al., 2015). However, two recent perceptual decision-making studies in macaques found that irrelevant sensory inputs are not filtered out before the integration stage (Mante et al., 2013; Siegel et al., 2015). To reconcile these findings, naive human participants performed context-dependent choices in the same perceptual decision-making task used in macaques. This showed that information about both relevant and irrelevant input dimensions was present in premotor cortex, but irrelevant inputs were weaker than relevant inputs, consistent with top-down suppression.

One major difference in terms of how human and animal studies are conducted, however, is that animals are trained for thousands of trials before neural recordings are performed. Differences in input selection mechanisms could therefore be a consequence of brain circuit reorganization caused by extensive over-training. In an attempt to reconcile our findings with those obtained in macaques, we next mirrored NHP training conditions, and our human participants underwent extensive training on over 20,000 trials of the task. However, this did not change the input selection mechanisms evident in the MEG recordings taken afterwards. Irrelevant inputs were present, but still more weakly than relevant inputs. This was true when examining information processed in premotor cortex, or when decoding from whole-brain activity.

Participants (n = 22) performed a modified version of a dynamic random-dot-motion (RDM) decision task, in which stimuli simultaneously contained information about colour (red vs. green) and motion (left vs. right, as in Mante et al., 2013). An additional flanker stimulus displayed 150 ms before the RDM stimulus instructed participants about the relevant stimulus dimension: arrows indicated the choice in the current trial was about the dominant motion direction, whereas coloured dots instructed participants to respond based on the dominant colour (Figure 1A).

Figure 1

Download asset Open asset

MEG activity can track the temporal evolution of a choice process

When recording activity from premotor cortex in NHPs performing the same task, neural activity shows a clear evolution from an initial processing of the choice input (e.g., colour in a colour trial) to a later processing of the choice direction (motor response: right or left; Mante et al., 2013). Our first analysis, therefore, aimed to establish whether the time-resolved nature of MEG data would allow us, in a similar way, to watch how a choice evolved from the processing of inputs to the processing of the motor response in human premotor cortex.

In order to get a handle on neurons that process colour, motion, or response in MEG recordings measuring the summed activity of many neurons, we incorporated an additional task manipulation. The main RDM stimulus on each trial (‘test stimulus’: TS) was preceded by another RDM stimulus, the ‘adaptation stimulus’ (AS), so that each trial contained two RDM stimuli in quick succession (Figure 1A). Unlike the TS, the AS contained only one input feature (strong red or strong green colour, or strong leftward or strong rightward motion).

Participants responded to both stimuli using the corresponding finger of the right hand (index finger for leftward motion or green colour, middle finger for rightward motion or red colour). We hypothesized that presentation of the AS meant that MEG activity recorded at the time of the TS would have suppressed responses to features already engaged at the time of the AS. For example, when a red and leftward TS was preceded by a red AS, MEG activity measured at the time of the TS would still contain the activity of neurons responding to leftward motion, but neurons responding to the red colour input would be suppressed, compared to a situation where the AS was green (Figure 1B). Furthermore, if participants were asked to attend to the colour of the TS, they would produce a middle finger response twice in quick succession (to AS and TS, both red), and thus the ‘response-direction (middle finger)’ selective neurons would also be suppressed (Figure 1C). This would not be the case for the identical stimuli when leftward motion was attended at the time of the test stimulus. In this case, only colour-sensitive neurons but not response-selective neurons would be selectively suppressed at the time of the TS. In summary, by creating and comparing situations with and without suppression for input (colour or motion) and response, we aimed to establish whether we could measure premotor cortex activity transition from a processing of input to a processing of choice output using non-invasive human MEG.

We focused this analysis on the time-course of activity from an area in left premotor cortex (PMd). PMd was identified by first projecting the MEG scalp signal back into source space, which relies on linearly combining MEG sensors to create ‘virtual sensors’ reflecting the signal from specific cortical locations (linearly constrained minimum variance (LCMV)-beamforming; see 'Materials and methods'). At the source level, we then computed a contrast between TS trials with repetition suppression to any input (colour or motion) or response at any time point between [−250,750] ms around the test stimulus and TS trials containing no such suppression (see 'Materials and methods'). This comparison made use of all trials in the experiment and matched the two sides of the contrast in terms of the visual stimulus properties of the TS and the required motor response. The only difference between the two sides of the contrast was which AS preceded a given TS and whether, as a result, we expected repetition suppression to a shared feature or not. Consequently, this contrast was agnostic to any differences between input and response suppression, any differences between relevant and irrelevant input suppression, and any timing differences between these effects. PMd, the premotor region responsible for selecting hand motor responses, was our a priori region of interest for this study. Its role in selecting finger responses is equivalent to the role of the frontal eye fields (FEF) for guiding eye movement responses, the region where NHP recordings were performed (Mante et al., 2013). Indeed, left PMd, in a cluster together with left M1, was the strongest peak at the whole-brain level for this contrast (p(FWE)=0.018 cluster-level corrected; left M1: z = 6.10, peak MNI coordinate x=-36, y=-18, z = 48; left PMd: z = 5.02, peak MNI coordinate x=-37, y=-6, z = 55; right PMd: z = 3.11 at x = 37, y = 06, z = 55; Figure 2A). A further analysis performed on a 38-region parcellation using beamforming with orthogonalization (Colclough et al., 2016; Colclough et al., 2015) confirmed that the linear regression coefficients for both input and response suppression were strongest in the parcel containing premotor cortex (Figure 2—figure supplement 1).

Figure 2 with 3 supplements see all

Download asset Open asset

Figure 2—source data 1

Contains 'pre' and 'post' [Time x Regressors x Subjects] for Figure 2B.

https://cdn.elifesciences.org/articles/60988/elife-60988-fig2-data1-v2.mat

Download elife-60988-fig2-data1-v2.mat

Figure 2—source data 2

Contains 'dat' [Regressors x Subjects] for Figure 2B.

https://cdn.elifesciences.org/articles/60988/elife-60988-fig2-data2-v2.mat

Download elife-60988-fig2-data2-v2.mat

To examine how the choice evolved over time in PMd, we extracted the timeseries from left PMd (x=-37, y=-6, z = 55) and used an L2-regularized linear regression (ridge regression) containing regressors that described which task variables were being suppressed on a trial-by-trial basis (see 'Materials and methods'). The regression was applied to each time point around the presentation of the TS ([−500,1200] ms) using a sliding-window approach (window size: 150 ms). Simulations showed that all regressors were estimable despite dependencies between some (maximum shared variance: r² = 0.34; Figure 2—figure supplements 2 and 3). Figure 2B demonstrates the temporal evolution of the choice for each of the regressors. The two contextual variables ‘context’ (motion or colour instruction flankers) and ‘switch’ (attended dimension same or different from AS) showed the earliest significant representation, and this response was sustained throughout the TS presentation (both first significant at 8 ms). This is unsurprising as contextual information was available −150 ms before the TS. Both of these variables were unrelated to the repetition suppression manipulation. Slightly later, starting from 108 ms, the input representation emerged (probing whether or not the same input feature was already shown in the AS and thus suppressed). The regressor that was represented latest in PMd was the motor response (probing whether the same finger was/was not used to respond to the AS and thus suppressed) and the choice direction (left or right; both starting from 275 ms). Input and response regressors infer sensory and motor processing using the repetition suppression manipulation; choice direction, by contrast, simply captures the TS response direction, independent of any repetition.

Statistical tests performed on the peaks of the linear regression coefficients obtained for the five regressors showed a significant difference in processing latency (one-way repeated-measures ANOVA: F(4, 84)=128.71, p=5.5154e-35, η²_p = 0.86; Figure 2C), and a pairwise post-hoc test between the two critical regressors, input and response repetition suppression, confirmed a significant timing difference with the choice input representations preceding the response representations (pairwise t-test on peak linear regression coefficients for input [429.5 ± 35 ms; mean ± s.t.d.] and response [531.1 ± 41.6 ms]: t(21)=11.07, p=3.1654e-9, after Bonferroni familywise error correction; Hedges’ g = 2.577, 95% CI = [1.8647, 3.4166]; see 'Materials and methods'; Figure 2C). Thus, activity recorded using MEG repetition suppression in human premotor cortex can track the evolution of a choice from sensory input to motor response.

This transition from input-to-choice processing can also be plotted as a continuous trace that progresses along the two task-axes. This is achieved by plotting the linear regression coefficients capturing input suppression, as a marker of input processing, on the first axis and the linear regression coefficients capturing response suppression, as a marker of motor response processing, on the second axis (Figure 3A). The obtained choice trace mimics the state space trajectories derived from neuronal population responses in previous NHP recordings (Mante et al., 2013). This way of plotting the data does not add information but helps visualize the choice dynamics in the way usually done when projecting neural populations into state space. A key difference, however, is that our approach did not rely on invasively recorded firing rates but was possible using non-invasive repetition suppression along multiple features.

Figure 3

Download asset Open asset

Figure 3—source data 1

Contains 'pre.relirrel' and 'post.relirrel' [Time x Regressors x Subjects].

https://cdn.elifesciences.org/articles/60988/elife-60988-fig3-data1-v2.mat

Download elife-60988-fig3-data1-v2.mat

Figure 3—source data 2

Contains structs of 'pre' and 'post'.

These structs have four regressors [Time x 1].

https://cdn.elifesciences.org/articles/60988/elife-60988-fig3-data2-v2.mat

Download elife-60988-fig3-data2-v2.mat

Over-training does not abolish the filtering out of irrelevant inputs

We speculated that one potential cause for seeing attenuated irrelevant sensory input processing in our human participants who spontaneously performed the task, but unattenuated irrelevant input processing in recordings from NHPs, may be rooted in the fact that monkeys have been trained to perform this task for many months and thousands of trials. This may have caused plasticity in the neural pathways that process and pass forward the input information. To test this hypothesis, we trained our human participants for 4 weeks on over 20,000 trials of the task (see 'Materials and methods'; Figure 4—figure supplement 1A). Unsurprisingly, this led to a speeding up of reaction times and higher performance scores (2×2×3 repeated-measures ANOVA with factors context [motion vs. colour], training [pre vs. post], and coherence [70% (AS) vs. 25.6% vs. 12.8%]; effect of training on log-reaction times (RTs): F(1,20)=9.23, p=6.46e-3, η²_p = 0.32, mean RT difference 29 ± 10 ms; effect of training on % correct: F(1,20)=36.39, p=6.75e-6, η²_p = 0.65, mean difference 4.76 ± 0.8; Figure 4A and Figure 4—figure supplement 1B,C). Upon completion of the training, we again recorded MEG data during performance of the task in the same way as done in pre-training. As before, the unfolding of the choice computation was evident in signals recorded from PMd (Figure 4C–D). Crucially, however, when we repeated the analysis focussing on any differences between the processing of relevant and irrelevant inputs, the difference was retained (Figure 4E–G). Even after having performed thousands of trials on the task, irrelevant inputs were, therefore, still filtered out from PMd activity. There was no interaction between input processing and training (two-way repeated-measures ANOVA with factors input [relevant vs. irrelevant] and training [pre vs. post]: all p>0.05 for effect of input x training; Bonferroni correction for familywise error rate; a two-way repeated-measures Bayesian ANOVA provided further evidence in favour of the null hypothesis: the input x training interaction was not part of the winning model at any time point).

Figure 4 with 3 supplements see all

Download asset Open asset

Figure 4—source data 1

Contains 'dat' [Regressors x Subjects].

https://cdn.elifesciences.org/articles/60988/elife-60988-fig4-data1-v2.mat

Download elife-60988-fig4-data1-v2.mat

Consistent with faster behavioural reaction times, we did observe changes in peak processing latency of the different components of the choice process after completion of an extensive training regime (Figure 4D; two-way repeated-measures ANOVA with factors regressors [Context, Switch, Input repetition suppression, Response repetition suppression, Choice] and training [pre vs. post]: effect of regressors [F(4,84) = 39.35, p=1.5824e-18, η²_p = 0.65]; effect of training [F(1,84)=75.81, p=2.0638e-08, η²_p = 0.78]; effect of regressors x training [F(4,84)=35.38, p=2.6436e-17, η²_p = 0.63]). Post-hoc tests confirmed that peak timings changed for four out of five regressors following the training, including the processing of context switches, input adaptation, response adaptation, and choice direction (pairwise t-test between pre vs. post on the peak timings of context: t(21) = 0.18, p=1.00, Hedges’ g = 0.046, 95% CI = [−0.481, 0.575]; switch: t(21) = 5.92, p=3.5641e-05, Hedges’ g = 1.98, 95% CI = [−0.481, 0.575]; input repetition suppression: t(21) = 5.91, p=3.6398e-05, Hedges’ g = 1.57, 95% CI = [0.94, 2.27]; response repetition suppression: t(21) = 12.15, p=2.8914e-10, Hedges’ g = 3.08, 95% CI = [2.28, 4.02]; choice: t(21) = 11.09, p=1.5246e-09, Hedges’ g = 3.28, 95% CI = [2.35, 4.37]; Bonferroni correction for familywise error rate; Figure 4D).

Behavioural effects predicted by neural adaptation mechanisms

We have shown that repeated exposure to sensory inputs or motor responses has an impact on the MEG signal recorded at the time of the TS. However, if repetition has such a clear effect on neural representations, it is possible that it might also impact participants’ behaviour. While the precise neuronal mechanisms underlying repetition suppression are not fully understood, one possibility is that a fatigue mechanism attenuates the inputs of the repeated feature (Barron et al., 2016a; Grill-Spector et al., 2006; Vidyasagar et al., 2010). This would affect the neural information content that is being processed and might translate into behavioural change.

To test this, we repeated the regression conducted at every time step in PMd on log-reaction times (RTs) and response accuracies (% correct) of TS choices (see 'Materials and methods'). We hypothesized that suppression of a relevant sensory feature or motor representation would slow RTs and reduce accuracy. For example, if a green AS meant that green inputs were attenuated at the time of the TS, we would predict TS reaction times and accuracy to be impacted when green is the relevant TS feature to attend to.

Consistent with our predictions, participants were slower and less accurate when the relevant input had already been processed at the time of the AS and was therefore suppressed at the time of the TS (RT: t(21)=6.80, p=1.00e-6, Hedges’ g = 2.00, 95% CI = [1.18, 2.92]; accuracy: t(21)=-8.65, p=2.28e-8, Hedges’ g = −2.55, 95% CI = [−3.60,–1.62]). Similarly, they were less accurate when the motor response was repeated (accuracy; t(21)=-2.28, p=0.03, Hedges’ g = −0.67, 95% CI = [−1.33,–0.05]; Figure 5; see Figure 5—figure supplement 1 for separated results of pre- and post-training), akin to choice history biases that tend to prefer alternating choices (Urai et al., 2019). However, neither RTs nor choice accuracies were affected by repetition of the irrelevant sensory feature (RT: t(21)=0.13, p=0.90, Hedges’ g = 0.04, 95% CI = [−0.57, 0.64]; accuracy: t(21)=0.03, p=0.98, Hedges’ g = 0.01, 95% CI = [−0.60, 0.61]). Other main effects included context (RT: t(21)=-2.31, p=0.03, Hedges’ g = −0.68, 95% CI = [−1.34,–0.06]; accuracy: t(21)=2.48, p=0.02, Hedges’ g = 0.73, 95% CI = [0.10, 1.39]) and context-switch (RT: t(21)=7.72, p=1.46e-7, Hedges’ g = 2.27, 95% CI = [1.40, 3.25]; accuracy: t(21)=-9.58, p=4.10e-9, Hedges’ g = −2.82, 95% CI = [−3.95,–1.82]).

Figure 5 with 1 supplement see all

Download asset Open asset

Filtering out of irrelevant inputs is present across the brain

Finally, it is possible that our targeted focus on PMd as the ‘output’ stage of the decision process might have obscured the fact that other MEG sensors processed irrelevant input information more strongly. To investigate this, we performed a multivariate decoding analysis on the data from 38 parcels covering the whole brain in source space. This analysis focused on stimulus properties of the TS, disregarding any repetition suppression. It also accounted for shared variance between relevant sensory inputs and choice direction by working on the residuals of the MEG signal after regressing out variance related to choice direction (see 'Materials and methods' for details). Separate decoders were constructed to decode relevant and irrelevant sensory input strength (coded as [−2,–1, 1, 2]) from these residuals. Consistent with earlier analyses, decoding across all parcels showed weaker decoding of irrelevant compared to relevant inputs (Figure 6). There was some significant decoding of irrelevant inputs (p<0.001 from 455 to 708.3 ms post-TS) but it was again less extended in time and weaker compared to the decoding of relevant inputs (p<0.001 from 201.7 to 961.7 ms post-TS; Figure 6A). There was again no difference between the data acquired pre- vs. post-training (two-way repeated-measures ANOVA with factors input [relevant vs. irrelevant] and training [pre vs. post]: all p>0.5; minimum p=0.52, F(1,21) = 5.49, η²_p = 0.21, for effect of training; Bonferroni correction for familywise error rate; a two-way repeated-measures Bayesian ANOVA provided further evidence in favour of the null hypothesis: the input x training interaction was not part of the winning model at any time point) apart from differences in peak timing (Figure 6B; pairwise t-test between pre vs. post on the peak timings of input repetition suppression: t(21) = 3.12, p=0.026, Hedges’ g = 1.00, 95% CI = [0.33, 1.73]; context: t(21) = 2.35, p=0.14, Hedges’ g = 0.62, 95% CI = [0.08, 1.20]; switch: t(21) = −0.60, p=1.00, Hedges’ g = −0.19, 95% CI = [−0.86, 0.46]; response repetition suppression: t(21) = 1.59, p=0.63, Hedges’ g = 0.52, 95% CI = [−0.15, 1.22]; choice: t(21) = 2.30, p=0.16, Hedges’ g = 0.67, 95% CI = [0.07, 1.30]), as observed in PMd (Figure 4D). This further supports an interpretation of increased efficiency as a result of extensive over-training. Thus, while this analysis cannot rule out that there may be a brain region that represents irrelevant inputs more strongly, we confirmed that even after extensive training and when considering the activity across the whole brain, the processing of irrelevant inputs was attenuated compared to the processing of relevant inputs when making a choice.

Figure 6

Download asset Open asset

Figure 6—source data 1

Contains structs of 'pre' and 'post'.

These structs have two structs 'relv' and 'irrel', having two matrices of 'motion' and 'colour' [Subjects x Time].

https://cdn.elifesciences.org/articles/60988/elife-60988-fig6-data1-v2.mat

Download elife-60988-fig6-data1-v2.mat

Figure 6—source data 2

Contains 'pre' and 'post' [1 x Subjects].

https://cdn.elifesciences.org/articles/60988/elife-60988-fig6-data2-v2.mat

Download elife-60988-fig6-data2-v2.mat

Here we have shown that non-invasively recorded MEG activity can be used to track a choice process on a millisecond timescale. As decisions unfolded, MEG activity in premotor cortex transitioned from processing sensory inputs to processing the motor response. Watching these dynamics has previously only been possible by projecting high-dimensional neural population recordings onto a low-dimensional set of axes or ‘neural state space’. Here, we took a different approach to assimilate this state space, in which we selectively suppressed the processing of sensory inputs or response features in the average neural data recorded with MEG. This allowed us to selectively index (or ‘squash’) neural activity along the different task-axes that would define the neural state space in direct neuronal recordings. We found that in human premotor cortex, sensory inputs irrelevant to the current choice were processed more weakly compared to relevant choice inputs. This partial filtering of irrelevant choice features was observed even after extensive over-training and when considering activity present across the brain.

Studying neural population responses, as opposed to the responses of individual neurons, has received increasing attention because it provides a window into larger-scale neural dynamics (Yuste, 2015). It has provided crucial insights, for example, into the evolution of choice (Harvey et al., 2012; Hunt et al., 2018; Mante et al., 2013; Morcos and Harvey, 2016; Raposo et al., 2014), movement preparation and execution (Churchland et al., 2012; Kaufman et al., 2014; Li et al., 2016), and the mechanisms underlying working memory (Murray et al., 2017). The approach we used to derive representations linked to multiple features, or task-axes, draws on the observation that electrophysiological responses are suppressed when repeatedly exposed to features to which they are sensitive. In MEG data recorded non-invasively from human participants, each sensor’s signal is driven by the summed postsynaptic potentials of millions of neurons. This bulk signal can be manipulated by repeated exposure to a specific feature, causing suppression in the subset of neurons sensitive to this feature and thus reduce the average MEG signal that is measured (Figure 1). This repetition approach has greater spatial specificity than multivariate approaches because it manipulates activity in subsets of neurons and can thus be measured within a single sensor, rather than relying on variations in the patterns of activity observed across sensors. While this insight has been exploited for understanding a wide range of cognitive processes, with both fMRI (Barron et al., 2013; Barron et al., 2016b; Chong et al., 2008; Doeller et al., 2010; Garvert et al., 2017; Garvert et al., 2015; Jenkins et al., 2008; Klein-Flügge et al., 2019; Klein-Flügge et al., 2013; Piazza et al., 2007) and MEG/EEG (Fritsche et al., 2020; Henson, 2016; Henson et al., 2004; Todorovic and de Lange, 2012), here we extended this framework in one unusual way. As in previous work, using MEG allowed us to measure temporally resolved repetition suppression effects, thus giving insight into precisely when particular features (e.g., input versus choice direction) are processed. This has been exploited previously, for example, to characterize the temporal tuning of repetition suppression (Fritsche et al., 2020) or to temporally dissociate stimulus repetition from stimulus expectation effects (Todorovic and de Lange, 2012). However, importantly, here we suppressed the signal to multiple different features within the same experiment (colour and motion – when relevant or irrelevant – and motor direction). This can be conceptualized as squashing the neural response onto multiple task-axes, with each feature representing one axis, and allowed us to visualize choice traces in a two-dimensional task space.

The obtained trajectories are comparable to population traces from invasive recordings in NHPs (Figures 3 and 4; Mante et al., 2013) in that they show a progression from a processing of choice inputs to a processing of the motor response. However, in Mante et al., 2013, the influence of input returned to baseline before choice execution. This may be because monkeys were only allowed to indicate a response after a delay, once the motion stimulus had been turned off. By contrast, we allowed participants to respond at any time while the stimulus was presented, and thus the trial could end before input processing returned to zero. Indeed, in our human choice trajectories, input representations started to return to baseline just before the time of response but did not fully return to zero by the time participants responded. Overall, we believe our approach provides an exciting new opportunity, allowing researchers to measure something like the state space trajectories obtained from direct recordings but in tasks that might be difficult for NHPs to do (e.g., involving language), or in human disease.

The neural mechanisms underlying repetition suppression are not fully understood to date. Hypothesized mechanisms include neuronal sharpening, fatigue, and facilitation, with current evidence favouring the fatigue model according to which suppression is caused by attenuated synaptic inputs (Barron et al., 2016b; Grill-Spector et al., 2006). Another possibility for the observed signal modulations might be related to change detection or prediction-error-like processes caused by novelty or unexpectedness. However, we observe adaptation of not just sensory inputs but also the motor response, and of not just attended but also unattended sensory inputs (Larsson and Smith, 2012). This lends support to an interpretation as a suppression rather than a prediction-error-like effect. The temporal dynamics of repetition suppression (e.g., the influence of time-lag) can vary between regions and is likely determined by the neural dynamics and recurrent processing of a given region. Importantly, however, the key effects reported here all come from within the same region. This eliminates inter-regional differences in neural dynamics as a possible explanation for the timing differences we observed between sensory and motor suppression effects. Ultimately, the arbitrary nature of the orientation of the MEG dipoles fitted means that we cannot be certain about whether feature repetition causes a suppression or enhancement in the underlying neural signal. For that reason, we report absolute linear regression coefficients throughout. However, single-unit studies provide strong and consistent evidence for a suppression, rather than enhancement, of neural responses after repeated exposure to a sensory feature (Barron et al., 2016b; Grill-Spector et al., 2006).

Our behavioural results are consistent with the interpretation that inputs are suppressed as a result of repeated exposure to a feature. If repeated exposure attenuates the received neuronal inputs, thus affecting the processed information content, this might translate into behavioural change because suppressed information cannot be processed as efficiently for making a choice. However, this should only be the case if the relevant sensory dimension is repeated. Our behavioural analyses confirmed this prediction (Figure 5 and Figure 5—figure supplement 1). Repetition of the relevant sensory feature or response, but not the irrelevant sensory feature, reduced choice accuracies, and repetition of relevant, but not irrelevant, sensory features slowed RTs. These results provide further evidence that the repetition suppression approach employed here was effective at manipulating the inputs to the decision process. It seems likely that the same adaptation process that changes the MEG signal in PMd is causing the changes in behaviour we identified. In prior work, choice biases were shown to depend on the precise choice history, and in the majority of participants, choices were biased away from the previous response (Urai et al., 2019). This finding could relate to effects observed here, showing improved performance for response alternation and reduced performance for response repetition.

Our data speak to an important controversy about the mechanisms underlying feature-based, or context-dependent, choice selection. A substantial body of evidence both in human and non-human primates has shown that selection of relevant sensory features occurs through top-down modulations from prefrontal and parietal regions onto early sensory regions (Buschman and Miller, 2007; de Lange et al., 2010; Desimone and Duncan, 1995; Everling et al., 2002; Friston, 2005; Friston and Kiebel, 2009; Kok et al., 2016; Michalareas et al., 2016; Moore and Zirnsak, 2017; Noudoost et al., 2010; Pezzulo and Cisek, 2016; Rao and Ballard, 1999; Richter et al., 2017; Squire et al., 2013; Summerfield et al., 2014; van Wassenhove et al., 2005; Wyart et al., 2015). By contrast, the study by Mante et al., 2013 found that irrelevant sensory features were not filtered out but passed forward to the output stage (see also Siegel et al., 2015). The authors provided a compelling neural network model that solved feature selection and evidence integration within a single recurrent network. However, the monkeys in Mante et al., 2013 were extensively over-trained at performing the task and this might have caused the circuits to reorganize. We hypothesized that top-down influences might be more important when someone is naive at doing the task. Indeed, in a task with only relevant features, training changes neuronal responses responsible for interpreting sensory evidence, but not those processing the sensory evidence itself (Law and Gold, 2010; Law and Gold, 2008). We, therefore, compared participants before and after >20,000 trials of training on the task. To our surprise, the only difference identified between naive and extensively over-trained participants was a shift in peak timings (Figures 4 and 5). We did not observe any changes in feature selection. In other words, even following the training, irrelevant features were present but weaker than relevant features in premotor cortex and across the brain. Our data thus suggest partial but incomplete filtering of irrelevant inputs in PMd. This is broadly in line with the effects present in participant’s behaviour, where we observed a weak but observable influence of the irrelevant feature (Figure 5A and Figure 5—figure supplement 1A). Consistent with predictive coding principles and theories of top-down control, the processing of irrelevant inputs was diminished, but partly consistent with Mante et al., 2013, they were not completely filtered out. Thus, it seems plausible that some but not all of the feature selection happens at the output stage, in PMd. Overall, the strength of top-down control, or the extent to which task-irrelevant information was filtered out, seemed unaffected by the amount of prior experience on the task.

One important difference between studies in humans and NHPs is the brain area under study. We focused on the premotor area responsible for selecting digit responses, the response modality in our task. By contrast, a large number of studies in NHPs, including Mante et al., 2013, record from FEF because choices are indicated using saccades. In terms of their anatomical connections, FEF and PMd can be considered at similar levels of the processing hierarchy for saccadic and digit-motor responses, respectively. FEF has direct projections to the saccade-initiating superior colliculus, while PMd directly projects to the region of M1 that controls hand motor responses. However, based on pyramidal neuron spine count, a good indicator of the hierarchical position of a cortical area, FEF may be at a lower level of the hierarchy than PMd (Chaudhuri et al., 2015). Whether or not the positions of PMd and FEF in the processing hierarchy are comparable, this potential difference is unlikely to explain the discrepant findings. Other work in NHPs has shown attentional filtering in regions as ‘low’ in the processing hierarchy as V4/MT (Noudoost et al., 2010 and references therein), which would imply sensory information is filtered out before FEF/PMd, consistent with our work.

This leaves outstanding the question as to why equally strong processing of relevant and irrelevant features was observed in NHP state space trajectories (Mante et al., 2013), but not in our data. It has been proposed that the site of feature selection may depend on the level of detail afforded by the prediction (de Lange et al., 2018; Hochstein and Ahissar, 2002). One difference between tasks was that the relevant dimension changed from trial to trial in our experiment but was blocked in Mante et al., 2013. There is evidence that FEF shows long-term selection history effects (Bichot et al., 1996; Bichot and Schall, 1999), which may be promoted by blocking trials. However, Siegel et al., 2015 do not mention any blocking of trials and report results consistent with Mante et al., 2013, making this an unlikely possibility. Recent work has shown attenuation of expected stimulus features when they are attended (Richter and de Lange, 2019). However, attenuation of expected information is usually thought to help filter out predictable objects (Duncan et al., 1997; Everling et al., 2002), for example, via representations of pre-stimulus sensory templates (Kok et al., 2017). Indeed, generally, processing is biased in favour of behaviourally relevant input (for review, see Desimone and Duncan, 1995; Stokes and Duncan, 2014). There is also a discrepancy in terms of the methods used here and in Mante et al., 2013. Mante et al. used multivariate approaches across cells, while we used a repetition suppression manipulation on non-invasive univariate data. However, again, this difference is unlikely to fully account for the discrepancy in findings. A large body of work in NHPs is consistent with our findings but used similar recording, analysis, and training approaches as in Mante et al., 2013 (Noudoost et al., 2010). Furthermore, when we implemented a comparable analysis approach to Mante et al., 2013, albeit using the univariate signal of a single sensor and ignoring the repetition suppression manipulation our design was optimized for (Figure 4—figure supplement 3), we were able to confirm that signal variation related to sensory stimulus strength was weaker for irrelevant compared to relevant features. Ultimately, the discrepancy between different findings remains to be addressed and highlights a general need for a better understanding of decision-making in environments that require dynamic changes (Glaze et al., 2015; Gold and Stocker, 2017; Ossmy et al., 2013).

Overall, our results reinforce the importance of inter-species translational research, whereby tasks and techniques are used across species, e.g., by using comparable analysis pipelines (Hunt et al., 2015; Kriegeskorte et al., 2008), by obtaining direct recordings from humans when possible (e.g., neurosurgical patients: Ekstrom et al., 2003; Fell et al., 2001; Miller et al., 2013; Rutishauser et al., 2010; Watrous et al., 2018), or by recording whole-brain fMRI in NHP species (Bongioanni et al., 2021; Chau et al., 2015; Fouragnan et al., 2019). Our work also emphasizes the importance of developing more mechanistic approaches in human neuroscience, and it shows that the generalizability from NHPs to humans can and should be tested but not assumed (Passingham, 2009).

The code used in the current study and the datasets generated and/or analyzed during the current study are available at the OSF repository (https://doi.org/10.17605/OSF.IO/RJY3Z).

1. 1. Takagi Y
   2. Hunt LT
   3. Woolrich MW
   4. Behrens TEJ
   5. Klein-Flügge MC

   (2021) Open Science Framework

   Adapting non-invasive human recordings along multiple task-axes shows unfolding of spontaneous and over-trained choice.

   https://doi.org/10.17605/OSF.IO/RJY3Z

1. 1. Barron HC
   2. Dolan RJ
   3. Behrens TE

   (2013) Online evaluation of novel choices by simultaneous representation of multiple memories

   Nature Neuroscience 16:1492–1498.

   https://doi.org/10.1038/nn.3515

   • PubMed
   • Google Scholar
2. 1. Barron HC
   2. Garvert MM
   3. Behrens TE

   (2016a) Repetition suppression: a means to index neural representations using BOLD?

   Philosophical transactions of the Royal Society of London. Series B, Biological Sciences 371:20150355.

   https://doi.org/10.1098/rstb.2015.0355

   • PubMed
   • Google Scholar
3. 1. Barron HC
   2. Vogels TP
   3. Emir UE
   4. Makin TR
   5. O'Shea J
   6. Clare S
   7. Jbabdi S
   8. Dolan RJ
   9. Behrens TE

   (2016b) Unmasking latent inhibitory connections in human cortex to reveal dormant cortical memories

   Neuron 90:191–203.

   https://doi.org/10.1016/j.neuron.2016.02.031

   • PubMed
   • Google Scholar
4. 1. Bichot NP
   2. Schall JD
   3. Thompson KG

   (1996) Visual feature selectivity in frontal eye fields induced by experience in mature macaques

   Nature 381:697–699.

   https://doi.org/10.1038/381697a0

   • PubMed
   • Google Scholar
5. 1. Bichot NP
   2. Schall JD

   (1999) Effects of similarity and history on neural mechanisms of visual selection

   Nature Neuroscience 2:549–554.

   https://doi.org/10.1038/9205

   • PubMed
   • Google Scholar
6. 1. Bongioanni A
   2. Folloni D
   3. Verhagen L
   4. Sallet J
   5. Klein-Flügge MC
   6. Rushworth MFS

   (2021) Activation and disruption of a neural mechanism for novel choice in monkeys

   Nature 591:270–274.

   https://doi.org/10.1038/s41586-020-03115-5

   • PubMed
   • Google Scholar
7. 1. Brainard DH

   (1997) The psychophysics toolbox

   Spatial Vision 10:433–436.

   https://doi.org/10.1163/156856897X00357

   • PubMed
   • Google Scholar
8. 1. Brookes MJ
   2. Stevenson CM
   3. Barnes GR
   4. Hillebrand A
   5. Simpson MI
   6. Francis ST
   7. Morris PG

   (2007) Beamformer reconstruction of correlated sources using a modified source model

   NeuroImage 34:1454–1465.

   https://doi.org/10.1016/j.neuroimage.2006.11.012

   • PubMed
   • Google Scholar
9. 1. Buschman TJ
   2. Miller EK

   (2007) Top-down versus bottom-up control of attention in the prefrontal and posterior parietal cortices

   Science 315:1860–1862.

   https://doi.org/10.1126/science.1138071

   • PubMed
   • Google Scholar
10. 1. Chau BK
    2. Sallet J
    3. Papageorgiou GK
    4. Noonan MP
    5. Bell AH
    6. Walton ME
    7. Rushworth MF

    (2015) Contrasting Roles for Orbitofrontal Cortex and Amygdala in Credit Assignment and Learning in Macaques

    Neuron 87:1106–1118.

    https://doi.org/10.1016/j.neuron.2015.08.018

    • PubMed
    • Google Scholar
11. 1. Chaudhuri R
    2. Knoblauch K
    3. Gariel MA
    4. Kennedy H
    5. Wang XJ

    (2015) A Large-Scale Circuit Mechanism for Hierarchical Dynamical Processing in the Primate Cortex

    Neuron 88:419–431.

    https://doi.org/10.1016/j.neuron.2015.09.008

    • PubMed
    • Google Scholar
12. 1. Chong TT
    2. Cunnington R
    3. Williams MA
    4. Kanwisher N
    5. Mattingley JB

    (2008) fMRI adaptation reveals mirror neurons in human inferior parietal cortex

    Current Biology: CB 18:1576–1580.

    https://doi.org/10.1016/j.cub.2008.08.068

    • PubMed
    • Google Scholar
13. 1. Churchland MM
    2. Cunningham JP
    3. Kaufman MT
    4. Foster JD
    5. Nuyujukian P
    6. Ryu SI
    7. Shenoy KV

    (2012) Neural population dynamics during reaching

    Nature 487:51–56.

    https://doi.org/10.1038/nature11129

    • PubMed
    • Google Scholar
14. 1. Colclough GL
    2. Brookes MJ
    3. Smith SM
    4. Woolrich MW

    (2015) A symmetric multivariate leakage correction for MEG connectomes

    NeuroImage 117:439–448.

    https://doi.org/10.1016/j.neuroimage.2015.03.071

    • PubMed
    • Google Scholar
15. 1. Colclough GL
    2. Woolrich MW
    3. Tewarie PK
    4. Brookes MJ
    5. Quinn AJ
    6. Smith SM

    (2016) How reliable are MEG resting-state connectivity metrics?

    NeuroImage 138:284–293.

    https://doi.org/10.1016/j.neuroimage.2016.05.070

    • PubMed
    • Google Scholar
16. 1. Cunningham JP
    2. Yu BM
    3. Bm Y

    (2014) Dimensionality reduction for large-scale neural recordings

    Nature Neuroscience 17:1500–1509.

    https://doi.org/10.1038/nn.3776

    • PubMed
    • Google Scholar
17. 1. de Lange FP
    2. Jensen O
    3. Dehaene S

    (2010) Accumulation of evidence during sequential decision making: the importance of top-down factors

    Journal of Neuroscience 30:731–738.

    https://doi.org/10.1523/JNEUROSCI.4080-09.2010

    • PubMed
    • Google Scholar
18. 1. de Lange FP
    2. Heilbron M
    3. Kok P

    (2018) How Do Expectations Shape Perception?

    Trends in Cognitive Sciences 22:764–779.

    https://doi.org/10.1016/j.tics.2018.06.002

    • PubMed
    • Google Scholar
19. 1. Desimone R
    2. Duncan J

    (1995) Neural mechanisms of selective visual attention

    Annual Review of Neuroscience 18:193–222.

    https://doi.org/10.1146/annurev.ne.18.030195.001205

    • PubMed
    • Google Scholar
20. 1. Doeller CF
    2. Barry C
    3. Burgess N

    (2010) Evidence for grid cells in a human memory network

    Nature 463:657–661.

    https://doi.org/10.1038/nature08704

    • PubMed
    • Google Scholar
21. 1. Dum RP
    2. Strick PL

    (2002) Motor areas in the frontal lobe of the primate

    Physiology & Behavior 77:677–682.

    https://doi.org/10.1016/s0031-9384(02)00929-0

    • PubMed
    • Google Scholar
22. 1. Duncan J
    2. Humphreys G
    3. Ward R

    (1997) Competitive brain activity in visual attention

    Current Opinion in Neurobiology 7:255–261.

    https://doi.org/10.1016/s0959-4388(97)80014-1

    • PubMed
    • Google Scholar
23. 1. Egner T
    2. Hirsch J

    (2005) Cognitive control mechanisms resolve conflict through cortical amplification of task-relevant information

    Nature Neuroscience 8:1784–1790.

    https://doi.org/10.1038/nn1594

    • PubMed
    • Google Scholar
24. 1. Ekstrom AD
    2. Kahana MJ
    3. Caplan JB
    4. Fields TA
    5. Isham EA
    6. Newman EL
    7. Fried I

    (2003) Cellular networks underlying human spatial navigation

    Nature 425:184–188.

    https://doi.org/10.1038/nature01964

    • PubMed
    • Google Scholar
25. 1. Everling S
    2. Tinsley CJ
    3. Gaffan D
    4. Duncan J

    (2002) Filtering of neural signals by focused attention in the monkey prefrontal cortex

    Nature Neuroscience 5:671–676.

    https://doi.org/10.1038/nn874

    • PubMed
    • Google Scholar
26. 1. Fell J
    2. Klaver P
    3. Lehnertz K
    4. Grunwald T
    5. Schaller C
    6. Elger CE
    7. Fernández G

    (2001) Human memory formation is accompanied by rhinal-hippocampal coupling and decoupling

    Nature Neuroscience 4:1259–1264.

    https://doi.org/10.1038/nn759

    • PubMed
    • Google Scholar
27. 1. Fouragnan EF
    2. Chau BKH
    3. Folloni D
    4. Kolling N
    5. Verhagen L
    6. Klein-Flügge M
    7. Tankelevitch L
    8. Papageorgiou GK
    9. Aubry JF
    10. Sallet J
    11. Rushworth MFS

    (2019) The macaque anterior cingulate cortex translates counterfactual choice value into actual behavioral change

    Nature Neuroscience 22:797–808.

    https://doi.org/10.1038/s41593-019-0375-6

    • PubMed
    • Google Scholar
28. 1. Friston K

    (2005) A theory of cortical responses

    Philosophical transactions of the Royal Society of London. Series B, Biological Sciences 360:815–836.

    https://doi.org/10.1098/rstb.2005.1622

    • PubMed
    • Google Scholar
29. 1. Friston K
    2. Kiebel S

    (2009) Predictive coding under the free-energy principle

    Philosophical transactions of the Royal Society of London. Series B, Biological sciences 364:1211–1221.

    https://doi.org/10.1098/rstb.2008.0300

    • PubMed
    • Google Scholar
30. 1. Fritsche M
    2. Lawrence SJD
    3. de Lange FP

    (2020) Temporal tuning of repetition suppression across the visual cortex

    Journal of Neurophysiology 123:224–233.

    https://doi.org/10.1152/jn.00582.2019

    • PubMed
    • Google Scholar
31. 1. Garvert MM
    2. Moutoussis M
    3. Kurth-Nelson Z
    4. Behrens TE
    5. Dolan RJ

    (2015) Learning-induced plasticity in medial prefrontal cortex predicts preference malleability

    Neuron 85:418–428.

    https://doi.org/10.1016/j.neuron.2014.12.033

    • PubMed
    • Google Scholar
32. 1. Garvert MM
    2. Dolan RJ
    3. Behrens TE

    (2017) A map of abstract relational knowledge in the human hippocampal-entorhinal cortex

    eLife 6:e17086.

    https://doi.org/10.7554/eLife.17086

    • PubMed
    • Google Scholar
33. 1. Gazzaley A
    2. Rissman J
    3. Cooney J
    4. Rutman A
    5. Seibert T
    6. Clapp W
    7. D'Esposito M

    (2007) Functional interactions between prefrontal and visual association cortex contribute to Top-Down modulation of visual processing

    Cerebral Cortex 17:i125–i135.

    https://doi.org/10.1093/cercor/bhm113

    • Google Scholar
34. 1. Glaze CM
    2. Kable JW
    3. Gold JI

    (2015) Normative evidence accumulation in unpredictable environments

    eLife 4:e08825.

    https://doi.org/10.7554/eLife.08825

    • Google Scholar
35. 1. Gold JI
    2. Stocker AA

    (2017) Visual Decision-Making in an uncertain and dynamic world

    Annual Review of Vision Science 3:227–250.

    https://doi.org/10.1146/annurev-vision-111815-114511

    • PubMed
    • Google Scholar
36. 1. Grill-Spector K
    2. Henson R
    3. Martin A

    (2006) Repetition and the brain: neural models of stimulus-specific effects

    Trends in Cognitive Sciences 10:14–23.

    https://doi.org/10.1016/j.tics.2005.11.006

    • PubMed
    • Google Scholar
37. 1. Gross CG
    2. Schiller PH
    3. Wells C
    4. Gerstein GL

    (1967) Single-unit activity in temporal association cortex of the monkey

    Journal of Neurophysiology 30:833–843.

    https://doi.org/10.1152/jn.1967.30.4.833

    • PubMed
    • Google Scholar
38. 1. Harvey CD
    2. Coen P
    3. Tank DW

    (2012) Choice-specific sequences in parietal cortex during a virtual-navigation decision task

    Nature 484:62–68.

    https://doi.org/10.1038/nature10918

    • PubMed
    • Google Scholar
39. 1. Henson RN
    2. Rylands A
    3. Ross E
    4. Vuilleumeir P
    5. Rugg MD

    (2004) The effect of repetition lag on electrophysiological and haemodynamic correlates of visual object priming

    NeuroImage 21:1674–1689.

    https://doi.org/10.1016/j.neuroimage.2003.12.020

    • PubMed
    • Google Scholar
40. 1. Henson RN

    (2016) Repetition suppression to faces in the fusiform face area: a personal and dynamic journey

    Cortex 80:174–184.

    https://doi.org/10.1016/j.cortex.2015.09.012

    • PubMed
    • Google Scholar
41. 1. Hochstein S
    2. Ahissar M

    (2002) View from the top: hierarchies and reverse hierarchies in the visual system

    Neuron 36:791–804.

    https://doi.org/10.1016/s0896-6273(02)01091-7

    • PubMed
    • Google Scholar
42. 1. Hunt LT
    2. Kolling N
    3. Soltani A
    4. Woolrich MW
    5. Rushworth MF
    6. Behrens TE

    (2012) Mechanisms underlying cortical activity during value-guided choice

    Nature Neuroscience 15:470–476.

    https://doi.org/10.1038/nn.3017

    • PubMed
    • Google Scholar
43. 1. Hunt LT
    2. Behrens TE
    3. Hosokawa T
    4. Wallis JD
    5. Kennerley SW

    (2015) Capturing the temporal evolution of choice across prefrontal cortex

    eLife 4:e11945.

    https://doi.org/10.7554/eLife.11945

    • PubMed
    • Google Scholar
44. 1. Hunt LT
    2. Malalasekera WMN
    3. de Berker AO
    4. Miranda B
    5. Farmer SF
    6. Behrens TEJ
    7. Kennerley SW

    (2018) Triple dissociation of attention and decision computations across prefrontal cortex

    Nature Neuroscience 21:1471–1481.

    https://doi.org/10.1038/s41593-018-0239-5

    • PubMed
    • Google Scholar
45. 1. Jenkins AC
    2. Macrae CN
    3. Mitchell JP

    (2008) Repetition suppression of ventromedial prefrontal activity during judgments of self and others

    PNAS 105:4507–4512.

    https://doi.org/10.1073/pnas.0708785105

    • PubMed
    • Google Scholar
46. 1. Kaufman MT
    2. Churchland MM
    3. Ryu SI
    4. Shenoy KV

    (2014) Cortical activity in the null space: permitting preparation without movement

    Nature Neuroscience 17:440–448.

    https://doi.org/10.1038/nn.3643

    • PubMed
    • Google Scholar
47. 1. Klein-Flügge MC
    2. Barron HC
    3. Brodersen KH
    4. Dolan RJ
    5. Behrens TE

    (2013) Segregated encoding of reward-identity and stimulus-reward associations in human orbitofrontal cortex

    Journal of Neuroscience 33:3202–3211.

    https://doi.org/10.1523/JNEUROSCI.2532-12.2013

    • PubMed
    • Google Scholar
48. 1. Klein-Flügge MC
    2. Wittmann MK
    3. Shpektor A
    4. Jensen DEA
    5. Rushworth MFS

    (2019) Multiple associative structures created by reinforcement and incidental statistical learning mechanisms

    Nature Communications 10:1–15.

    https://doi.org/10.1038/s41467-019-12557-z

    • Google Scholar
49. 1. Kok P
    2. Bains LJ
    3. van Mourik T
    4. Norris DG
    5. de Lange FP

    (2016) Selective Activation of the Deep Layers of the Human Primary Visual Cortex by Top-Down Feedback

    Current Biology: CB 26:371–376.

    https://doi.org/10.1016/j.cub.2015.12.038

    • PubMed
    • Google Scholar
50. 1. Kok P
    2. Mostert P
    3. de Lange FP

    (2017) Prior expectations induce prestimulus sensory templates

    PNAS 114:10473–10478.

    https://doi.org/10.1073/pnas.1705652114

    • PubMed
    • Google Scholar
51. 1. Kriegeskorte N
    2. Mur M
    3. Ruff DA
    4. Kiani R
    5. Bodurka J
    6. Esteky H
    7. Tanaka K
    8. Bandettini PA

    (2008) Matching categorical object representations in inferior temporal cortex of man and monkey

    Neuron 60:1126–1141.

    https://doi.org/10.1016/j.neuron.2008.10.043

    • PubMed
    • Google Scholar
52. 1. Larsson J
    2. Smith AT

    (2012) fMRI repetition suppression: neuronal adaptation or stimulus expectation?

    Cerebral Cortex 22:567–576.

    https://doi.org/10.1093/cercor/bhr119

    • Google Scholar
53. 1. Law CT
    2. Gold JI

    (2008) Neural correlates of perceptual learning in a sensory-motor, but not a sensory, cortical area

    Nature Neuroscience 11:505–513.

    https://doi.org/10.1038/nn2070

    • PubMed
    • Google Scholar
54. 1. Law CT
    2. Gold JI

    (2010) Shared mechanisms of perceptual learning and decision making

    Topics in Cognitive Science 2:226–238.

    https://doi.org/10.1111/j.1756-8765.2009.01044.x

    • PubMed
    • Google Scholar
55. 1. Li N
    2. Daie K
    3. Svoboda K
    4. Druckmann S

    (2016) Robust neuronal dynamics in premotor cortex during motor planning

    Nature 532:459–464.

    https://doi.org/10.1038/nature17643

    • PubMed
    • Google Scholar
56. 1. Love J
    2. Selker R
    3. Marsman M
    4. Jamil T
    5. Dropmann D
    6. Verhagen J
    7. Ly A
    8. Gronau QF
    9. Smíra M
    10. Epskamp S
    11. Matzke D
    12. Wild A
    13. Knight P
    14. Rouder JN
    15. Morey RD
    16. Wagenmakers E-J

    (2019) JASP: Graphical Statistical Software for Common Statistical Designs

    Journal of Statistical Software 88:1–17.

    https://doi.org/10.18637/jss.v088.i02

    • Google Scholar
57. 1. Mante V
    2. Sussillo D
    3. Shenoy KV
    4. Newsome WT

    (2013) Context-dependent computation by recurrent dynamics in prefrontal cortex

    Nature 503:78–84.

    https://doi.org/10.1038/nature12742

    • PubMed
    • Google Scholar
58. 1. Michalareas G
    2. Vezoli J
    3. van Pelt S
    4. Schoffelen JM
    5. Kennedy H
    6. Fries P

    (2016) Alpha-Beta and Gamma Rhythms Subserve Feedback and Feedforward Influences among Human Visual Cortical Areas

    Neuron 89:384–397.

    https://doi.org/10.1016/j.neuron.2015.12.018

    • PubMed
    • Google Scholar
59. 1. Miller JF
    2. Neufang M
    3. Solway A
    4. Brandt A
    5. Trippel M
    6. Mader I
    7. Hefft S
    8. Merkow M
    9. Polyn SM
    10. Jacobs J
    11. Kahana MJ
    12. Schulze-Bonhage A

    (2013) Neural activity in human hippocampal formation reveals the spatial context of retrieved memories

    Science 342:1111–1114.

    https://doi.org/10.1126/science.1244056

    • PubMed
    • Google Scholar
60. 1. Moore T
    2. Zirnsak M

    (2017) Neural mechanisms of selective visual attention

    Annual Review of Psychology 68:47–72.

    https://doi.org/10.1146/annurev-psych-122414-033400

    • PubMed
    • Google Scholar
61. 1. Morcos AS
    2. Harvey CD

    (2016) History-dependent variability in population dynamics during evidence accumulation in cortex

    Nature Neuroscience 19:1672–1681.

    https://doi.org/10.1038/nn.4403

    • PubMed
    • Google Scholar
62. 1. Murray JD
    2. Bernacchia A
    3. Roy NA
    4. Constantinidis C
    5. Romo R
    6. Wang XJ

    (2017) Stable population coding for working memory coexists with heterogeneous neural dynamics in prefrontal cortex

    PNAS 114:394–399.

    https://doi.org/10.1073/pnas.1619449114

    • PubMed
    • Google Scholar
63. 1. Noudoost B
    2. Chang MH
    3. Steinmetz NA
    4. Moore T

    (2010) Top-down control of visual attention

    Current Opinion in Neurobiology 20:183–190.

    https://doi.org/10.1016/j.conb.2010.02.003

    • PubMed
    • Google Scholar
64. 1. Ossmy O
    2. Moran R
    3. Pfeffer T
    4. Tsetsos K
    5. Usher M
    6. Donner TH

    (2013) The timescale of perceptual evidence integration can be adapted to the environment

    Current biology: CB 23:981–986.

    https://doi.org/10.1016/j.cub.2013.04.039

    • PubMed
    • Google Scholar
65. 1. Passingham R

    (2009) How good is the macaque monkey model of the human brain?

    Current Opinion in Neurobiology 19:6–11.

    https://doi.org/10.1016/j.conb.2009.01.002

    • PubMed
    • Google Scholar
66. 1. Pezzulo G
    2. Cisek P

    (2016) Navigating the Affordance Landscape: Feedback Control as a Process Model of Behavior and Cognition

    Trends in Cognitive Sciences 20:414–424.

    https://doi.org/10.1016/j.tics.2016.03.013

    • PubMed
    • Google Scholar
67. 1. Piazza M
    2. Pinel P
    3. Le Bihan D
    4. Dehaene S

    (2007) A magnitude code common to numerosities and number symbols in human intraparietal cortex

    Neuron 53:293–305.

    https://doi.org/10.1016/j.neuron.2006.11.022

    • PubMed
    • Google Scholar
68. 1. Rao RP
    2. Ballard DH

    (1999) Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects

    Nature Neuroscience 2:79–87.

    https://doi.org/10.1038/4580

    • PubMed
    • Google Scholar
69. 1. Raposo D
    2. Kaufman MT
    3. Churchland AK

    (2014) A category-free neural population supports evolving demands during decision-making

    Nature Neuroscience 17:1784–1792.

    https://doi.org/10.1038/nn.3865

    • PubMed
    • Google Scholar
70. 1. Richter CG
    2. Thompson WH
    3. Bosman CA
    4. Fries P

    (2017) Top-Down Beta Enhances Bottom-Up Gamma

    Journal of Neuroscience 37:6698–6711.

    https://doi.org/10.1523/JNEUROSCI.3771-16.2017

    • PubMed
    • Google Scholar
71. 1. Richter D
    2. de Lange FP

    (2019) Statistical learning attenuates visual activity only for attended stimuli

    eLife 8:e47869.

    https://doi.org/10.7554/eLife.47869

    • PubMed
    • Google Scholar
72. 1. Rouder JN
    2. Morey RD
    3. Speckman PL
    4. Province JM

    (2012) Default bayes factors for ANOVA designs

    Journal of Mathematical Psychology 56:356–374.

    https://doi.org/10.1016/j.jmp.2012.08.001

    • Google Scholar
73. 1. Rushworth MF
    2. Johansen-Berg H
    3. Göbel SM
    4. Devlin JT

    (2003) The left parietal and premotor cortices: motor attention and selection

    NeuroImage 20:S89–S100.

    https://doi.org/10.1016/j.neuroimage.2003.09.011

    • PubMed
    • Google Scholar
74. 1. Rutishauser U
    2. Ross IB
    3. Mamelak AN
    4. Schuman EM

    (2010) Human memory strength is predicted by theta-frequency phase-locking of single neurons

    Nature 464:903–907.

    https://doi.org/10.1038/nature08860

    • PubMed
    • Google Scholar
75. 1. Siegel M
    2. Buschman TJ
    3. Miller EK

    (2015) Cortical information flow during flexible sensorimotor decisions

    Science 348:1352–1355.

    https://doi.org/10.1126/science.aab0551

    • PubMed
    • Google Scholar
76. 1. Squire RF
    2. Noudoost B
    3. Schafer RJ
    4. Moore T

    (2013) Prefrontal contributions to visual selective attention

    Annual Review of Neuroscience 36:451–466.

    https://doi.org/10.1146/annurev-neuro-062111-150439

    • PubMed
    • Google Scholar
77. Book

    1. Stokes M
    2. Duncan J

    (2014) Dynamic Brain States for Preparatory Attention and Working Memory

    Oxf Handb Atten.

    https://doi.org/10.1093/oxfordhb/9780199675111.013.03

    • Google Scholar
78. 1. Summerfield C
    2. de Lange FP
    3. de LFP

    (2014) Expectation in perceptual decision making: neural and computational mechanisms

    Nature reviews. Neuroscience 15:745–756.

    https://doi.org/10.1038/nrn3838

    • PubMed
    • Google Scholar
79. 1. Todorovic A
    2. de Lange FP

    (2012) Repetition suppression and expectation suppression are dissociable in time in early auditory evoked fields

    Journal of Neuroscience 32:13389–13395.

    https://doi.org/10.1523/JNEUROSCI.2227-12.2012

    • PubMed
    • Google Scholar
80. 1. Urai AE
    2. de Gee JW
    3. Tsetsos K
    4. Donner TH

    (2019) Choice history biases subsequent evidence accumulation

    eLife 8:e46331.

    https://doi.org/10.7554/eLife.46331

    • PubMed
    • Google Scholar
81. 1. van Wassenhove V
    2. Grant KW
    3. Poeppel D

    (2005) Visual speech speeds up the neural processing of auditory speech

    PNAS 102:1181–1186.

    https://doi.org/10.1073/pnas.0408949102

    • PubMed
    • Google Scholar
82. 1. Vidyasagar R
    2. Stancak A
    3. Parkes LM

    (2010) A multimodal brain imaging study of repetition suppression in the human visual cortex

    NeuroImage 49:1612–1621.

    https://doi.org/10.1016/j.neuroimage.2009.10.020

    • PubMed
    • Google Scholar
83. 1. Watrous AJ
    2. Miller J
    3. Qasim SE
    4. Fried I
    5. Jacobs J

    (2018) Phase-tuned neuronal firing encodes human contextual representations for navigational goals

    eLife 7:e32554.

    https://doi.org/10.7554/eLife.32554

    • PubMed
    • Google Scholar
84. 1. Woolrich M
    2. Hunt L
    3. Groves A
    4. Barnes G

    (2011) MEG beamforming using Bayesian PCA for adaptive data covariance matrix regularization

    NeuroImage 57:1466–1479.

    https://doi.org/10.1016/j.neuroimage.2011.04.041

    • PubMed
    • Google Scholar
85. 1. Wyart V
    2. Myers NE
    3. Summerfield C

    (2015) Neural mechanisms of human perceptual choice under focused and divided attention

    Journal of Neuroscience 35:3485–3498.

    https://doi.org/10.1523/JNEUROSCI.3276-14.2015

    • PubMed
    • Google Scholar
86. 1. Yuste R

    (2015) From the neuron doctrine to neural networks

    Nature Reviews. Neuroscience 16:487–497.

    https://doi.org/10.1038/nrn3962

    • PubMed
    • Google Scholar

Author details

1. Yu Takagi

   1. Wellcome Centre for Integrative Neuroimaging (WIN), Department of Experimental Psychology, University of Oxford, Oxford, United Kingdom
   2. Department of Psychiatry, University of Oxford, Warneford Hospital, Oxford, United Kingdom
   3. Department of Neuropsychiatry, Graduate School of Medicine, University of Tokyo, Tokyo, Japan

   Contribution

   Formal analysis, Visualization, Methodology, Writing - original draft, Writing - review and editing

   For correspondence

   yutakagi322@gmail.com

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-0503-785X

2. Laurence Tudor Hunt

   1. Department of Psychiatry, University of Oxford, Warneford Hospital, Oxford, United Kingdom
   2. Wellcome Centre for Integrative Neuroimaging (WIN), Centre for Functional MRI of the Brain (FMRIB), University of Oxford, Nuffield Department of Clinical Neurosciences, John Radcliffe Hospital, Oxford, United Kingdom

   Contribution

   Methodology, Writing - review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-8393-8533

3. Mark W Woolrich

   1. Department of Psychiatry, University of Oxford, Warneford Hospital, Oxford, United Kingdom
   2. Wellcome Centre for Integrative Neuroimaging (WIN), Centre for Functional MRI of the Brain (FMRIB), University of Oxford, Nuffield Department of Clinical Neurosciences, John Radcliffe Hospital, Oxford, United Kingdom

   Contribution

   Methodology, Writing - review and editing

   Competing interests

   No competing interests declared

4. Timothy EJ Behrens

   1. Wellcome Centre for Integrative Neuroimaging (WIN), Centre for Functional MRI of the Brain (FMRIB), University of Oxford, Nuffield Department of Clinical Neurosciences, John Radcliffe Hospital, Oxford, United Kingdom
   2. Wellcome Trust Centre for Neuroimaging, UCL Institute of Neurology, University College London (UCL), London, United Kingdom

   Contribution

   Conceptualization, Supervision, Writing - review and editing

   Contributed equally with

   Miriam C Klein-Flügge

   Competing interests

   Senior/Deputy editor, eLife

   "This ORCID iD identifies the author of this article:" 0000-0003-0048-1177

5. Miriam C Klein-Flügge

   1. Wellcome Centre for Integrative Neuroimaging (WIN), Department of Experimental Psychology, University of Oxford, Oxford, United Kingdom
   2. Wellcome Centre for Integrative Neuroimaging (WIN), Centre for Functional MRI of the Brain (FMRIB), University of Oxford, Nuffield Department of Clinical Neurosciences, John Radcliffe Hospital, Oxford, United Kingdom

   Contribution

   Conceptualization, Data curation, Formal analysis, Supervision, Validation, Investigation, Visualization, Methodology, Writing - original draft, Project administration, Writing - review and editing

   Contributed equally with

   Timothy EJ Behrens

   For correspondence

   miriam.klein-flugge@psy.ox.ac.uk

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-5156-9833

• 1,655

  views

• 154

  downloads

• 11

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.