Representations of evidence for a perceptual decision in the human brain

In perceptual decision making the brain extracts and accumulates decision evidence from a stimulus over time and eventually makes a decision based on the accumulated evidence. Several characteristics of this process have been observed in human electrophysiological experiments, especially an average build-up of motor-related signals supposedly reflecting accumulated evidence, when averaged across trials. A more direct approach to investigate the representation of decision evidence in brain signals is to correlate the trial-to-trial fluctuations of a model-based prediction of evidence with the measured signals. We here report results for an experiment in which we applied this approach to human magnetoencephalographic recordings. These results consolidate a range of previous findings and suggest that decision evidence is processed in three consecutive phases in the human brain: In an early phase around 120 ms after the evidence became visible on the screen, in a transition phase around 180 ms and a plateau phase roughly from 300 to 500 ms. We located sources of evidence representations in these phases in early visual (early), parietal (transition) and motor (plateau) regions of the brain while signals in posterior cingulate cortex represented decision evidence in all three phases. These findings imply that parietal cortex is only transiently involved in the processing of decision evidence, that motor areas represent accumulated evidence throughout decision making and that posterior cingulate cortex may have a central role in processing and maintaining decision evidence.


Abstract 13
In perceptual decision making the brain extracts and accumulates decision evidence from 14 a stimulus over time and eventually makes a decision based on the accumulated evidence. 15 Several characteristics of this process have been observed in human electrophysiological 16 experiments, especially an average build-up of motor-related signals supposedly reflecting 17 accumulated evidence, when averaged across trials. A more direct approach to investigate 18 the representation of decision evidence in brain signals is to correlate the trial-to-trial 19 fluctuations of a model-based prediction of evidence with the measured signals. We here 20 report results for an experiment in which we applied this approach to human 21 magnetoencephalographic recordings. These results consolidate a range of previous 22 findings and suggest that decision evidence is processed in three consecutive phases in the 23 human brain: In an early phase around 120 ms after the evidence became visible on the 24 screen, in a transition phase around 180 ms and a plateau phase roughly from 300 to 500 25 ms. We located sources of evidence representations in these phases in early visual (early), 26 parietal (transition) and motor (plateau) regions of the brain while signals in posterior 27 cingulate cortex represented decision evidence in all three phases. These findings imply 28 that parietal cortex is only transiently involved in the processing of decision evidence, that 29 motor areas represent accumulated evidence throughout decision making and that 30 posterior cingulate cortex may have a central role in processing and maintaining decision 31 evidence. 32 During perceptual decision making observers reason about the state of their environment. Supported 33 by findings in single neurons of non-human primates, the underlying mechanism has been 34 characterised as an accumulation-to-bound process 1 . Specifically, the current consensus is that during 35 perceptual decision making the brain accumulates noisy pieces of sensory evidence across time until it 36 reaches a confidence bound. In experimental settings where one condition may be associated with a 37 particular type of stimulus providing a fixed amount of evidence for a decision, this process predicts 38 that brain signals representing accumulated evidence exhibit a continuous build-up reaching a 39 maximum close to the response, when averaged across trials. Further, the slope of this average build-40 up is expected to depend on the amount of evidence provided by the stimulus. 41 In humans, evidence of this kind of average build-up have been found using magneto-and 42 encephalography (M/EEG). For example, lateralised oscillatory signals in the beta band measured with 43 magnetoencephalography exhibit this build-up, where sources were located to dorsal premotor and 44 primary motor cortex 2 . In EEG, there are similar findings of a build-up for lateralised readiness 45 potentials and oscillations 3,4 . Furthermore, when human participants have to detect the presence of 46 stimuli in noise, a centro-parietal positivity shows the characteristics of an evidence-dependent build-47 up independently of the type of stimulus used and the kind of response made 3,5 . Together these 48 findings suggest that the human parietal and motor cortices are involved in perceptual decision making 49 and in particular represent accumulated evidence. This view is compatible with electrophysiological 50 recordings in non-human animals 6 and an active role of the motor system during decision making 7 . 51 It has long been known that electromagnetic signals over motor areas build up towards a motor 52 response and can signal an eventual choice even before the response 8 . This means that the crucial 53 aspect of decision evidence representations is not the build-up as such, but its covariance with the 54 theoretically available evidence. In most previous experiments the amount of evidence provided by the 55 stimulus has only been varied across trials and only across few discrete levels 1,3-5 , but both of these 56 restrictions can be relaxed, if one has a model to predict the amount of decision evidence provided by 57 the stimulus at any given time 9-11 . Such model-based analyses are more specific than standard 58 analyses, because one can directly assess the covariation between decision evidence and neural signals 59 across a much richer sample of decision evidences. 60 In the present work we applied this approach to human MEG signals recorded during execution of a 61 two-alternative forced choice reaction time task. Using a computational model we derived the decision 62 evidence provided by stimuli which induced random evidence changes every 100 ms within a trial. We 63 assessed the covariation between the resulting decision evidence dynamics and MEG signals in sensor 64 and source space using regression. This analysis showed that MEG signals most strongly co-varied with 65 individual pieces of decision evidence 120 ms, 180 ms and within 300-500 ms after evidence updates 66 revealing for the first time that the processing of decision evidence may proceed in three consecutive 67 phases in the human brain. At the sensor level we found the strongest co-variation for these effects at 68 occipital, centro-parietal and central locations, respectively, mirroring the presumed transfer of 69 information from visual over parietal to motor areas, as supported by source reconstruction results. In 70 addition, our results implicate posterior cingulate cortex at all of these time points suggesting a central 71 role of this brain region in the transformation of sensory signals to decision evidence in our task. 72

73
While MEG was recorded, 34 human participants observed a single white dot on the screen changing 74 its position every 100 ms and had to decide whether a left or a right target (two yellow dots) was the 75 centre of the white dot movement (Figure 1). Participants indicated their choice with a button press 76 using the index finger of the corresponding hand. The distance of the target dots on the screen was 77 chosen in behavioural pilots so that participants had an intermediate accuracy around 75% while being 78 told to be as accurate and fast as possible. The average median response time across participants was 79 1.1 s with an average accuracy of 78% (cf. Figure 1). 80 81 An ideal observer model for inference about the target given a sequence of single dots has been 82 described before 12,13 . This model identifies the x-coordinates of the white dot positions as momentary 83 decision evidence while the y-coordinate only provides irrelevant perceptual information and acts as a 84 decision-unrelated control variable. Further, the sum of x-coordinates across single dot positions 85 reflects accumulated evidence and corresponds to the average state of a discrete-time drift-diffusion 86 model 13 . 87 Participants integrate evidence provided by single dot positions to make decisions 88 As the task required and the model predicted, participants made their decision based on the provided 89 evidence. In Figure 2 we show this as the correlation of participants' choices with momentary and 90 accumulated evidence. Momentary evidence was mildly correlated with choices throughout the trial 91 (correlation coefficients around 0.3) while the correlation between accumulated evidence and choices 92 increased to a high level (around 0.7) as more and more dot positions were presented. This result 93 indicates that participants accumulated the momentary evidence, here the x-coordinate of the dot, to 94 make their choices. In contrast, as expected, the y-coordinates had no influence on the participants' 95 choices as indicated by correlation coefficients around 0 ( Figure 2B). 96 Previous work has investigated the influence of individual stimulus elements on the eventual decision 97 and whether this influence differed across elements 9,14 . In our analysis this corresponds to checking 98 whether the correlations with momentary evidence shown in Figure 2A  to the small influence of the 4 th dot) and a manipulation of the 5 th dot to create large variation in x-105 coordinates (see Methods for further details). Taken together these results confirm that the used 106 stimuli were effective in driving the decisions of the participants and that the model predictions of 107 momentary and accumulated evidence integrate well with observed behaviour. 108 Figure 1. Course of events within a trial in the single dot task (A) and behaviour of individual participants (B). Each trial started with the presentation of a fixation cross, followed by the appearance of the two yellow targets after about 1 s. 700 ms after the appearance of the targets the fixation cross disappeared and a single white dot was presented at a random position on the screen (drawn from a 2D-Gaussian distribution centred on one of the targets). Every 100 ms the position of the white dot was changed to a new random draw from the same distribution. Participants were instructed to indicate the target which they thought was the centre of the observed dot positions. and downs of variables such as the momentary evidence, across trials. 120 As a first result, we found that correlations between momentary evidence and MEG signals followed a 121 stereotypical temporal profile after each dot position update (cf. Supplementary Figure 1). Therefore, 122 we performed an expanded regression analysis where we explicitly modelled the time from each dot 123 position update, which we call 'dot onset' in the following. To exclude the possibility that effects 124 signalling the button press motor response influence the results of the dot onset aligned analysis, we 125 only included data, for each trial, up until at most 200 ms prior to the participant's response. 126 We first identified time points at which the MEG signal correlated most strongly with the momentary 127 evidence. To do this we performed separate regression analyses for each time point from dot onset, 128 magnetometer sensor and participant, computed the mean regression coefficients across participants, 129 took their absolute value to yield a magnitude and averaged them across sensors. Figure 3 shows that 130 the strongest correlations between momentary evidence and magnetometer signals occurred at 120 131 ms, 180 ms and in a prolonged period from roughly 300 to 500 ms after dot onset. In contrast, 132 correlations with the control, that is, the dot y-coordinates, were significantly lower in this period from 133   (two-tailed Wilcoxon test for absolute average coefficients across all sensors and times  134 within 300-500 ms, W = 382781, p << 0.001). 135

136
The sensor topographies shown in Figure 3 indicate for the momentary evidence a progression of the 137 strongest correlations from an occipital positivity over a centro-parietal positivity to a central positivity. 138 y-coordinate correlations, on the other hand, remained spatially at occipito-parietal sensors. 139 Correlations with accumulated evidence 140 Guided by the model we used dot x-coordinates as representation of momentary evidence, but dot x-141 coordinates also do have purely perceptual interpretations as they simply measure the horizontal 142 location of a visual stimulus. Correlations with x-coordinates, therefore, may reflect early visual 143 processes independent of the decision. Contrasting the strength of significant effects for x-and y-144 coordinates ( Figure 3) already suggested that at least from 400 ms after dot onset x-coordinates indeed 145 represented momentary evidence. To further corroborate this supposition we turned to a form of 146 decision evidence that has no direct purely perceptual interpretation and is more closely related to the 147 decision itself, the accumulated evidence. 148 As the accumulated evidence is simply the sum of previously observed momentary evidences, the two 149 are strongly entangled. We therefore opted for running an independent regression analysis including 150 accumulated evidence as regressor of interest instead of momentary evidence and refer the reader to 151 Supplementary Material for motivation of this approach and a discussion of the caveats. Accumulated 152 The y-coordinate correlations are visibly and significantly weaker than for the evidence, but there are two prominent peaks from about 120 ms to 210 ms and at 320 ms after dot onset. There is no sustained correlation with the y-coordinate beyond 400 ms and the topographies of magnetometers differ strongly between evidence and y-coordinates. Specifically, the evidence exhibits occipital, centro-parietal and central topographies whereas the y-coordinate exhibits strong correlations only in lateral occipito-parietal sensors. evidence is, through the final choice, more strongly related to the motor response than the momentary 153 evidence (cf. Figure 2A, Supplementary Figure 2) which means that some effects indicated by the 154 accumulated evidence regressor may be attributed to the motor response and not the accumulated 155 evidence. To account for this potential confound we excluded from this analysis all data later than 200 156 ms before the response so that the results only contain effects unrelated to the motor response. 157 Figure 4 depicts the time course of overall correlation magnitudes for accumulated evidence together 158 with effect topographies at chosen time points. We found correlations between the MEG signal and 159 accumulated evidence at all of peri-stimulus time until about 550 ms after dot onset. Crucially, at all 160 time points of that period we observed centro-parietal and, especially, central sensor topographies 161 suggesting that these represent specifically decision-relevant information such as momentary or 162 accumulated evidence, as hypothesised based on the correlations with x-coordinates shown in Figure  163 3. For further discussion of the time course of accumulated evidence, see Supplementary material. 164 165 Sources of stimulus-aligned momentary evidence effects 166 By investigating the sources of the evidence correlations at sensor level, we aimed to better understand 167 the nature of these effects and to confirm their locations in the brain suggested by the shown sensor 168 topographies. In particular, we were interested in linking the time points at which we found strong 169 momentary evidence correlations to potential functional stages in the processing of decision evidence, 170 such as sensory processing, relating sensory information to the decision and integrating momentary 171 evidence with previous evidence. 172 We reconstructed source currents along the cerebral cortex for each participant and subsequently 173 repeated our regression analysis on the estimated sources. Specifically, we performed source 174 reconstruction on the preprocessed MEG data using noise-normalised minimum norm estimation 175 within 180 brain areas defined by a recently published brain atlas 20 . This resulted in average time 177 courses for each experimental trial in each of the 180 brain areas defined per hemisphere for each 178 participant. We then applied the expanded regression analysis to these source-reconstructed time 179 courses instead of onto MEG sensors. Following the summary statistics approach we identified time 180 points and areas with significant second-level correlations by performing t-tests across participants and 181 applying multiple comparison correction using false discovery rate 21 simultaneously across all time 182 points and brain areas. 183 The time course of correlation magnitudes shown in Figure 3 suggested three time windows at which 184 particularly strong correlations with momentary evidence were present in the brain. Consequently, we 185 concentrated on these time windows in our analysis. We call these time windows "early" (110 ms -130 186 ms), "transition" (160 ms -200 ms) and "plateau" (300 ms -500 ms). Figure 5 depicts the brain areas 187 with significant correlations for each of these time windows with colour scale indicating the average t-188 value magnitudes within the time window (we chose to display t-value magnitudes instead of 189 correlation magnitudes here, because the estimated correlation values had larger second-level 190 variability differences across brain areas than sensors). We provide a full list of significant effects across 191 all investigated time points in Supplementary Table 1. 192 193 As the sensor topographies suggested, we observed that in the early window the strongest correlations 194 were located in visual areas such as V3, V1 and areas in the lateral occipital cortex (e.g., FST, MST,LO3 195 according to 20 ), but also in a small area of posterior cingulate cortex (v23ab) and there was an effect in 196 a parietal area of the left hemisphere (MIP). In the transition window most of the correlations in visual 197 areas, especially those in lateral occipital areas, vanished. Instead, more parietal areas exhibited 198 significant correlations with momentary evidence, especially in the right inferior (IP0, PGp) and superior 199 parietal cortex (VIP, 7AL, 7Am). Additionally, we found strong correlations in posterior cingulate cortex 200 (POS2 and DVT). In the plateau window some correlations in parietal areas persisted, but only focal at 201 some time points so that on average across the time window correlations were weak compared to 202 other brain areas. Specifically, the strongest correlations were spread across the posterior cingulate 203 cortex in both hemispheres (especially areas v23ab, 31pd, 7m, 31pv, d23ab). Further strong 204 correlations occurred in motor areas, especially in the left hemisphere, including somatosensory areas 205 (3a, 3b, 1), primary motor cortex (area 4) and premotor areas (6a, 6d). Note that we excluded from the 206 analysis all time points later than 200 ms before the trial-specific motor response. Additionally, we 207 observed weaker correlations in mid and anterior cingulate motor areas (e.g., 24dv, p24pr). These 208 results confirm that decision-relevant information shifts from visual over parietal areas towards motor 209 areas where momentary evidence appears to be represented over a longer time period. The results 210 also reveal that source currents of brain areas in posterior cingulate cortex had strong correlations with 211 momentary evidence throughout all three time windows. Accordingly, the areas with the largest 212 correlation magnitudes on average across all time points within 0 to 500 ms were predominantly 213 located in posterior cingulate cortex (5 areas with strongest average effects in that order: left -v23ab, 214 3a, 31pd, 3b, 1; right -v23ab, DVT, d23ab, 31pv, 7m). This suggests a potentially central role of posterior 215 cingulate cortex in the processing of momentary evidence in the task. 216 Sources of stimulus-aligned accumulated evidence effects 217 The sensor topographies for the accumulated evidence effects suggested that accumulated evidence 218 was represented in common brain sources across the whole time window of 0 to 550 ms from dot 219 onset. Therefore, we used this full time window to investigate the underlying sources. As for the 220 momentary evidence, cf. Figure 5, we identified brain areas with significant correlations after FDR 221 correction across locations and times (p < 0.05, no significant effects for p < 0.01) in at least one time 222 point and then averaged the t-value magnitudes across time points within the time window in these 223 areas. Given the similarity of sensor topographies of momentary evidence in the plateau window and 224 the sensor topographies of accumulated evidence we expected their sources to overlap. 225 226 In Figure 6, one can see that, although the estimated correlation magnitudes were slightly higher for 227 the accumulated evidence than for the momentary evidence, fewer effects were statistically significant 228 for accumulated evidence. This is most likely because the variability of correlation magnitudes across 229 participants increased relative to momentary evidence effects (results not shown). Otherwise, the 230 identified brain areas were consistent with those of the momentary evidence in the plateau window. 231 In particular, we observed consistently strong correlations with accumulated evidence in motor, 232 premotor, cingulate motor and posterior cingulate areas. 233 Correlations with choice reveal response-aligned build-up and separate motor response 234 Our finding that momentary or accumulated evidence appears is represented in motor areas is 235 consistent with a wide range of previous work 2-4,10,22,23 . If motor areas are involved in processing 236 momentary or accumulated evidence prior to a response, as these results indicate, the question arises 237 how these processes relate to motor processes linked to the response itself. More specifically, we were 238 interested in how the patterns of correlations with momentary and accumulated evidence related to 239 correlation patterns representing the motor response and whether these could be linked to the 240 absence or presence of the involvement of certain brain areas. To investigate correlation patterns 241 representing the motor response we computed choice-dependent effects centred on the response 242 time of the participants. We did this with a regression analysis using the participant choice as a 243 regressor of interest (see Methods). The choice regressor provides a measure for how well the choice 244 of the participants can be decoded from univariate brain signals. 245 Figure 6. Sustained correlations with accumulated evidence in motor and cingulate areas. Following the procedure in Figure  5, we coloured only areas with a significant correlation with accumulated evidence (p < 0.05 FDR corrected) with colour indicating the average t-value magnitude in the extended time window from 0 ms to 550 ms after dot onset. The 5 largest effects were (marked by black boundaries): left -3a, 6d, 1, 2, v23ab; right -V6, 6a, 7m, p24pr, 24dv. sensors. From about 500 ms before the response, correlations between choice and MEG data became 248 gradually stronger culminating in an expected peak centred slightly after the response. The sensor 249 topographies of the build-up period before the response strongly resembled those we found for 250 accumulated evidence in our previous analyses. In fact, these results most likely correspond to the same 251 effect, because the participant choice itself was increasingly correlated with accumulated evidence as 252 the trial progressed (cf. Figure 2). That is, the build-up seen in the figure only indirectly visualises an 253 increasing evidence signal by depicting an increasing alignment of the final choice with the internal 254 representation before the response (presumably accumulated evidence). 255 The motor response itself (peak around 30 ms) was, as expected, much more strongly represented in 256 the MEG signals than the accumulated evidence, see Figure 7. Although the motor response also had a 257 predominantly central topography, its topography visibly differed from that prior to the response 258 (at -300 and -120 ms). Specifically, the topography before the response exhibited stronger anti-259 Figure 7. The button press motor response is also represented most strongly in central magnetometers, but the corresponding topography differs slightly from that associated with momentary and accumulated evidence. We computed the correlation between participant choices and MEG magnetometers using linear regression for data aligned at response time. Following the format of Figure 3 we here show the time course of the mean (across sensors) magnitude of grand average regression coefficients (β). Sensor topographies for time points indicated by the black dots are shown below the main panel. Note that for the time points before the response we use a different scaling of colours than for time points around the response and later. This is to more clearly visualise the topography around the response which contains larger values. The colour scaling for the time points before the response is equal to that of Figure 3 and Figure 4. The topography at -300 ms strongly resembled that for accumulated evidence, but the topography around the response (30 ms) additionally exhibited stronger fronto-lateral and weaker occipital anti-correlations (p < 0.01 corrected, cf. Supplementary Figure 6). Positive values / correlations mean that measured sensor values tended to be high for a right choice (button press) and low for a left choice and vice-versa for negative values. See Supplementary Figure 5 to see how the central topography at 30 ms shown here results as the difference of the topographies associated with right and left choices. correlation in occipital sensors than around the response while the topography around the response 260 exhibited stronger anti-correlations in fronto-lateral sensors (p < 0.01 corrected, cf. Supplementary 261 Figure 6). Furthermore, the correlation with choice was relatively higher over central sensors at 30 ms 262 than at -120 ms (Supplementary Figure 6). 263 To analyse this difference at the source level we applied the regression analysis to the reconstructed 264 source currents. Figure 8 depicts the results of an analysis of two time windows: the "build-up" window 265 from -500 ms to -120 ms (when a dip before the response indicates an end of the build-up) and the 266 "response" window capturing the response peak from -30 ms to 100 ms. We only show brain areas 267 with a significant effect within the time window after correcting for multiple comparisons (FDR with α 268 = 0.01 across brain areas and the two time windows). The shown colours indicate normalised second-269 level t-value magnitudes (see Methods). 270 As expected, in the response window, the effects were dominated by choice correlations in bilateral 271 primary motor and somatosensory cortices, but also choice correlations in cingulate motor areas 272 (around Brodmann area 24) were among the effects with the strongest magnitudes. Other significant 273 correlations with choice within the response window occurred in premotor and posterior cingulate 274 cortices. In the build-up window, the strongest correlations occurred predominantly in cingulate motor 275 cortex and premotor areas (especially 6a). 276 We further aimed at identifying brain areas with significantly different correlation magnitudes in the 277 two time windows. Specifically, we were interested in the difference of the spatial patterns of 278 correlation magnitudes, across brain areas, between the two time windows. To do this, we normalised 279 correlation magnitudes across brain areas within the time windows and computed the differences 280 between time windows within each brain area and participant (see Methods for details). Figure 8, 281 bottom panel, shows that across participants the only statistically significant differences occurred in 282 the primary motor and somatosensory cortices and, with smaller effect size, in cingulate motor areas. 283 In all these areas correlation magnitudes were larger in the response as compared to the build-up 284 window. 285

286
In summary, the response-centred analysis of choice correlations suggests that the build-up of choice-287 correlations leading towards a response is related to the accumulation of momentary evidence, 288 because sensor topographies and brain areas were highly consistent across choice-and evidence-based 289 analyses. The correlation topographies for the build-up and the response windows shown in Figure 7 290 had significant differences in central, occipital and fronto-lateral sensors. When analysing these 291 differences at the source level (Figure 8), the only sources with significant differences were located in 292 motor areas. These results together suggest that the brain areas representing decision evidence are 293 largely overlapping with those representing the upcoming choice and the motor response. The 294 difference in correlation patterns at the source level between the upcoming choice and motor response 295 could be explained by an increase in choice correlations in motor areas. 296 Discussion 297 Using MEG, we have analysed the dynamics of evidence representations in the human brain during 298 perceptual decision making. Instead of using average, evidence-dependent build-ups of MEG signals 299 across trials, we directly assessed trial-wise MEG signals that correlated with stimulus-induced, model-300 based measures of rapidly changing decision evidence. Specifically, we induced fast, within-trial 301 evidence fluctuations using a visual stimulus in which new, momentary evidence appeared every 100 302 ms and correlated the resulting model-predicted momentary evidence dynamics with MEG signals. We 303 Figure 8. Around the response time strongest correlations with choice occurred in primary motor, somatosensory and cingulate motor cortex (BA 24) while during the build-up period we found the strongest effects in premotor and cingulate motor cortex. The 5 largest effects per hemisphere were: build-up, left -24dv, 6a, 24dd, 3a, SCEF; right -24dv, p24pr, 6a, SCEF, OP2-3; response, left -4, 3b, 24dv, 3a, SCEF; right -3b, 4, p24pr, 24dv, 2. When testing for differences in the spatial pattern of correlation magnitudes (see Methods) between the two time windows, we only found significant differences in the motor and cingulate areas: 1, 24dv, 2, 31a, 3a, 3b, 4, 6d, 6mp, SCEF, p24pr. All of these effects indicated that correlations with choice were stronger in the response window (blue). The build-up and response panels show spatially normalised t-value magnitudes while the difference panel shows t-values of spatially normalised correlation magnitude differences.
found that each update of momentary evidence elicited a stereotyped response in the MEG signal that 304 lasted until about 600 ms after the update onset, meaning that the brain processed incoming pieces of 305 momentary evidence in parallel. We identified three main phases of the representation of momentary 306 evidence: an early phase around 120 ms after an evidence update, a transition phase around 180 ms 307 and a plateau phase from about 300 to 500 ms. These phases exhibited different sensor topographies 308 with positive correlations shifting from occipital to centro-parietal to central sensors during the three 309 phases. Using source reconstruction, we localized these representations of momentary evidence in 310 early visual, parietal and motor areas, respectively, with significant correlations in posterior cingulate 311 cortex occurring in all three phases. Significant correlations with accumulated evidence including the 312 most recent evidence update occurred continuously until about 550 ms after update onset and 313 exhibited a central topography similar to that in the plateau phase of momentary evidence 314 representations with corresponding sources. Additionally, response-aligned correlations of the MEG 315 signal with the final choice of the participants shared a similar topography in a build-up phase hundreds 316 of milliseconds before the response. The correlation analysis at the source level further showed that 317 the only significant differences between build-up phase and motor response were higher choice 318 correlations in motor areas during the response. 319 These results consolidate a wide range of separate previous findings: Although it has previously been 320 shown that the human brain exhibits evidence-dependent responses to stimuli carrying individual 321 pieces of momentary evidence 9,24-26 , these studies had either rather long stimulus presentation times 322 atypical for fast perceptual decisions 25,26 , or did not employ a reaction time paradigm 9,24,26 . Crucially, 323 these studies did not identify the three different phases of momentary evidence processing, although 324 at least the early and plateau phases were previously hinted at 9 . 325 A large proportion of previous work investigating the dynamics of evidence representations in the 326 human brain focused on oscillatory signals 2,4,26,27 . For example, it has been found that the average 327 amount of evidence in a trial is represented in the power of oscillations in occipital and parietal cortex 328 27 . Further, the difference in the power of oscillations between central-left and central-right sensors 329 exhibits an evidence-dependent build-up towards the response that appears to be generated in motor 330 areas 2,4 . We here made corresponding observations, but directly in the trial-wise MEG signals reflecting 331 trial-wise signal variations correlated with decision evidence that are believed to result from minute, 332 event-related fluctuations in the voltage potentials of neuronal populations. 333 There is overwhelming evidence that motor areas including areas in the premotor and primary motor 334 cortex are involved in perceptual decision making, e.g. 6,28 . Specifically, it has been shown that some 335 single neurons in primary motor cortex represent momentary evidence 10 , that the strength of muscle 336 reflex gains is proportional to the average amount of momentary evidence within a trial 22 , that motor-337 evoked potentials can be related to accumulated evidence 23,29 and that classical lateralised readiness 338 potentials which are thought to represent motor processes 8 also exhibit evidence-dependent build-up 339 in a detection task 3 . Our results further substantiate these findings by showing that human motor areas 340 represent each update of momentary evidence roughly within 300 to 500 ms after the update onset 341 and that accumulated evidence is represented in motor areas throughout the decision making process. 342 Using a response-aligned analysis of choice-dependent effects in the same reference frame as the 343 analyses of evidence, we could further show that the stimulus-aligned evidence representations 344 resemble closely the representation of the final choice during a build-up phase before the motor 345 response. This supports the hypothesis that previous observations of pre-response representations of 346 an upcoming choice, such as the lateralised readiness potential, should be interpreted as expressions 347 of an ongoing decision making process about the next sensible motor response. In sum, the present 348 and previous findings strongly affirm a tight coupling between decision making and motor processes, 349 as, for example, formulated in the affordance competition hypothesis 7,30 , but also other theories in 350 cognitive computational neuroscience 31-33 . 351 One potential caveat of our correlation results in motor areas is that due to the specifics of our task 352 participants may actually have executed micro-movements trying to track the changes of the 353 perceptual stimulus either with their eyes, or with minimal finger movements close to the response 354 buttons. In this scenario the observed correlations in motor areas would be possibly explained by motor 355 signals to the muscles. Although we cannot completely exclude this possibility we deem it unlikely, 356 because: i) Stimuli were shown only very centrally at visual angles within about 10° visual angle with 357 most stimuli within 5° diameter from fixation meaning that most of them were well within the foveal 358 visual field. ii) The sensor topographies representing evidence were very similar to that associated with 359 the motor response, that is, the evidence representations do not appear to be specifically related to 360 eye movements. iii) As mentioned above, a large body of work employing a wide variety of different 361 tasks already supports the reverse interpretation that motor areas represent decision evidence before 362 motor execution. In conclusion, we do not believe that the correlations with momentary or 363 accumulated evidence observed in motor areas of the brain are merely an expression of motor control 364 signals that caused stimulus-correlated micro-movements. Even if such micro-movements existed, we 365 deem it likely that these follow the time-course of decision evidence rather than decision-irrelevant 366 stimulus properties, as suggested by recent results about the adaptation of reflex gains and motor 367 evoked potentials during decision making 22,23,29 . 368 The early, transition and plateau phases of momentary evidence representations mirrored the 369 presumed general transfer of behaviourally relevant visual information through the brain 34 . In the early 370 phase around 120 ms after evidence updates we found the strongest representations of momentary 371 evidence in early visual cortex and occipito-temporal areas while in the transition phase around 180 ms 372 momentary evidence representations included areas in inferior and superior parietal cortex. In the 373 plateau phase the momentary evidence was predominantly represented in pre-/motor, somatosensory 374 and cingulate areas while we only found one weak significant correlation with momentary evidence in 375 one area of parietal cortex (right PFt). The same was true for representations of accumulated evidence. 376 Taken together, these results indicate that in our task parietal cortex was only transiently involved in 377 the processing of momentary evidence and that it did not accumulate evidence for the decision, or at 378 least did not represent accumulated evidence over an extended period of time. 379 These results appear to be at odds with previous findings in non-human primates which had identified 380 neurons in inferior parietal cortex that seemed to represent accumulated evidence 1 . More recent work, 381 however, suggests that the firing of these neurons is more diverse than originally thought 35-37 . It is 382 possible that the signal from only few evidence accumulating neurons in inferior parietal cortex is too 383 weak to be recorded with MEG. 384 To manipulate decision evidence in our task we changed the position of a single dot presented on a 385 screen. Only the x-coordinates of these dot positions represented momentary decision evidence while 386 the decision-irrelevant y-coordinates acted as a perceptual control variable. We have shown that 387 correlations of MEG signals with the perceptual control variable, in contrast to momentary evidence, 388 were strongly diminished in the period from 300 to 500 ms after dot onset. This suggests that the brain 389 ceases to represent perceptual information that is behaviourally irrelevant around this time and that 390 brain areas with strong correlations with momentary evidence in this time window indeed are involved 391 in the decision making process. This interpretation is further supported by previous work which has 392 shown that purely perceptual stimulus information is represented in electrophysiological signals only 393 until about 400 ms after stimulus onset 9,38,39 while specifically decision-related information is 394 represented longer starting around 170 ms after stimulus onset 9,38-42 . 395 We further validated this interpretation by investigating correlations with accumulated evidence, that 396 is, the current sum of momentary evidences within a trial. In contrast to the momentary evidence, this 397 sum is more specifically related to the decision and has no simple, purely perceptual interpretation. 398 The similarity of the topographies for accumulated evidence correlations and for momentary evidence 399 correlations in the plateau phase suggests that specifically decision-relevant evidence is represented in 400 the plateau phase, that is, within 300 to 500 ms after evidence updates. Our results do not allow to 401 clearly state whether momentary, or accumulated, or both types of decision evidence were 402 represented in the brain in this time window, because both types of evidence are correlated, especially 403 early within a trial. However, we also found that accumulated evidence exhibited the corresponding 404 central topography more consistently throughout peri-stimulus time than momentary evidence, so it 405 appears reasonable to assume that predominantly accumulated evidence is represented in the plateau 406 phase. 407 follow one (press left) or another (press right) behaviour. 425

426
This study has been approved by the ethics committee of the Participants were seated in a dimly lit shielding room during the training and the MEG measurement. 460 Visual stimuli were presented using Presentation® software (Version 16.0, Neurobehavioral Systems, 461 Inc., Berkeley, CA, www.neurobs.com). The display was a semi-transparent screen onto which the 462 stimuli were back-projected from a projector located outside of the magnetic shielding room 463 (Vacuumschmelze Hanau, Germany). The display was located 90 cm from the participants. The task was 464 to find out which target (left or right) was the centre of the white dot positions, but participants were 465 instructed with a cover story: Each target represented a bee hive and the white dot represented a bee. 466 Participants should tell which bee hive is more likely the home of the bee. They were additionally 467 instructed to be both accurate and fast, but not too fast at the expense of being inaccurate, and not 468 too slow that the trial times out. was the mean signal in the period from -300 ms to 0 ms (first dot onset). 506 Source reconstruction 507 We reconstructed the source currents underlying the measured MEG signals using noise-normalised 508 minimum norm estimation 19 implemented in the MNE software. To create participant-specific forward 509 models we semi-automatically co-registered the head positions of participants with the MEG 510 coordinate frame while at the same time morphing the participants' head shape to that of Freesurfer's 511 fsaverage by aligning the fsaverage head surface to a set of head points recorded for each participant. 512 We defined a source space along the white matter surface of the average subject with 4098 equally 513 spaced sources per hemisphere and an approximate source spacing of about 5 mm (MNE's "oct6" 514 option). For minimum norm estimation we assumed a signal-to-noise ratio of 3 (lambda 2 = 0.11). We 515 estimated the noise covariance matrix for noise normalisation 19 from the MEG signals in the baseline 516 period spanning from 300 ms before to first dot onset in each trial. We further used standard loose 517 orientation constraints (loose=0.2), but subsequently picked only the currents normal to the cortical 518 mantle. We employed standard depth weighting with a value of 0.8 to overcome the bias of minimum 519 norm estimates towards superficial sources. We computed the inverse solution from all MEG sensors 520 (magnetometers and the two sets of gradiometers) returning dynamic statistical parametric maps for 521 each participant. Before some of the subsequent statistical analyses we averaged the reconstructed 522 source signals across all sources of a brain area as defined trial. See Figure 9 for an illustration. For each time after dot onset and for each participant we pooled 555 all of these data points across trials and inferred regression coefficients on these expanded data sets. 556 Note that this approach can equally be interpreted as statistical inference over how strongly the 557 sequence of momentary evidence caused by the dot updates is represented in the signal at 100 ms 558 wide steps with a delay given by the chosen time from dot onset. 559 These analyses included two regressors of interest: momentary evidence (x-coordinate) and y-560 coordinate of the associated dots. We additionally included the following nuisance regressors: an 561 intercept capturing average effects, the absolute values of x-and y-coordinates, perceptual update 562 Figure 9. Diagram demonstrating the selection of data points entering the expanded regression analyses. Dot positions (d1, d2, d3, …) changed every 100 ms in the experiment (black). Coloured dots indicate times at which signal data points entered the analysis for a given time from dot position change (dot onset, shown exemplarily for 80 and 220 ms from dot onset). We only considered time points up to 200 ms before the response in each trial. Coloured d1, d2, d3 above the points indicate the dot positions associated with the corresponding signal data points for the given time from dot onset. For each trial, these pairs of signal data and dot positions entered the expanded regression analyses.
variables for x-and y-coordinates 9 defined as the magnitude of the change from one dot position to 563 another and accumulated values of x-and y-coordinates. Because we found that the accumulated 564 values can be strongly correlated with the individual x-and y-coordinates (cf. Supplementary Figure 2), 565 we only used accumulated values up to the previous dot in the regressor. For example, if a data point 566 was associated with the y-coordinate of the 4 th dot, the accumulated regressor would contain the sum 567 of only the first three y-coordinates. This accumulated regressor is equal to the regressor resulting from 568 Gram-Schmidt orthonormalisation of the full sum of y-coordinates with respect to the last shown y-569 coordinate. The accumulated evidence regressor was derived from the ideal observer model as the log 570 posterior odds of the two alternatives, but this was almost 100% correlated with the simple sum of x-571 coordinates. The small differences between model-based accumulated evidence and sum of x-572 coordinates after normalisation resulted from a small participant-specific offset representing the 573 overall bias of the participant towards one decision alternative. 574 Because of only being able to use accumulated evidence up to the previous dot as a regressor together 575 with the x-coordinate regressors and because this meant that we could not get results for accumulated 576 evidence for the first 100 ms after dot onset, we decided to repeat the expanded regression analysis 577 replacing the x-coordinate regressor with the sum of x-coordinates (and dropping the previous 578 accumulated evidence regressor). We report results from this regression analysis in Figure 4, Figure 6, 579 Supplementary Figure 3 and Supplementary Figure 4. 580

Response-aligned regression analyses 581
Additional to the first-dot onset and dot onset aligned analyses, we conducted response-aligned 582 analyses in which time was referenced to trial-specific response times of participants. The regressors in 583 this analysis were the trial-specific choice of the participant, trial-time and an intercept. Choice was 584 encoded as -1 for left and +1 for right so that the direction of correlations was compatible with that for 585 the evidence regressors. The trial-time regressor simply counted the trial number within the 586 experiment per participant. Timed out trials were excluded from analysis. As in the other regression 587 analyses we z-scored regressors and data across trials before estimating the regression coefficients, 588 except for trial-time which was only scaled to standard deviation equal to 1. 589 We ran two different analyses in sensor and source space. In sensor space (magnetometers) we ran 590 independent univariate regressions for each combination of sensor and time so that we ran 102 * 70 591 regressions with maximally 480 data points (one per trial, minus excluded trials). We report results of 592 this analysis in Figure 7 and Supplementary Figure 6. After having identified time windows of interest 593 based on the sensor level results, we aggregated data from the identified times into a common 594 regression on source data. To do this we simply pooled the data from all times in the time window and 595 ran the regression on this expanded data set, then including maximally number of trials * number of 596 time points data points. This approach meant that we were automatically estimating the mean 597 regression coefficients across the selected time window for each brain area and participant. We report 598 results of this analysis in Figure 8. 599 Identification of significant source-level effects 600 To identify significant correlations between regressors of interest and source signals we followed the 601 summary statistics approach 47 and performed two-sided t-tests on the second level (group-level, t-602 tests across participants). We corrected for multiple comparisons across time points and brain areas by 603 controlling the false discovery rate using the Benjamini-Hochberg procedure 21 . Specifically, for 604 identifying significant effects reported in Figure 5 we corrected across 25,340 tests covering 70 time 605 points (0 to 690 ms from dot onset in 10 ms steps) and 362 brain areas (180 brain areas of interest per 606 hemisphere plus one collection of sources per hemisphere that fell between the area definitions 607 provided by the atlas). We report all significant effects of this analysis in Supplementary Table 1. 608 Identification of significant differences in correlation patterns 609 We formally investigated the differences in correlation patterns of the response-aligned analysis 610 between the two time windows of interest (Figure 8, Supplementary Figure 6). As we were interested 611 in the differences between spatial patterns, we accounted for the overall increase in correlation 612 magnitudes from build-up to response window by normalising the correlation magnitudes. This 613 normalisation consisted of first shifting the minimum magnitude to 0 and then scaling the resulting 614 magnitudes so that their mean equals 1 across sensors or brain areas. The initial shift of the magnitudes 615 prevents excessive shrinking of magnitude variances for magnitude patterns with overall large 616 magnitudes and ensures that the magnitudes have similar distributions across the involved sensors or 617 brain areas in both considered time periods. We subsequently computed the differences between the 618 selected time periods on the first level and report second-level (across participant) statistics. 619 The analysis on the source level in principle equalled that of the sensor level, but additionally accounted 620 for the fact that most brain areas were not involved in encoding the choice. We achieved this by 621 computing the normalisation parameters for a time window only across brain areas with a significant 622 effect in this time window. However, we then computed magnitude differences for all brain areas with 623 a significant effect in at least one of the time windows and proceeded with second-level statistics for 624 these areas as before.  Figure 3A) and two time-shifted replicas of it, one shifted by 100 ms to the past (mid grey) and another shifted by 200 ms to the past (light grey). These time courses, therefore, are associated with the representation of the momentary evidence / x-coordinates of the current and the previous two dots in the brain. This visualisation shows that peaks in accumulated evidence tend to coincide with peaks in momentary evidences presented at subsequent time points. Larger discrepancies between correlation magnitudes of momentary and accumulated evidence only occurred at 20 ms, 120 ms and from about 450 ms after dot onset. At 80 ms and 180 ms topographies slightly shifted towards parietal sensors otherwise effects were located centrally.
At 80 ms and 180 ms the sensor topographies for accumulated evidence deviated somewhat from a 784 central to a more centro-parietal positivity. We reported above that the peaks of accumulated evidence 785 coincided with peaks of momentary evidence at these time points. These momentary evidence peaks 786 corresponded to the 180 ms momentary evidence peak in relation to dot onset which had a centro-787 parietal topography as shown in Figure 3A.     Numerical values plotted in Figure 6 804 Supplementary  We formally tested the apparent differences in topographies of choice correlations shown in Figure 7 for time points -120 ms and 30 ms. As we were interested in the spatial patterns and not absolute value differences within sensors, we scaled the coefficient estimates (β) across sensors, but within time points and participants for this analysis so that their mean magnitude across sensors was equal to 1. We then computed the difference between time points within each participant and sensor.
The colouring in the plot shows the mean of these differences across participants. We further applied a t-test across participants within each sensor and corrected the resulting p-values for false discovery rate at α = 0.01 across sensors. The white dots in the figure indicate sensors which exhibited a significant difference after multiple comparison correction.
Together with the topographies shown in Figure 7 the results of this analysis confirm that before the response occipital sensors had stronger anti-correlation with choice than around the response. In contrast, fronto-lateral sensors exhibited stronger anti-correlation around the response than before the response. Furthermore, the strongest difference occurred in central sensors which exhibited a relatively stronger correlation with choice around the response than before the response.
Supplementary Figure 7. Time courses from the expanded regression analysis in dot onset aligned time in source space. Values are mean (across brain areas) magnitude of second-level regression coefficients (β). Shown are effects for all used regressors (excluding the intercept). All regressors were derived from the dot positions. Regressors derived from x-coordinates (evidence) are shown in blue while y-coordinate regressors are shown in orange. For reference we also plotted regression coefficients obtained from permuted data as dotted lines. We included 4 different measures in our analysis: 'momentary' (evidence) are the original x-and y-coordinates (cf. Figure 3), 'accumulated' corresponds to summed coordinates, but only up to the previous dot (see Methods), 'absolute' are the absolute coordinates measuring only the displacement of the dot from 0 in both directions and 'perceptual update' is the absolute difference between the latest and the previous dot positions (cf. Wyart et al., 2012). Only the regressors respecting the sign of the coordinates (momentary and accumulated) exhibit strong effects. Note that we shifted the effects for the accumulated regressors 100 ms to the right to account for them being defined for the previous dot position instead of the current ones as for the other regressors. Also, the effects measured with the accumulated regressors, although accumulated and momentary regressors were uncorrelated, are a mixture of effects attributable to raw x-, y-coordinates and their cumulative sum, because accumulated regressors of dot d-1, although being uncorrelated to momentary regressors of dot d, are correlated with momentary regressors of dot d-1 and earlier.