DANDI:000579

assigned coordinates to 93,881 layer 2/3 neurons imaged from 141 sessions of 8 mice ... we obtained 14,373 neurons in V1, 9,564 neurons in PM, 14,974 neurons in AM, 9,885 neurons in MM, 23,036 neurons in RSC, and 22,049 neurons in area A

Results, Calcium imaging of neural activity across posterior cortex, paragraph 1; STAR Methods, Area parcellation

With session fraction correct over all trials, 150 layer 2/3 sessions above 0.65 hold 97,989 neurons from 8 mice (V1 15,497, PM 10,313, AM 15,225, MM 10,654, RSC 23,728, A 22,572), each within its tolerance; the area shares are within 0.006 of the paper's. With visually guided trials excluded from fraction correct, 100 sessions and 69,707 neurons remain and only the mouse count is within tolerance. The files hold 278,155 neurons in 300 sessions, the paper's total, so the stored sources are the paper's neurons; the paper's list of 141 sessions cannot be recovered.

How the test differed from the paper’s method

• Sessions above 0.65 are selected on fraction correct over all trials, an assumption (Methods, Classification of the sources, gives 141 layer 2/3 sessions above 65% correct; Behavioral training says visually guided trials are excluded from all analyses). Only fraction correct over all trials gives counts near the paper's (150 layer 2/3 sessions, 97,989 neurons); with visually guided trials excluded there are 100 and 69,707. Both selections are shown.
• The paper's cell-body classifier is not in the files; the stored sources are taken as the cell bodies, since their total equals the paper's 278,155 neurons.

Their behavioral performance was high before rule switches and typically dropped below chance after switches before gradually recovering to high accuracy by the end of a block, without signs of anticipating rule switches

Results, Mice learned flexible cue-choice associations, paragraph 2; Figure 1D; Figure S1B

Over 601 switches in 297 sessions above 0.5 (fraction correct over all trials), the end of the block before the switch is 0.807, the fitted initial performance after the switch 0.226 ± 0.016 (no bootstrap sample at or above 0.5), the fitted end performance 0.727 ± 0.026, and end minus initial 0.501 ± 0.038 (p < 0.0001). With visually guided trials excluded from the selection, 589 switches give 0.225 and 0.730. The end mean is above 0.7 only just: 15% of bootstrap samples (12% with the other selection) are at or below 0.7, and with the other selection two of the 8 mice end below 0.7 (0.613 and 0.666). With a free sigmoid midpoint the initial value is 0.317. The paper reports 506 to 513 switches. The fit overestimates initial performance a little on synthetic blocks (verifier), which works against the below-0.5 test.

How the test differed from the paper’s method

• The sigmoid's t0 is fixed at the block's first trial, as the Methods state (Task performance analysis); a free midpoint is shown as a variant. Rates are kept in (0, 1) by the fit's parametrisation, the claims file's standard choice.
• The session selection above 0.5 matches no reading of the paper's 253 sessions (291 or 297 here); both readings are shown and agree.

The LSTM accurately predicted a mouse's choices in held-out sessions (80.2% ± 5.1%, mean ± SD)

Results, Decision-making strategy varied, paragraph 1; STAR Methods, LSTM model

Accuracy of the stored predictions is 0.772 over all 300 sessions (SD 0.064 over sessions; 0.771, SD 0.043 over mice). Above 0.65 it is 0.792 (visually guided excluded from the selection, 154 sessions) and 0.769 (all trials, 238 sessions). Every reading is within 0.051 of 0.802. Per mouse 0.705 to 0.826.

How the test differed from the paper’s method

• The LSTM was not refit. The association matrices stored in the trial tables are taken as the paper's held-out (leave-one-session-out) predictions, an assumption, as the files do not say how they were produced; the Methods (LSTM model) describe only held-out predictions being used for analysis.

Log likelihood of the LSTM was greater than the RL model (p = 5.95 × 10-22, Wilcoxon signed rank test)

Figure S1H

What the paper reports

• Wilcoxon signed rank test, p = 6 × 10⁻²² (n = 216 sessions)

On the 238 sessions above 0.65 (all trials), held-out log2 likelihood per trial is -0.710 for the LSTM and -0.732 for RL model 4; the LSTM is higher in 185 sessions, Wilcoxon p = 4.65e-22 (paper 5.95e-22, 216 sessions). On the 154 sessions of the other selection, -0.667 against -0.691, p = 1.42e-17. The LSTM is ahead in all 8 mice. The optimiser reported convergence for the best of three starts in all 238 folds, and the starts reach the same likelihood in the median fold (spread 0.000, at most 480 of a median 8,262). The learning rate for unrewarded trials is at its lower bound of 0 in 198 folds and the inverse temperature at its upper bound of 50 in 50 folds.

How the test differed from the paper’s method

• RL model 4 was refit in Python (L-BFGS-B with bounds, three starting points) in place of Matlab's fminunc, leave-one-session-out over all sessions of each mouse (Methods, Reinforcement learning model). The lapse is the mixture (1 - lambda) P + lambda/2 (claims file), the bias inside the inverse temperature as printed, and visually guided trials update the model but are left out of the likelihood, as for the LSTM.
• The LSTM side uses the stored predictions (see lstm-choice-accuracy).

On average, prior to a rule switch, mice performed with little choice bias and followed the correct rule ... After a rule switch, the randomness of choices increased rapidly, as seen by near chance levels of the model's probability of the actual choice, and bias-following increased moderately. Then, rule-following recovered gradually for the new rule belief

Results, Decision-making strategy varied, paragraph 2; Figure 1I

Under both windows (10 and 20 trials) and both session selections, rule-following before the switch is 0.23 to 0.26 (p < 1e-4 against zero) with choice bias 0.13 to 0.16; the probability of the actual choice drops by 0.14 to 0.17 (p < 1e-4) with a switch-aligned minimum of 0.570 to 0.578; bias-following rises by 0.029 to 0.045 (p <= 4e-4); rule-following rises from 0.04 to 0.07 in the first window to 0.21 to 0.22 at the end of the block (p < 1e-4). Recovered switches number 320 (all-trial selection) and 251, against the paper's 265.

How the test differed from the paper’s method

• Strategy variables are those stored in the files, not refit.
• The end of the block is taken as its last W non-visually guided trials; the claims file says only 'by the end of the block'.

This variability in running trajectories increased following a rule switch

Results, Running trajectories reflected the within-trial dynamics, paragraph 1; Figure 2C

Δ trajectory correlation after minus before the switch is -0.077 ± 0.013 (11 bins) and -0.074 ± 0.012 (20 bins) with the all-trial selection (320 and 317 switches), -0.090 and -0.087 with the other (251 and 248), p < 1e-4 in each, and negative in all 8 mice. Trials with an empty position bin are dropped (1,066 to 2,556 trials).

How the test differed from the paper’s method

• Trials with an empty position bin are dropped; the Methods do not say how such trials are handled.

As expected, the latency to dynamic choice crossing a threshold was shorter on trials with higher rule- or bias-following

Results, Running trajectories reflected the within-trial dynamics, paragraph 3; Figure 2F

The maze epoch matches the legend's trial duration (9.31 ± 2.14 s, all-trial selection; 8.98 ± 1.93 s, the other; expected 8.95 ± 2.04), and trial start_time is maze onset (first frame with a position at median -0.004 s). With the all-trial selection (70,410 trials crossing threshold, paper 85,463), latency falls with rule-following (Spearman rho -0.113; highest against lowest bin p = 0.0045 with 5 bins and p < 0.0001 with 10) and with bias-following (rho -0.102; p = 0.0011 and p < 0.0001). With the other selection every test holds except bias-following with 5 bins (0.587 against 0.656, p = 0.058), so the outcome rests on the session selection; the hierarchical bootstrap with 8 mice is liberal near 0.05 (verifier), so p values near it are fragile. One mouse has a positive rho for rule-following.

How the test differed from the paper’s method

• Dynamic choice is the stored velocity LSTM output, not refit.
• Trial duration is the maze epoch, the duration that matches the legend's 8.95 s (the whole trial is 14.9 s).
• Trials whose dynamic choice never crosses threshold are left out (4,015 of 74,425).
• Sessions above 0.65 are selected on fraction correct over all trials, an assumption (Methods, Classification of the sources, gives 141 layer 2/3 sessions above 65% correct; Behavioral training says visually guided trials are excluded from all analyses). Only fraction correct over all trials gives counts near the paper's (150 layer 2/3 sessions, 97,989 neurons); with visually guided trials excluded there are 100 and 69,707. Both selections are shown.

On trials with high rule-following, choice commitment started near chance but increased rapidly during maze traversal ... when rule-following was low, choice commitment was low until late in the trial ... when mice followed a bias, choice commitment was high at trial onset

Results, Running trajectories reflected the within-trial dynamics, paragraph 3; Figure 2G; Figure S2L

In the first position bin the bias-following coefficient is 0.28 to 0.42 (p <= 0.001) and the rule-following coefficient -0.016 to 0.000 (p >= 0.36, at most 0.05 of the bias coefficient); in the last stem bin the rule-following coefficient is 0.85 to 0.90 (p < 0.0001), above its first-bin value (p < 0.0001). Holds for 11 and 20 bins, both session selections, and the local reward rate with or without the current trial.

How the test differed from the paper’s method

• Choice commitment uses the stored dynamic choice.
• Local reward rate is the fraction correct over the 9 non-visually guided trials centred on the trial (Figure S2L legend says 'adjacent 9 trials'); the 8 neighbours without the trial itself are run as a variant and give the same result.

which was corroborated with decoding of these variables from the activity of ~100 simultaneously recorded neurons

Results, Distinct encoding gradients, paragraph 2; Figure 3Q

V1 and PM come from every decoder of the 41 pool sessions with a V1 or PM field (decoders_v1_pm_offset.ipynb; 149 V1 and 71 PM decoders with k-means groups, 142 and 80 with tiles), the other four areas from the main notebook's 116 decoders of 18 sessions. Cue decoding is held-out log2 likelihood, chance -1. V1 is above PM under all 12 readings, by 0.027 to 0.062 (V1 -0.139 to -0.246, PM -0.187 to -0.282); the difference is significant only for tiles over the stem bins (p = 0.039 and 0.042), p = 0.098 to 0.474 otherwise. The highest of the other areas, AM (-0.276 to -0.382, 16 to 17 decoders in 3 mice), is 0.118 to 0.157 below V1. A is the lowest area in all 12 readings and V1 is above A by 0.355 to 0.409 (p < 0.001). So V1 is the highest area and A the lowest under every reading, with V1 above A at p < 0.05, which the criterion calls reproduced. On the subsample alone V1 was highest in only 2 of 12 readings (PM above V1 by 0.001 to 0.027 on 12 to 14 PM decoders). For k-means groups, frame folds, the minimum penalty and stem bins, the only reading printed per mouse and per field, V1 is above PM in each of the 6 mice with both areas (mouse 3 by 0.012) and within each field. Decoders labelled PM in sessions of other fields are not fitted, so PM has 71 to 80 decoders against the paper's 106; in the subsample such decoders do not average higher than the PM field's. Frame folds and trial folds give the same ranking. Proposed after round 2 and agreed in round 3.

How the test differed from the paper’s method

• The decoders were fit on a random subsample of 18 of the 142 multi-plane layer 2/3 sessions of the all-trial selection (3 per imaged field's area, seeded), 116 decoders against the paper's 974, as at the notebook's rate (488 s on 4 processes for the 18 sessions, 108 CPU-s per session) all 142 sessions take about 4.5 CPU hours, 67 minutes on 4 cores, over the 30-minute notebook limit. The 8 single-plane layer 2/3 sessions of this selection (150 - 142) were left out. A second notebook, decoders_v1_pm_offset.ipynb, fits every pool session with a field in V1 or PM (41 sessions, 23,375 neurons, 922 s on 4 processes) and the 6 pool sessions with the cue offset at 0.76.
• Decoders follow the Methods (Trial phase-specific logistic decoders): groups of about 100 neurons, PCA to more than 90% of variance, logistic regression with elastic net (10% L1) per each of 11 position bins on single frames (6 Hz volumes), 5-fold cross-validation. The solver is FISTA written here in place of Matlab's lassoglm (checked against scikit-learn's saga to 2e-5), with 10 penalties from the largest useful one down to 1e-4 of it (lassoglm's range, fewer values), a standard choice. Dynamic choice is fit as a binomial proportion.
• Readings: penalty at the minimum CV deviance and one-SE; groups by k-means and by equal-count tiles (claims file). The Methods do not say how folds are formed; lassoglm's default (random frames) and folds of whole trials are both run, the latter with k-means groups only. Areas are the label of the neuron nearest each group's centroid (97% agree with the members' majority area).
• Frames are maze traversal of non-visually guided trials; stem frames are those below each session's stem length.

decoders trained to predict cue or choice at one maze position degraded in accuracy when tested at other maze positions

Results, Flexible, high-dimensional representations, paragraph 1; Figure 8C, 8D

Decoders trained in one position bin and tested on held-out trials at every bin lose accuracy with distance for both variables in every area: cue log2 likelihood falls from -0.438 at zero distance to -1.663 at ten bins (largest minus zero distance -0.71 to -1.88 by area, p < 0.001 in each), dynamic choice Spearman correlation from 0.599 to 0.146 (-0.415 to -0.483 by area, p < 0.001 in each); every decoder drops. 116 decoders from 18 sessions, k-means groups, trial folds, minimum-CV penalty. Under the claims file's own reading of the cue test (decoders_v1_pm_offset.ipynb), the 53 decoders of the 6 pool sessions with the cue offset at 0.76 (3 mice) also lose accuracy with distance in every area: cue -0.465 to -1.086 and dynamic choice -0.450 to -0.526 from zero to the largest distance, p < 0.001 in each, every decoder dropping; V1 (2 decoders), AM (5) and A (11) each come from one mouse. The outcome therefore does not depend on the assumption.

How the test differed from the paper’s method

• The decoders were fit on a random subsample of 18 of the 142 multi-plane layer 2/3 sessions of the all-trial selection (3 per imaged field's area, seeded), 116 decoders against the paper's 974, as at the notebook's rate (488 s on 4 processes for the 18 sessions, 108 CPU-s per session) all 142 sessions take about 4.5 CPU hours, 67 minutes on 4 cores, over the 30-minute notebook limit. The 8 single-plane layer 2/3 sessions of this selection (150 - 142) were left out. A second notebook, decoders_v1_pm_offset.ipynb, fits every pool session with a field in V1 or PM (41 sessions, 23,375 neurons, 922 s on 4 processes) and the 6 pool sessions with the cue offset at 0.76.
• Decoders follow the Methods (Trial phase-specific logistic decoders): groups of about 100 neurons, PCA to more than 90% of variance, logistic regression with elastic net (10% L1) per each of 11 position bins on single frames (6 Hz volumes), 5-fold cross-validation. The solver is FISTA written here in place of Matlab's lassoglm (checked against scikit-learn's saga to 2e-5), with 10 penalties from the largest useful one down to 1e-4 of it (lassoglm's range, fewer values), a standard choice. Dynamic choice is fit as a binomial proportion.
• Readings: penalty at the minimum CV deviance and one-SE; groups by k-means and by equal-count tiles (claims file). The Methods do not say how folds are formed; lassoglm's default (random frames) and folds of whole trials are both run, the latter with k-means groups only. Areas are the label of the neuron nearest each group's centroid (97% agree with the members' majority area).
• Frames are maze traversal of non-visually guided trials; stem frames are those below each session's stem length.
• Assumed. The Figure 8C legend uses only sessions with the cue offset at 0.76 of the maze. The files' cue delay length gives that offset for 6 of the 142 sessions in the pool and none of the 18 decoder sessions (offsets 0.96 and 1.0), and none with a field in V1 or AM (findings), so the cue test uses every decoder session, with the cue-present bins set by each session's own offset. The 6 sessions with the 0.76 offset are decoded in the second notebook and give the same result.
• Cross-position tests use folds of whole trials, so that a test frame never comes from a training trial.

this behavior was variable across mice and sessions and was uncorrelated with task performance and the time course of dynamic choice

Results, Running trajectories reflected the within-trial dynamics, paragraph 2; Figure S2H, S2I, S2K

Cue-biased running varies across mice (mean cue log likelihood by mouse -0.92 to -0.36) and in direction across and within mice (roll coefficient positive in 79 of 238 sessions; all 30 negative in mouse 5; 27 positive and 11 negative in mouse 10). Per-mouse slopes of performance on cue decoding are not different from zero (mean 0.024, bootstrap p = 0.35; other selection -0.048, p = 0.35), but the pooled correlation over sessions is positive (r = 0.224, p = 0.0005; other selection r = 0.207, p = 0.010). The criterion's two readings disagree, so the claim holds in part. The Figure S2G part was reported, not judged; choice decoding rises with cue decoding in the third quadrant of the maze, from -0.675 to -0.509 over the 7 bins of cue decoding; in the second it rises from -0.817 to -0.726 over the first four bins and stays at -0.768 to -0.776 in the last three.

How the test differed from the paper’s method

• Cue decoding uses the stored cue-biased running. The pooled-correlation and per-mouse readings are the claims file's; the per-mouse test is a bootstrap over mice (the Methods' default), with a t-test shown.
• The logistic decoder is fit with scikit-learn (unpenalised in effect, C = 1e6) on z-scored velocities in place of Matlab's fitglm.

RSC was also most active in the maze stem and least active during feedback and ITI periods, when area A was more active

Results, Encoding of dynamic choice was enriched in RSC, paragraph 4; Figure 4H

Two parts of the claim hold and one does not. In the feedback period/ITI RSC is the least active area (-0.042) and A the most (0.023), A above RSC (p < 0.001) in both selections and in all 8 mice; within RSC the feedback period/ITI is the lowest epoch (p < 0.001). RSC is not shown to be most active in the stem: within RSC the arm mean exceeds the stem mean (0.042 against 0.023, p = 0.07; other selection p = 0.18; arm higher in 7 of 8 mice), and between areas RSC's stem mean is the highest and above V1, PM, MM and A (p <= 0.016) but not AM (p = 0.29); with the other selection only PM and A differ. The criterion asks for every comparison under every reading, and the well-fit reading was not run. The verifier would call this not reproduced; it is kept as partially reproduced because two of the claim's three parts hold (see verifications).

How the test differed from the paper’s method

• Only the all-neuron reading was run; the well-fit reading of the claims file needs the GLM fits (blocked).
• Epochs are taken from the frame-aligned position and time from the choice point in the files: stem below the stem length, arm from there to the choice point (including 913,421 frames just past the 2.5 m maze end), feedback/ITI after the choice point. Frames of the 1 s feedback delay have no position (0 of 3,673,732), so they fall in feedback/ITI.
• Sessions above 0.65 are selected on fraction correct over all trials, an assumption (Methods, Classification of the sources, gives 141 layer 2/3 sessions above 65% correct; Behavioral training says visually guided trials are excluded from all analyses). Only fraction correct over all trials gives counts near the paper's (150 layer 2/3 sessions, 97,989 neurons); with visually guided trials excluded there are 100 and 69,707. Both selections are shown.
• In the single-plane sessions (13 in the dandiset; 8 of the 150 layer 2/3 sessions of the all-trial selection) the deconvolved activity timestamps run 2000 times the behaviour clock (findings); their frames are matched to behaviour by index, one frame per behaviour frame.

Partial correlations were greater for dynamic choice than reported choice, revealing that even decoders trained to predict reported choice were more closely related to dynamic choice

Results, Encoding of dynamic choice was enriched in RSC, paragraph 2; Figure 4D

What the paper reports

• difference of the two partial correlations against zero: bootstrap, p < 0.001 (n = 974 decoders)

Under every reading the reported-choice decoders' output is more related to dynamic choice than to reported choice (difference 0.217 to 0.326, p < 0.001, positive in 97% to 100% of decoders). The partial correlation with dynamic choice is 0.428 to 0.444, above 0.35 ± 0.02 under every reading; with reported choice 0.101 to 0.217, within 0.14 ± 0.02 only for trial folds at the minimum-CV penalty (0.130); the difference is within 0.21 ± 0.02 for three of the four frame-fold readings. No reading has all three within tolerance, so only the sign and test hold. Measured values are k-means groups, frame folds, minimum-CV penalty.

How the test differed from the paper’s method

• The decoders were fit on a random subsample of 18 of the 142 multi-plane layer 2/3 sessions of the all-trial selection (3 per imaged field's area, seeded), 116 decoders against the paper's 974, as at the notebook's rate (488 s on 4 processes for the 18 sessions, 108 CPU-s per session) all 142 sessions take about 4.5 CPU hours, 67 minutes on 4 cores, over the 30-minute notebook limit. The 8 single-plane layer 2/3 sessions of this selection (150 - 142) were left out. A second notebook, decoders_v1_pm_offset.ipynb, fits every pool session with a field in V1 or PM (41 sessions, 23,375 neurons, 922 s on 4 processes) and the 6 pool sessions with the cue offset at 0.76.
• Decoders follow the Methods (Trial phase-specific logistic decoders): groups of about 100 neurons, PCA to more than 90% of variance, logistic regression with elastic net (10% L1) per each of 11 position bins on single frames (6 Hz volumes), 5-fold cross-validation. The solver is FISTA written here in place of Matlab's lassoglm (checked against scikit-learn's saga to 2e-5), with 10 penalties from the largest useful one down to 1e-4 of it (lassoglm's range, fewer values), a standard choice. Dynamic choice is fit as a binomial proportion.
• Readings: penalty at the minimum CV deviance and one-SE; groups by k-means and by equal-count tiles (claims file). The Methods do not say how folds are formed; lassoglm's default (random frames) and folds of whole trials are both run, the latter with k-means groups only. Areas are the label of the neuron nearest each group's centroid (97% agree with the members' majority area).
• Frames are maze traversal of non-visually guided trials; stem frames are those below each session's stem length.

Decoding of both dynamic and reported choice was above chance (p < 10-3, bootstrap), but decoding of reported choice was less accurate than dynamic choice

Figure S5F

What the paper reports

• each decoder type against chance: bootstrap, p < 0.001 (n = 974 decoders)
• dynamic minus reported choice decoding: bootstrap, p < 0.001 (n = 974 decoders)

Under every reading both decoders are above chance (p < 0.001) and dynamic choice is decoded better than reported choice (difference 0.238 to 0.297, p < 0.001, within 0.33 ± 0.13). Both values are above the paper's: dynamic choice 0.636 to 0.683 against 0.59 ± 0.02 and reported choice 0.343 to 0.444 against 0.26 ± 0.02, under no reading within tolerance. Trial folds give lower reported-choice values (0.343 to 0.373) than frame folds (0.431 to 0.444). Measured values are k-means groups, frame folds, minimum-CV penalty.

How the test differed from the paper’s method

• The decoders were fit on a random subsample of 18 of the 142 multi-plane layer 2/3 sessions of the all-trial selection (3 per imaged field's area, seeded), 116 decoders against the paper's 974, as at the notebook's rate (488 s on 4 processes for the 18 sessions, 108 CPU-s per session) all 142 sessions take about 4.5 CPU hours, 67 minutes on 4 cores, over the 30-minute notebook limit. The 8 single-plane layer 2/3 sessions of this selection (150 - 142) were left out. A second notebook, decoders_v1_pm_offset.ipynb, fits every pool session with a field in V1 or PM (41 sessions, 23,375 neurons, 922 s on 4 processes) and the 6 pool sessions with the cue offset at 0.76.
• Decoders follow the Methods (Trial phase-specific logistic decoders): groups of about 100 neurons, PCA to more than 90% of variance, logistic regression with elastic net (10% L1) per each of 11 position bins on single frames (6 Hz volumes), 5-fold cross-validation. The solver is FISTA written here in place of Matlab's lassoglm (checked against scikit-learn's saga to 2e-5), with 10 penalties from the largest useful one down to 1e-4 of it (lassoglm's range, fewer values), a standard choice. Dynamic choice is fit as a binomial proportion.
• Readings: penalty at the minimum CV deviance and one-SE; groups by k-means and by equal-count tiles (claims file). The Methods do not say how folds are formed; lassoglm's default (random frames) and folds of whole trials are both run, the latter with k-means groups only. Areas are the label of the neuron nearest each group's centroid (97% agree with the members' majority area).
• Frames are maze traversal of non-visually guided trials; stem frames are those below each session's stem length.

The rule update per trial was 0.013 ± 0.0055 for visually guided trials and 0.0080 ± 0.0023 for normal trials (difference = 0.0052 ± 0.0057, p = 0.34, hierarchical bootstrap mean ± SEM, n = 265 switches)

STAR Methods, Decision-making strategy variables, last paragraph

What the paper reports

• visually guided against normal trials, all trials: hierarchical bootstrap, p = 0.34 (n = 265 switches)
• visually guided against normal trials, correct trials only: hierarchical bootstrap, p = 0.89 (n = 265 switches)

With the update of trial t taken as rule belief on t + 1 minus that on t (the adopted reading), visually guided trials update more than normal trials over the 30 trials after a switch, 0.0545 ± 0.0075 against 0.0215 ± 0.0029, difference 0.0329, p < 0.0001 (302 switches; the other selection gives 0.0591 against 0.0236, p = 0.0022), so the paper's test finds the effect the claim says is absent. On correct trials only there is no difference (0.068 against 0.076, p = 0.22; paper p = 0.89). The change from t - 1 to t gives no difference (0.0295 against 0.0267, p = 0.48), but it is not the adopted reading and is not taken as support. Correct trials move belief toward the new rule (0.0765 normal, 0.0690 visually guided) and incorrect trials away from it (-0.0243, -0.0129); 80% of visually guided trials are correct against 46% of normal trials, which accounts for the forward difference. Under every all-trial reading both values are 2 to 5 times the paper's 0.013 and 0.0080, outside the tolerances (the correct-only rows, which have no paper values, give 0.076 to 0.080 for normal trials); the stored switch-aligned rule belief itself rises by 0.028 per trial over the 30 trials (from -0.619 to 0.195). The question of how the update was defined is put to the paper's authors (findings).

How the test differed from the paper’s method

• Assumed. The Methods (Decision-making strategy variables) do not say which difference of rule belief is a trial's update. The LSTM's belief on trial t is conditioned on trials 1 to t - 1 (Methods, LSTM model), so the change from t to t + 1 is the one trial t causes; this reading is adopted. The change from t - 1 to t is shown.
• Rule belief is recomputed from the stored association matrix with the Methods' sign, as the stored column has the opposite sign (findings).
• Sessions above 0.65 are selected on fraction correct over all trials, an assumption (Methods, Classification of the sources, gives 141 layer 2/3 sessions above 65% correct; Behavioral training says visually guided trials are excluded from all analyses). Only fraction correct over all trials gives counts near the paper's (150 layer 2/3 sessions, 97,989 neurons); with visually guided trials excluded there are 100 and 69,707. Both selections are shown.

Inhibiting either RSC or PPC led to markedly lower task performance compared to control trials or trials in which part of primary somatosensory cortex (S1) was inhibited

Results, Photoinhibition of posterior cortical areas, paragraph 1; Figure 1L

What the paper reports

• control against RSC or PPC: hierarchical bootstrap, p < 10⁻⁴ (n = 164 sessions from 7 mice)
• control against S1: hierarchical bootstrap, p = 2 × 10⁻⁴ (n = 164 sessions from 7 mice)
• S1 against RSC or PPC: hierarchical bootstrap, p < 10⁻⁴ (n = 164 sessions from 7 mice)
• RSC against PPC: hierarchical bootstrap, p = 0.058 (n = 164 sessions from 7 mice)

Needs task performance under photoinhibition. The dandiset's 308 files are 300 imaging sessions and 8 widefield retinotopy files; none is a photoinhibition session of the 7 VGAT-ChR2 mice, no path names inhibition, and no trial table has an inhibition column. The paper names no other archive for these data.

PPC and RSC inhibition decreased rule-following compared to control trials and S1 inhibition

Results, Photoinhibition of posterior cortical areas, paragraph 2; Figure 1M; Figure S1K

What the paper reports

• rule-following, each target against control: hierarchical bootstrap, p < 10⁻⁴ (n = 164 sessions from 7 mice)
• rule-following, RSC and PPC against S1: hierarchical bootstrap, p < 10⁻⁴ (n = 164 sessions from 7 mice)
• bias-following, S1 against control: hierarchical bootstrap, p = 0.025 (n = 164 sessions from 7 mice)
• bias-following, RSC against control: hierarchical bootstrap, p = 0.018 (n = 164 sessions from 7 mice)
• bias-following, PPC against control: hierarchical bootstrap, p < 10⁻⁴ (n = 164 sessions from 7 mice)

Needs strategy variables under photoinhibition. The dandiset's 308 files are 300 imaging sessions and 8 widefield retinotopy files; none is a photoinhibition session of the 7 VGAT-ChR2 mice, no path names inhibition, and no trial table has an inhibition column. The paper names no other archive for these data.

Pearson correlation 0.18 (p = 0.10 vs. zero) for RSC and -0.03 (p = 0.74) for PPC

Figure S1M

What the paper reports

• RSC: Pearson correlation against zero, r = 0.18, p = 0.1 (n = 164 sessions from 7 mice)
• PPC: Pearson correlation against zero, r = -0.03, p = 0.74 (n = 164 sessions from 7 mice)

Needs per-session changes under photoinhibition. The dandiset's 308 files are 300 imaging sessions and 8 widefield retinotopy files; none is a photoinhibition session of the 7 VGAT-ChR2 mice, no path names inhibition, and no trial table has an inhibition column. The paper names no other archive for these data.

The GLM explained a substantial amount of moment-to-moment variability in a neuron's activity (fraction of Poisson deviance explained: 0.25 ± 0.17, mean ± SD

Results, Calcium imaging of neural activity, paragraph 2; Figure 3I; Figure S3E

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

42,998 neurons were selected, including 7,775 neurons in V1, 4,934 neurons in PM, 6,878 neurons in AM, 3,991 neurons in MM, 10,329 neurons in RSC, and 9,091 neurons in area A

STAR Methods, Generalized Linear Models, Fitting procedure; Figure S3D

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

We confirmed that the difference in the fraction explained deviance evaluated on the CV held-out data and test data was sufficiently small (0.0010 ± 0.0006, mean ± SEM

STAR Methods, Quantification of fraction explained deviance (without re-fitting); Figure S3E

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

the encoding of task variables exhibited an anterior-posterior gradient with highest strength in V1, intermediate in RSC, and lowest in area A, whereas movement encoding strength exhibited a gradient in the opposite direction

Results, Distinct encoding gradients, paragraph 1; Figure 3M

What the paper reports

• A against every other area, movement (higher) and task variables (lower): hierarchical bootstrap, p < 0.001 (n = 42,998 well-fit neurons from 8 mice)
• V1 against every other area, task variables: hierarchical bootstrap, p < 0.001 (n = 42,998 well-fit neurons from 8 mice)

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

Encoding of the visual cue identity was strongest in V1 and neighboring areas AM and PM, weaker in RSC, and weakest in area A

Results, Distinct encoding gradients, paragraph 2; Figure 3O; Figure S4K

What the paper reports

• V1 higher than every other area: hierarchical bootstrap, p < 0.001
• A lower than every other area: hierarchical bootstrap, p < 0.001

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

Variables related to the decision-making strategy, including their interactions with other task variables, collectively exhibited a moderate encoding strength distributed evenly across posterior cortex

Results, Distinct encoding gradients, paragraph 3; Figure 3R; Figure S4L

What the paper reports

• all area pairs except AM against A: hierarchical bootstrap, p > 0.03, Benjamini-Hochberg (not significant after correction)-corrected
• AM against A: hierarchical bootstrap, p < 0.001

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

The previous trial's outcome was much more strongly encoded than the previous trial's cue or choice

Results, Distinct encoding gradients, paragraph 3; Figure S4P to S4S

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

during traversal of the maze stem, encoding of dynamic choice was distributed across all areas but enriched in RSC and adjacent medial areas, accounting for a substantial amount of neural activity compared to near-zero encoding of reported choice

Results, Encoding of dynamic choice was enriched in RSC, paragraph 1; Figure 4A, 4B

What the paper reports

• dynamic against reported choice, each area: hierarchical bootstrap, p < 0.001
• dynamic choice, RSC against V1, PM or A: hierarchical bootstrap, p < 0.001
• dynamic choice, RSC against AM or MM: hierarchical bootstrap, p > 0.05
• reported choice, each area against zero: hierarchical bootstrap, p > 0.05

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

In contrast, during the feedback and ITI periods, we observed greater encoding of reported choice than dynamic choice

Results, Encoding of dynamic choice was enriched in RSC, paragraph 1; Figure 4C

What the paper reports

• reported against dynamic choice, each area: hierarchical bootstrap, p < 0.001
• reported choice, RSC against each other area: hierarchical bootstrap, p < 0.05

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

For all areas except V1 and area A, the mean fraction explained deviance for dynamic choice was greater than roll velocity, which controls left-right movements in the maze (AM: p = 0.006; PM, MM and RSC: p < 10-3)

Figure S5I

What the paper reports

• AM: hierarchical bootstrap, p = 0.006
• PM, MM and RSC: hierarchical bootstrap, p < 0.001

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

Encoding of dynamic choice in models with movement history was significantly greater than zero in all 6 areas (p < 10-3), and was higher in RSC compared to V1, PM and area A (p < 10-3)

Results, Encoding of dynamic choice was enriched in RSC, paragraph 3; Figure S5M

What the paper reports

• with against without movement history, each area: hierarchical bootstrap, p < 0.001
• each area against zero: hierarchical bootstrap, p < 0.001
• RSC against V1, PM and A: hierarchical bootstrap, p < 0.001

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

Encoding of maze position was present in all areas, but strongest in RSC, with a similar medial-to-lateral gradient as choice encoding

Results, Encoding of dynamic choice was enriched in RSC, paragraph 4; Figure 4F, 4G

What the paper reports

• RSC against V1, PM or A: hierarchical bootstrap, p < 0.001
• RSC against AM or MM: hierarchical bootstrap, p > 0.05

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

Cue and movement were the two variables with the highest mutual information between encoding strength and posterior cortical area. However, their equivalent random fractions were greater than 0.75 ... Similarly, their equivalent jitters were near 0.6

Results, Encoding of most variables is highly distributed, paragraph 2; Figure 5E, 5F; Figure S6A

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

The encoding of maze position and choice were even more distributed than cue and movement, with equivalent random fractions near 0.9. All other variables were also highly distributed, with decision-making strategy being the most distributed (random fraction near one)

Results, Encoding of most variables is highly distributed, paragraph 2; Figure 5E

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

Neurons in V1, A, and RSC were typically predicted to reside in their actual areas of origin, indicating that these neurons had distinctive encoding profiles, in contrast to neurons from AM and MM that were predicted to reside in all sites across posterior cortex

Results, Single-neuron encoding profiles confirm functional gradients, paragraph 1; Figure 6A, 6B

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

Strikingly, these factors formed three distinct spatial gradients that dominated the anatomical organization of encoding with peaks in V1, RSC and A, respectively

Results, Single-neuron encoding profiles confirm functional gradients, paragraph 2; Figure 6D; Figure S6I; STAR Methods, Non-negative matrix factorization

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

cue was greatest in factor 1, maze position and choice were greatest in factor 2, and movement was greatest in factor 3

Results, Single-neuron encoding profiles confirm functional gradients, paragraph 2; Figure S6G

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

The number of principal components needed to explain 70% of the variance in the three matrices was 8, 6, and 3, respectively, and further, 35, 23, and 16 principal components were needed to explain 95% of the variance

STAR Methods, Analysis of dimensionality of encoding across neurons versus encoding across cortical space

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

Strikingly, the correlations in encoding between pairs of cue, choice, and movement variables were similar across areas, despite major differences in the encoding strengths of these variables between areas ... In addition, the correlation coefficients were close to zero

Results, Conjunctive structure of encoding, paragraph 2; Figure 7D to 7F

What the paper reports

• cue and dynamic choice, all area pairs: bootstrap, p > 0.05
• cue and movement, A against each other area: bootstrap, p < 0.013
• cue and movement, pairs among the other 5 areas: bootstrap, p > 0.05
• dynamic choice and movement, all area pairs: bootstrap, p > 0.05

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

Although decoding based on encoding correlations allowed above-chance performance, it was poor relative to decoding from encoding strengths alone. Further, decoding with both encoding correlations and encoding strengths was barely improved from decoding with encoding strengths alone

Results, Conjunctive structure of encoding, paragraph 3; Figure 7G, 7H

What the paper reports

• strengths only against correlations only, all areas except AM: Wilcoxon signed-rank test, p < 0.05
• strengths only against strengths plus correlations, all areas: Wilcoxon signed-rank test, p > 0.03, not significant after multiple comparison correction-corrected

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

The performance of the models on test data was worse on average than those with group lasso (difference in fraction of explained deviance

STAR Methods, Quantification of fraction explained deviance (without re-fitting); Figure S8D to S8J

What the paper reports

• hierarchical bootstrap, p < 0.001 (n = 93,881 neurons)

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

for subsequent analyses, we focused on well fit neurons although results were robust to this criterion

Results, Calcium imaging of neural activity, paragraph 2; Figure S8K to S8O

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

Although the centroids of the distributions (i.e. most typical encoding profiles) were offset from one another, many neurons were functionally closer to the centroid of a different area than to the centroid of the area where they resided

Results, Single-neuron encoding profiles confirm functional gradients, paragraph 3; Figure 6H; Figure S6L

Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

For single neuron encoding magnitudes, dynamic choice was greater than cue-biased running (p < 10-3) for all areas except V1 and PM. For population decoding performance, dynamic choice was greater than cue-biased running (p < 10-3) for all areas

Figure S5L; Results, Encoding of dynamic choice was enriched in RSC, paragraph 3

What the paper reports

• single-neuron encoding, PM excluded with V1, the other four areas: hierarchical bootstrap, p < 0.001
• population decoding, all six areas: hierarchical bootstrap, p < 0.001 (n = 998 decoders)

Depends on well-fit-neurons, which needs the GLM (blocked, too large), and so is blocked with it, as a claim that depends on a blocked claim. Its decoder half could be run with the decoders of fig3q-cue-decoding but the criterion judges the single-neuron and decoder halves together. Needs the Poisson GLM of the Methods fit to every layer 2/3 neuron (5-fold cross-validation over 21 penalties plus a final fit, 106 fits per neuron). The dandiset holds no fitted coefficients, deviances or encoding strengths. The notebook times one L2 Poisson fit of the median session's training frames by 1,310 predictors at 0.2 s, which puts the 97,989 neurons at 652 CPU hours, a random subsample of 2,000 neurons at 13.3 CPU hours (3.3 h on the run's 4 cores) and one of 200 at 1.3 CPU hours (0.3 h), each a lower bound since the Methods' group lasso is slower. The fit time varies between runs on the same machine type: the round-2 run printed 0.4 s per fit (1,191, 24.3 and 2.4 CPU hours). Even the 200-neuron subsample, 20 to 36 minutes of fitting by these lower bounds, leaves no room in a 30-minute notebook for streaming the sessions and building the design, and 200 neurons give about 33 per area. The design matrix (27 task variables on raised-cosine bases, 336 movement predictors, drift terms) and the predictor grouping are not in the paper's public code. Not run.

Shattering dimensionality for all conjunctions in all areas was well above chance, and interestingly, shattering dimensionality for each conjunction was similar across posterior cortex ... Almost all differences between areas were not statistically significant and fell within a range of < 10% classification accuracy for each conjunction

Results, Flexible, high-dimensional representations, paragraph 2; Figure 8F, 8G; Figure S8A to S8C

What the paper reports

• all pairs among all-area and the 6 areas, Figure 8G: hierarchical bootstrap, p > 0.01, not significant after multiple comparison correction-corrected
• cue, dynamic choice and position, RSC against A (Figure S8B): hierarchical bootstrap, p < 0.001

Needs pseudo-populations of 1,000 neurons per area resampled 1,000 times, with linear SVCs on every marginally balanced dichotomy of each conjunction, 10 splits each, on stem frames of all layer 2/3 neurons, under two readings of choice. The notebook counts 126 marginally balanced dichotomies for cue by 10 positions alone and times one SVC on 20 conditions by 1,000 neurons at 1.04 ms, which puts that one conjunction at 17,640,000 fits and 5.1 CPU hours (1.3 h on 4 cores; earlier runs printed 0.69 and 0.81 ms, 3.4 and 3.9 CPU hours), a lower bound without building the trial samples and without the other conjunctions of Figure 8G. Not run.