A Neural Population Mechanism for Rapid Learning (DANDI:000688)

Claims

From the paper A Neural Population Mechanism for Rapid Learning

Reproducing Perich et al. (2018) from dandiset 000688 5

Notebook reproduce_perich_2018.ipynb

Reproduced2

ReproducedFigure S4D legend; STAR Methods, Dimensionality reduction, paragraph 2rests on an assumption
With random subsamples of equal, fixed numbers of neurons, estimated dimensionality fell for small samples but PMd stayed higher-dimensional than M1 for every neuron count tested.
With decreasing neuron counts, we began to underestimate the apparent dimensionality. However, critically, PMd (red) was consistently higher dimensional than M1 (black) for all tested neuron counts
Figure S4D legend; STAR Methods, Dimensionality reduction, paragraph 2
Under rule (b) the mean PMd dimensionality exceeds the mean M1 at 40, 50, 60, 70 and all neurons under both kernels, both windows and both session sets, by 1.30 to 4.63 (full-width kernel, window a, 1.88 at 40 and 4.52 with all neurons; standard-deviation kernel, window a, 1.62 and 3.48). Paired by session, PMd is higher in every session at every size except one session with all neurons in window (b) under both kernels (20 of 21) and one at 40 neurons with the full-width kernel in window (b) (16 of 17). Rule (a) gives PMd above M1 at every size except 40 neurons with the full-width kernel in window (b) (-0.13 and -0.30). The PMd mean falls at small sizes (window a, full-width kernel, 13.05 with all neurons and 9.90 at 40). The M1 mean does not fall consistently (window a, full-width kernel, 8.52 with all, 8.53 and 8.65 at 60 and 70, 8.03 at 40; window b, full-width kernel, 6.52 with all and 6.74 at 40), so the statement's fall is shown for PMd only, outside the criterion. The M1 means at 40 neurons cover 17 sessions, against 21 for all neurons. M1 sizes of 50 to 70 rest on 11 to 15 sessions, as Monkey M has 26 to 52 M1 units.
How the test differed from the paper’s method
- The paper's 16 sessions cannot be identified. All 21 dual-array center-out sessions of Monkeys C (2016-09-09 to 2016-10-21) and M (2014-02-03 to 2014-03-07) are used, and the 19 whose reward counts fit the paper's Baseline range (154 to 225) are shown as a second set. All reward trials of each file are taken as Baseline.
- Assumed that the noise eigenvalues are scaled to the noise of a trial average (the difference of two trials divided by the square root of 2K, K the mean number of trials per target). STAR Methods, Dimensionality reduction, says only that the activity of random pairs of trials was subtracted; taken as it is, or divided by the square root of two, the noise exceeds every component of the trial-averaged data and every estimate is 0, which cannot give the paper's medians. All three scalings are shown.
- 100 subsamples per size (the claims file's standard choice) with 100 noise repetitions per subsample, to fit the time limit; the paper's 1000 noise repetitions are kept for all neurons.
- Run with both kernel readings (standard deviation and full width at half maximum of 100 ms) in windows (a) and (b).
ReproducedFigure S4C legend; STAR Methods, Output-potent and output-null subspaces, paragraph 3rests on an assumption
Single neurons contributed to both output-potent and output-null subspaces; the null/potent index was centered on zero with virtually no neurons near -1 or 1.
The distribution of neurons was centered around zero and virtually no neurons had a null/potent index close to -1 / 1, indicating that the output-potent and output-null subspaces captured population-wide features
Figure S4C legend; STAR Methods, Output-potent and output-null subspaces, paragraph 3
Over 3263 PMd units of 21 sessions, with absolute values, the mean index is 0.017 to 0.056 across the window and kernel readings and no unit exceeds 0.9 in absolute value (range -0.701 to 0.741). With sums of squares three of four readings meet the criterion; window (a) with the full-width kernel gives a mean of 0.100085, just above 0.1, with 0.09% of units above 0.9. The shuffled control has mean 0.036, standard deviation 0.483 (data 0.217) and 3.2% of units above 0.9. The index leans positive in Monkey M, whose 599 units have a mean of 0.109 under the main reading (above 0.1 if judged alone; session means 0.081 to 0.143), against 0.028 for the 2664 Monkey C units (session means -0.007 to 0.081).
How the test differed from the paper’s method
- The paper's 16 sessions cannot be identified. All 21 dual-array center-out sessions of Monkeys C (2016-09-09 to 2016-10-21) and M (2014-02-03 to 2014-03-07) are used, and the 19 whose reward counts fit the paper's Baseline range (154 to 225) are shown as a second set. All reward trials of each file are taken as Baseline.
- Target acquisition (window a, the claims file's O3) is the trial's stop time minus the 0.5 s target hold, from STAR Methods, Behavioral task ("required to hold there for 0.5 s"); the cursor cannot be used, as in several sessions it never enters the stored target square.
- The readings of magnitude disagree in one of eight cases; the sum of absolute values, the claims file's first reading of "the magnitude ... summed across all dimensions" (STAR Methods, Output-potent and output-null subspaces), is adopted.
- The one failing reading (sums of squares, window a) uses the full-width kernel. Under the standard-deviation kernel, the reading adopted for every claim, all four window and magnitude readings meet the criterion.

Partially reproduced3

Partially reproducedSTAR Methods, Neural recordings, paragraph 2rests on an assumption
Per session, 137 to 256 PMd and 55 to 93 M1 neurons were isolated in Monkey C and 66 to 121 PMd and 26 to 51 M1 neurons in Monkey M; neurons above 5 Hz, pooled over all CF and VR sessions, number 918 in M1 and 2221 in PMd.
Across all sessions, we isolated between 137 - 256 PMd and 55 - 93 M1 neurons for Monkey C, and 66 - 121 PMd and 26 - 51 M1 neurons for Monkey M ... Pooled across all monkeys and CF and VR sessions, this gave a population of 918 M1 and 2221 PMd neurons
STAR Methods, Neural recordings, paragraph 2
Measured values are over all 21 candidate sessions. Monkey M's 7 sessions give 66 to 121 PMd and 26 to 52 M1 units, inside the widened ranges. 13 of 14 Monkey C candidates are inside (PMd 137 to 258, M1 55 to 95); C 2016-09-29 has 114 PMd units, below the lower bound of 123.3. The ranges hold for 715 of the 2002 choices of 9 Monkey C sessions, those without C 2016-09-29, which is the assumption the outcome rests on. The pooled counts above 5 Hz hold under no choice of 9 Monkey C sessions with the 7 Monkey M sessions (at most 634 M1 and 1224 PMd against 918 and 2221); all 21 sessions give 835 M1 (within 10%) and 1565 PMd (outside). Rates over all trials of a file change the counts by at most 10. Monkey M has 6 to 21 PMd units above 5 Hz per session. The pooled verdicts of the first table in the notebook are for each whole set (21 or 19 sessions); the test on 16 sessions is the second cell.
How the test differed from the paper’s method
- The paper's 16 sessions cannot be identified. All 21 dual-array center-out sessions of Monkeys C (2016-09-09 to 2016-10-21) and M (2014-02-03 to 2014-03-07) are used, and the 19 whose reward counts fit the paper's Baseline range (154 to 225) are shown as a second set. All reward trials of each file are taken as Baseline.
- The 5 Hz rule uses the trials in the file (Baseline only), from start to stop of each reward trial; all trials are shown as a variant. The paper's rate is over the whole session (claims file criterion).
- The paper excluded cells with high coincidence across channels; the units are counted as deposited.
Partially reproducedResults, Models to study adaptation-related changes ..., paragraph 3; Figure 3Brests on an assumption
In every session the PMd manifold had more dimensions than the M1 manifold (median 14 against 8.5).
In all sessions, the PMd manifold contained more dimensions than did the M1 manifold ... M1 median, 8.5; PMd median, 14
Results, Models to study adaptation-related changes ..., paragraph 3; Figure 3B
Over all 21 sessions, under rule (b), window (a), the noise of the trial average and a kernel of 100 ms standard deviation, the medians are 7 (M1) and 10 (PMd), outside the tolerance of 8.5 and 14, and PMd exceeds M1 in 21 of 21 sessions (19 of 19 in the second set, medians 7 and 11). Over the 2002 choices of 16 sessions (9 of the 14 Monkey C candidates with the 7 Monkey M sessions) the medians are 7 to 7.5 and 10 to 11, outside tolerance in every choice, with PMd above M1 in every session in all 2002; the closest session is M 2014-03-03 (PMd 9, M1 8). By monkey C has medians 7 and 10, M 7 and 11. Under this kernel no reading meets the criterion in full; window (b) gives PMd above M1 in 20 of 21, rule (a) in 19 and 18 of 21 with medians of 26 to 47 (M1) and 46 to 78 (PMd). With a kernel of 100 ms full width at half maximum the medians are 9 and 13 over 21 sessions, within tolerance in 910 of the 2002 choices, which would be reproduced. On synthetic rank-5 data rule (b) with the adopted noise scale returns 5 in 9 of 9 runs and rule (a) 16 to 45. Raw or root-two noise gives 0 dimensions in every session. The outcome was reproduced (full-width reading) through round 2 and changed in round 3, when the verifier agreed with the proposed change.
How the test differed from the paper’s method
- The paper's 16 sessions cannot be identified. All 21 dual-array center-out sessions of Monkeys C (2016-09-09 to 2016-10-21) and M (2014-02-03 to 2014-03-07) are used, and the 19 whose reward counts fit the paper's Baseline range (154 to 225) are shown as a second set, with every choice of 16 sessions. All reward trials of each file are taken as Baseline.
- Target acquisition (window a, the claims file's O3) is the trial's stop time minus the 0.5 s target hold, from STAR Methods, Behavioral task ("required to hold there for 0.5 s"); the cursor cannot be used, as in several sessions it never enters the stored target square.
- Assumed that the noise eigenvalues are scaled to the noise of a trial average (the difference of two trials divided by the square root of 2K, K the mean number of trials per target). STAR Methods, Dimensionality reduction, says only that the activity of random pairs of trials was subtracted; taken as it is, or divided by the square root of two, the noise exceeds every component of the trial-averaged data and every estimate is 0, which cannot give the paper's medians. All three scalings are shown. The choice was made after the raw and root-two scales gave 0; a synthetic check of known rank (notebook) shows rule (b) with this scale recovers the rank, but the raw and root-two scales also recover it there, so the check does not choose between the scales; the choice rests on the other scales giving 0 on the real data.
- The outcome depends on the kernel width, which the claims file expected not to matter (O4). Assumed a standard deviation of 100 ms for every claim, as the usual way a Gaussian smoothing kernel's width is stated and the way the same group states it in a later paper on Monkey C recordings ("a Gaussian kernel (s.d. 50 ms)", Gallego, Perich et al. 2020, Nat. Neurosci.). STAR Methods, Dimensionality reduction, says only "a non-causal Gaussian kernel of width 100 ms". With a full width at half maximum of 100 ms the medians are within tolerance and the claim would be reproduced, while fig6c-null-leads-potent would be not reproduced; neither reading favours every claim. The claims file's header says a claim whose open-detail readings disagree is blocked as methods_underspecified; the protocol (author, step 1.6) instead lets the author adopt the best-supported reading as an assumption, and that rule is followed.
- Trials of unequal length are aligned on the window start and cut to the session's shortest window before averaging per target; the paper does not say how they were averaged (a standard choice).
- All units of each area are used; the paper does not say whether a rate threshold applies here.
Partially reproducedFigure 6B and 6C (legend's first "d"); Results, Evolving changes in PMd output-null planning activity ..., paragraph 1rests on an assumption
In Baseline, output-null activity preceded output-potent activity by 35 ms on average across the 16 sessions, a lag significantly below zero (t-test, p = 0.02), and output-null also led at the end of Force.
Output-null activity preceded output-potent by 35 ms on average, and the distribution was significantly negative (t-test; p = 0.02)
Figure 6B and 6C (legend's first "d"); Results, Evolving changes in PMd output-null planning activity ..., paragraph 1
What the paper reports
- t-test, p = 0.02 (n = 16 CF and VR sessions)
Over the 21 candidate sessions the mean Baseline lag is negative under all 8 readings (-2.4 to -51.4 ms). It is significantly negative only with window (a) and the 100 ms standard-deviation kernel, at -18.1 ms (t = -2.29, p = 0.033, 10 ms resolution) and -51.4 ms (p = 0.035, interpolated 50 ms activity); in the latter two sessions peak within one step of the edge of the search (C 2016-10-06 at -350 ms, C 2016-10-11 at -340 ms), and without them the mean is -20.5 ms over 19 sessions (p = 0.034). The full-width kernel and window (b) readings give p from 0.18 to 0.82. No reading over 21 sessions is within 3.5 ms of -35 ms. Over the 2002 choices of 16 sessions (9 of the 14 Monkey C candidates with the 7 Monkey M sessions) the main reading is significantly negative in 763 (38%), the 50 ms reading in 478 (816 with edge maxima dropped), and at least one of the 8 readings in 1013 (51%; 1052 with edge maxima dropped); 12 choices, all with edge maxima dropped at 50 ms resolution, give a significant mean within tolerance of -35 ms. The 19-session set gives p = 0.040 for the main reading. By monkey, -22.9 ms (C) and -8.6 ms (M). The end-of-Force part of the statement cannot be tested, as no Force trials are in the dandiset; the outcome follows the criterion's partial rule on the 21 candidates (significantly negative under one reading, outside tolerance) and holds for about half of the possible 16-session sets.
How the test differed from the paper’s method
- The paper's 16 sessions cannot be identified. All 21 dual-array center-out sessions of Monkeys C (2016-09-09 to 2016-10-21) and M (2014-02-03 to 2014-03-07) are used, and the 19 whose reward counts fit the paper's Baseline range (154 to 225) are shown as a second set. All reward trials of each file are taken as Baseline.
- The t-test is run over all 21 candidate sessions, a superset of the paper's 16, which the outcome rests on; the same test on every choice of 16 sessions is shown.
- Target acquisition (window a, the claims file's O3) is the trial's stop time minus the 0.5 s target hold, from STAR Methods, Behavioral task ("required to hold there for 0.5 s"); the cursor cannot be used, as in several sessions it never enters the stored target square.
- The outcome depends on the kernel width, which the claims file expected not to matter (O4). Significance holds with a standard deviation of 100 ms and not with a full width at half maximum of 100 ms (p 0.18 to 0.38), which would make it not reproduced. The claims file's header says a claim whose open-detail readings disagree is blocked as methods_underspecified; the protocol (author, step 1.6) lets the author adopt the best-supported reading as an assumption, and that rule is followed. The standard deviation is adopted for every claim, as the usual way a Gaussian kernel's width is stated and the form the same group used later on Monkey C recordings ("s.d. 50 ms", Gallego, Perich et al. 2020). Under it fig3b-pmd-higher-dimensional is partially reproduced.
- The lag sign follows the Figure 6B legend (a peak left of zero means output-null leads).
- A maximum within one step of the edge of the lag search (340 or 350 ms) is flagged and the test is shown with such sessions kept and dropped; the paper does not say how edge maxima were treated.

Not tested in a notebook 23

23 data missing from the dataset

Blocked23

BlockedResults, Curl field adaptation involves widespread ..., paragraph 1; Figure 2C and 2D· Data missing from the dataset
On exposure to the curl field the monkeys made large reach direction errors that shrank gradually until behavior stabilized, and mirror-image after-effects appeared at the start of Washout.
Within each session, the monkeys exhibited large errors upon exposure to the CF, which were gradually reduced until behavior stabilized ... Further evidence of adaptation was revealed by behavioral after-effects (mirror image errors) that occurred upon return to normal reaching during the Washout epoch
Results, Curl field adaptation involves widespread ..., paragraph 1; Figure 2C and 2D
Needs the Force and Washout trials of the curl field sessions. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedResults, PMd to M1 mapping is unchanged throughout adaptation to a visuomotor rotation, paragraph 2; Figure 7B and 7C· Data missing from the dataset
In VR sessions behavioral errors were similar in magnitude and time course to those of CF sessions, with large initial errors in Rotation and strong after-effects in Washout.
In many respects, adaptation during the CF and VR sessions was quite similar. Behavioral errors were similar in magnitude and time course
Results, PMd to M1 mapping is unchanged throughout adaptation to a visuomotor rotation, paragraph 2; Figure 7B and 7C
Needs the Rotation and Washout trials of the visuomotor rotation sessions, and the curl field Force trials. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedResults, Models to study adaptation-related changes ..., paragraph 2; Figure S2E· Data missing from the dataset
In a representative CF session, the pairwise covariance of neurons in the Force and Washout blocks correlated with Baseline values at r of at least 0.85 in M1 and at least 0.93 in PMd.
representative session ... Pearson's correlation with Baseline values r ≥ 0.85 for M1 and r ≥ 0.93 for PMd
Results, Models to study adaptation-related changes ..., paragraph 2; Figure S2E
Needs the Force and Washout blocks; only the Baseline split could be computed. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedResults, Models to study adaptation-related changes ..., paragraph 2; Figure S3A· Data missing from the dataset
Across all CF sessions the similarity of the covariance between Baseline and the Force and Washout blocks resembled that between random subsamples of Baseline, so the covariance structure did not change with adaptation.
Across all sessions, the differences in covariance strongly resembled those obtained from random subsamples of Baseline data for which the covariance structure should be unchanged
Results, Models to study adaptation-related changes ..., paragraph 2; Figure S3A
Needs the Force and Washout blocks to compare with the Baseline subsamples. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedResults, PMd to M1 mapping is unchanged throughout adaptation to a visuomotor rotation, paragraph 2; Figure S3B (the Results cite it as Figure S2B, which shows CF example neurons; the VR covariance summary is Figure S3B)· Data missing from the dataset
In VR sessions as well, the changes in neural activity preserved the population covariance structure.
Additionally, changes in neural activity preserved the population covariance structure
Results, PMd to M1 mapping is unchanged throughout adaptation to a visuomotor rotation, paragraph 2; Figure S3B (the Results cite it as Figure S2B, which shows CF example neurons; the VR covariance summary is Figure S3B)
Needs the Rotation and Washout blocks of the visuomotor rotation sessions. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigures 3C and S5E (cell counts in the panels); STAR Methods, Selecting cells with significant GLM predictions· Data missing from the dataset
In the CF sessions pooled, the number of M1 or PMd neurons with significant GLM fits was 214 for M1-M1, 338 for PMd-PMd, 170 for PC-M1, 144 for Pot-M1 and 97 for Null-M1; the Basic model was significant for 522 of 918 M1 and 612 of 2221 PMd neurons.
This reduced the pool of candidate cells to 522 of 918 (57%) in M1 and 612 of 2221 (28%) in PMd
Figures 3C and S5E (cell counts in the panels); STAR Methods, Selecting cells with significant GLM predictions
The models are trained on the last half of Force, which is not in the dandiset. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedResults, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 2; Figure S5E· Data missing from the dataset
Pot-M1 and Null-M1 had similar cross-validated accuracy within the end-of-Force training trials.
Both models had similar accuracy when evaluated within the end-of-Force training trials using ten-fold crossvalidation
Results, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 2; Figure S5E
The models are trained and cross-validated on end-of-Force trials, not in the dandiset. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigure 3E (p-values in the panel); Results, Curl field adaptation is not associated ..., paragraph 2· Data missing from the dataset
The rpR2 distributions of the within-area models M1-M1 and PMd-PMd did not differ between Early and Late Force trials (Kolmogorov-Smirnov p = 0.72 and p = 0.19).
M1-M1 and PMd-PMd had similar distributions during Early and Late
Figure 3E (p-values in the panel); Results, Curl field adaptation is not associated ..., paragraph 2
What the paper reports
- M1-M1, Early against Late Force rpR2: Kolmogorov-Smirnov test, p = 0.72 (n = 214 M1 cells, nine CF sessions)
- PMd-PMd, Early against Late Force rpR2: Kolmogorov-Smirnov test, p = 0.19 (n = 338 PMd cells, nine CF sessions)
Needs the Early and Late Force trials. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigure 3G (legend "No effects were significant (two-sample t-test)"); Results, Curl field adaptation is not associated ..., paragraph 2· Data missing from the dataset
Model error of M1-M1 and PMd-PMd did not differ between Early and Late Force; the local models generalized throughout CF adaptation.
We found that both within-area models ( M1 -M1, PMd -PMd) generalized well, predicting the adaptation-related changes in single-neuron spiking nearly equally well in both blocks
Figure 3G (legend "No effects were significant (two-sample t-test)"); Results, Curl field adaptation is not associated ..., paragraph 2
What the paper reports
- two-sample t-test, p > 0.01 (n = 214 M1-M1 and 338 PMd-PMd cells, nine CF sessions)
Needs the Early and Late Force trials. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigure 4D legend (Results cite it as Figure 4E); Results, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 2· Data missing from the dataset
Null-M1 generalized significantly worse in Early than in Late Force, while Pot-M1 generalized well in both blocks.
The Pot -M1 model generalized well throughout learning, making accurate predictions in Early Force as well as Late Force ... However, the Null -M1 model generalized significantly worse in the Early block than in the Late block
Figure 4D legend (Results cite it as Figure 4E); Results, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 2
What the paper reports
- Null-M1, Early against Late Force model error: two-sample t-test, p < 0.01 (n = 97 M1 cells, nine CF sessions)
Needs the Early and Late Force trials. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigure 4C; Results, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 2· Data missing from the dataset
The Null-M1 model error changed over Force trials with a time course similar to that of behavioral adaptation, while Pot-M1 stayed near zero.
Interestingly, the Null-M1 error alone paralleled the time course of behavior
Figure 4C; Results, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 2
Needs the Force trials and their behavior. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedResults, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 3; Figure S6 (pink)· Data missing from the dataset
PC-M1, with all PMd latent activity as input, also failed to generalize during CF adaptation, its accuracy changing with a time course like that of behavior.
Since PC- M1 contained all of the output-potent and output-null subspace activity, we expected that, similarly to Null -M1, it would fail to generalize during adaptation. Indeed, its accuracy changed progressively during adaptation
Results, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 3; Figure S6 (pink)
Needs the Early and Late Force trials. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigure S6D and S6E (asterisks)· Data missing from the dataset
In each monkey separately, only Null-M1 and PC-M1 had significantly larger model error in Early than in Late Force; M1-M1, PMd-PMd and Pot-M1 did not.
Not quoted: Read from the asterisks of Figure S6D and S6E, whose legend says asterisks mark Early against Late differences at p < 0.01 (two-sample t-test); no sentence names the models.
What the paper reports
- Null-M1 and PC-M1, Early against Late, each monkey: two-sample t-test, p < 0.01 (n = Monkey C five CF sessions, Monkey M four CF sessions)
Needs the Early and Late Force trials. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigure S5F (legend and p-values in the panel)· Data missing from the dataset
Of the unnormalized rpR2 distributions of the five models in Early against Late CF trials, only Null-M1 (p = 0.04) and PC-M1 (p = 0.01) changed; Pot-M1 (p = 0.40), M1-M1 (p = 0.72) and PMd-PMd (p = 0.19) did not.
values for all predicted neurons for the five GLM models during Early and Late CF adaptation. Only Null-M1 and PMd-M1 were significantly changed between the two periods
Figure S5F (legend and p-values in the panel)
What the paper reports
- Null-M1: Kolmogorov-Smirnov test, p = 0.04 (n = 97 cells)
- PC-M1 (PMd-M1): Kolmogorov-Smirnov test, p = 0.01 (n = 170 cells)
- Pot-M1: Kolmogorov-Smirnov test, p = 0.4 (n = 144 cells)
Needs the Early and Late Force trials. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigure S7A and S7B (legend and p-values in the panels)· Data missing from the dataset
Without normalization by cross-validated performance, the change in rpR2 between Early and Late CF was significant only for Null-M1 (p = 0.01) and PC-M1 (p = 0.00); in VR no model changed (p from 0.28 to 0.80).
The distributions of change in rpR2 for the curl field sessions, but without the normalization by the cross-val- ... performance. Asterisks indicate significance at p < 0.01. (B) Same as Panel A, but for the visuomotor rotation sessions. There were no significant differences in the models
Figure S7A and S7B (legend and p-values in the panels)
What the paper reports
- CF, M1-M1: p = 0.13
- CF, PMd-PMd: p = 0.44
- CF, Pot-M1: p = 0.54
- CF, Null-M1: p = 0.01
- CF, PC-M1: p = 0
- VR, M1-M1: p = 0.33
- VR, PMd-PMd: p = 0.62
- VR, Pot-M1: p = 0.28
- VR, Null-M1: p = 0.72
- VR, PC-M1: p = 0.8
Needs the Force trials of the curl field sessions and the Rotation trials of the rotation sessions. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigure S7E; STAR Methods, Controls for the output-potent and output-null subspaces; Results, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 3· Data missing from the dataset
The Null-M1 failure and Pot-M1 generalization held for M1 dimensionalities from 7 to 20 (PMd twice that); Null-M1 was significant at all tested values except 9, Pot-M1 at none.
Our primary effect that Pot-M1, and not Null-M1, generalizes to early CF trials held for a wide range of dimensionalities
Figure S7E; STAR Methods, Controls for the output-potent and output-null subspaces; Results, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 3
What the paper reports
- two-sample t-test, p < 0.01
Needs the Early and Late Force trials. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigure S7F; Results, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 3· Data missing from the dataset
The main result held with Factor Analysis in place of PCA; in CF only Null-M1 and PC-M1 had larger Early error, and in VR no model did.
We thus replicated these main results using Factor Analysis, another common method for dimensionality reduction of neural data, instead of PCA
Figure S7F; Results, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 3
Needs the Force and Rotation trials. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigure S7C and S7D; Results, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 3· Data missing from the dataset
With single-neuron spiking as inputs, M1sn-M1 and PMdsn-PMd generalized well through CF adaptation while PMdsn-M1 did not, in each monkey; in Washout, PMdsn-M1 alone showed de-adaptation in CF (p = 0.00), and no model changed in VR.
Indeed, the local models ( M1sn -M1, PMdsn PMd) generalized well throughout adaptation, while the between area model ( PMdsn -M1) did not
Figure S7C and S7D; Results, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 3
What the paper reports
- Washout CF, M1sn-M1: p = 0.25
- Washout CF, PMdsn-PMd: p = 0.15
- Washout CF, PMdsn-M1: p = 0
- Washout VR, M1sn-M1: p = 0.07
- Washout VR, PMdsn-PMd: p = 0.64
- Washout VR, PMdsn-M1: p = 0.6
Needs the Force and Washout trials. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigure 5B; Results, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 4· Data missing from the dataset
Models trained on Baseline predicted the first eight Washout trials worse than later Washout trials for Null-M1 only; Pot-M1 predicted both blocks well.
We found that Null -M1 generalized poorly in the earliest trials, then rapidly improved with de-adaptation as output-null activity began to return to Baseline. There was no such effect for the Pot -M1 model, which accurately predicted both early and late Washout trials
Figure 5B; Results, Output-potent activity in PMd consistently predicts M1 neurons ..., paragraph 4
What the paper reports
- Null-M1, Early against Late Washout model error: two-sample t-test, p < 0.01
Needs the Washout trials; Baseline models could be trained but not tested. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigure 6D and 6E; Results, Evolving changes in PMd output-null planning activity ..., paragraph 2· Data missing from the dataset
Changes in preparatory (planning) activity from Baseline were largely confined to the output-null subspace and grew significantly from Early to Late Force; in the movement window, M1 activity changed significantly more in Late Force.
Changes in preparatory activity were largely confined to the output-null subspace. Furthermore, the amount of this altered preparatory activity increased significantly between Early and Late adaptation
Figure 6D and 6E; Results, Evolving changes in PMd output-null planning activity ..., paragraph 2
What the paper reports
- Null subspace, planning window, Early against Late Force: p < 0.05
- M1, movement window, Early against Late Force (asterisk on the M1 bars of Figure 6E): p < 0.05
Needs the Force trials to compare with Baseline. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigure 7D and 7E; Results, PMd to M1 mapping is unchanged ..., paragraph 2· Data missing from the dataset
During VR adaptation all four GLM models, including Null-M1, generalized to Early and Late Rotation trials, with no significant change in model error.
However, all GLM models, including Null -M1, accurately generalized throughout Rotation
Figure 7D and 7E; Results, PMd to M1 mapping is unchanged ..., paragraph 2
What the paper reports
- two-sample t-test, p > 0.01 (n = seven VR sessions)
Needs the Rotation trials and which sessions were rotation sessions. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedFigure 7F; Results, PMd to M1 mapping is unchanged ..., paragraph 2· Data missing from the dataset
In VR Washout neither Pot-M1 nor Null-M1 showed a change in performance during de-adaptation, despite behavioral after-effects.
Furthermore, we did not see any effect of de-adaptation in the neural predictions during Washout, despite the presence of behavioral after-effects
Figure 7F; Results, PMd to M1 mapping is unchanged ..., paragraph 2
Needs the Washout trials of the rotation sessions. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.
BlockedResults, PMd to M1 mapping is unchanged ..., paragraph 2; Discussion, Visuomotor rotation and curl field likely employ distinct adaptive processes· Data missing from the dataset
The Null-M1 failure to generalize is specific to CF adaptation; it is absent in VR adaptation of similar behavioral magnitude, so it is not a necessary consequence of changing behavior.
This result highlights a fundamental difference in the neural adaptation to these two perturbations
Results, PMd to M1 mapping is unchanged ..., paragraph 2; Discussion, Visuomotor rotation and curl field likely employ distinct adaptive processes
Depends on fig4d-null-m1-fails-early, fig5-washout-null-m1, fig7de-vr-all-generalize, fig7f-vr-washout-unchanged and fig7bc-vr-behavior-similar, all blocked. The dandiset holds only the Baseline epoch of the dual-array center-out sessions (about 200 reward trials each, with no perturbation block in the reach errors) and no label of a session as curl field or visuomotor rotation. The paper says its data may be shared upon request; no other public copy is named.

Findings about the data

What attempts found about the dataset itself, for its owner and the archive.

Data missingThe dandiset holds about 200 reward trials per dual-array center-out session of Monkeys C and M, the size of the paper's Baseline epoch, with steady reach errors. The Force, Rotation and Washout epochs that 23 of the paper's 28 findings rest on are not in it.
Description missingNo file or metadata field says which sessions were curl field or visuomotor rotation sessions, which epoch the trials belong to, or which 16 sessions the paper analysed. Monkey C has 14 candidate sessions for the paper's 9.
Metadata errorIn seven Monkey C sessions (2016-09-09 to 2016-09-23) the cursor ends rewarded reaches about 4.2 cm from the center, with peak speeds near 14 cm/s, while target_corners places every target at 8 cm and the later sessions end at about 7.9 cm with peak speeds near 25 cm/s. The paper says Monkey C's first three sessions used 6 cm targets, which fits neither.
Metadata errorThe session description of the seven Monkey C sessions from 2016-09-09 to 2016-09-23 ends in a cut-off warning ("Warning: This session with Monkey"), in the files and in the archive's metadata, so what it warns of is lost. These are the sessions whose cursor positions do not match the stored targets.
Metadata errorIn all seven early Monkey C sessions the cursor never enters the square given by target_corners on rewarded trials, and in Monkey M 2014-02-03 and 2014-02-18 it enters it on 1% of them.
File format errorThe cursor timestamps repeat 8531 times in C 2016-09-23 and 5014 times in C 2016-10-11, so velocities computed from them are undefined on many samples.
OtherUnits above 5 Hz are far fewer than the paper reports, at most 634 M1 and 1224 PMd for any choice of 16 candidate sessions, against 918 and 2221. Monkey M has 6 to 21 PMd units above 5 Hz per session.
Methods not availableQuestion for the authors. How was the noise scaled in the dimensionality estimate adapted from Machens et al. (2010)? The Methods subtract random pairs of trials; taken as it is, that noise exceeds every component of the trial-averaged rates and gives 0 dimensions. This attempt assumed the noise of the trial average (difference over the square root of 2K) and the 95% rule, which gives medians of 7 and 10 with a kernel of 100 ms standard deviation (9 and 13 with a full width at half maximum of 100 ms).
Methods not availableQuestion for the authors. Is the 100 ms width of the Gaussian kernel its standard deviation or its full width at half maximum, and how were trials of unequal length averaged? With the full width the dimensionality medians are 9 and 13 (paper 8.5 and 14) and the Baseline lag is not significant; with the standard deviation the medians are 7 and 10 and the lag is significantly negative over 21 sessions. This attempt assumed the standard deviation for every claim.
Selection not documentedThe dandiset holds 14 dual-array Monkey C sessions where the paper used 9, and does not say how they relate to the paper's sessions or why the perturbation epochs were left out.

Notebooks

Every registered entry that uses this dataset. An entry can hold several notebooks, and can use more than one dataset.

Reproduction PMd and M1 population structure in Baseline reaching (Perich, Gallego and Miller 2018)

Tests the findings of Perich, Gallego and Miller (2018) that can be tested on dandiset 000688. The dandiset holds the Baseline epoch only of the dual-array center-out sessions of Monkeys C and M, without the curl field, visuomotor rotation and Washout trials, and without labels saying which sessions are the paper's. It tests the unit counts, the dimensionality of the PMd and M1 manifolds, the null/potent index of PMd neurons, and the lag between output-null and output-potent activity in Baseline. All files

Reproducing Perich et al. (2018) from dandiset 000688
reproduce_perich_2018.ipynbExecuted, with outputs · Source

Verification

Round 1
2026-10-31Agreed with 28 of 28 outcomes. Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook). Reviewed: The handover of the first run (the executed notebook with the conclusions and summary removed), the helper module and the claims file; the verifier also queried the DANDI API for the metadata of all 111 assets.
Round 2
2026-10-31Agreed with 28 of 28 outcomes. Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook). Reviewed: The full executed notebook of the first revision (with conclusions and summary), the helper module, the claims file and the attempt record.
Round 3
2026-10-31Agreed with 1 of 1 outcomes. Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook). Reviewed: Round 3. The one proposed change of outcome (fig3b-pmd-higher-dimensional, reproduced to partially reproduced) and the analysis behind it in the executed notebook of the second revision (cells 16, 17, 19, 33 and 34), with the helper module, the claims file and the paper. Nothing else was reviewed. Report in transcripts/000688/verifier-3-report.md.

Method differences

How the tests differed from the paper’s method or data, and the claims each applies to.

The paper's 16 sessions cannot be identified. All 21 dual-array center-out sessions of Monkeys C (2016-09-09 to 2016-10-21) and M (2014-02-03 to 2014-03-07) are used, and the 19 whose reward counts fit the paper's Baseline range (154 to 225) are shown as a second set. All reward trials of each file are taken as Baseline.methods-neuron-counts figs4d-dimensionality-robust-to-size figs4c-null-potent-index-centered fig6c-null-leads-potent
Target acquisition (window a, the claims file's O3) is the trial's stop time minus the 0.5 s target hold, from STAR Methods, Behavioral task ("required to hold there for 0.5 s"); the cursor cannot be used, as in several sessions it never enters the stored target square.fig3b-pmd-higher-dimensional figs4c-null-potent-index-centered fig6c-null-leads-potent
The 5 Hz rule uses the trials in the file (Baseline only), from start to stop of each reward trial; all trials are shown as a variant. The paper's rate is over the whole session (claims file criterion).methods-neuron-counts
The paper excluded cells with high coincidence across channels; the units are counted as deposited.methods-neuron-counts
The paper's 16 sessions cannot be identified. All 21 dual-array center-out sessions of Monkeys C (2016-09-09 to 2016-10-21) and M (2014-02-03 to 2014-03-07) are used, and the 19 whose reward counts fit the paper's Baseline range (154 to 225) are shown as a second set, with every choice of 16 sessions. All reward trials of each file are taken as Baseline.fig3b-pmd-higher-dimensional
Assumed that the noise eigenvalues are scaled to the noise of a trial average (the difference of two trials divided by the square root of 2K, K the mean number of trials per target). STAR Methods, Dimensionality reduction, says only that the activity of random pairs of trials was subtracted; taken as it is, or divided by the square root of two, the noise exceeds every component of the trial-averaged data and every estimate is 0, which cannot give the paper's medians. All three scalings are shown. The choice was made after the raw and root-two scales gave 0; a synthetic check of known rank (notebook) shows rule (b) with this scale recovers the rank, but the raw and root-two scales also recover it there, so the check does not choose between the scales; the choice rests on the other scales giving 0 on the real data.fig3b-pmd-higher-dimensional
The outcome depends on the kernel width, which the claims file expected not to matter (O4). Assumed a standard deviation of 100 ms for every claim, as the usual way a Gaussian smoothing kernel's width is stated and the way the same group states it in a later paper on Monkey C recordings ("a Gaussian kernel (s.d. 50 ms)", Gallego, Perich et al. 2020, Nat. Neurosci.). STAR Methods, Dimensionality reduction, says only "a non-causal Gaussian kernel of width 100 ms". With a full width at half maximum of 100 ms the medians are within tolerance and the claim would be reproduced, while fig6c-null-leads-potent would be not reproduced; neither reading favours every claim. The claims file's header says a claim whose open-detail readings disagree is blocked as methods_underspecified; the protocol (author, step 1.6) instead lets the author adopt the best-supported reading as an assumption, and that rule is followed.fig3b-pmd-higher-dimensional
Trials of unequal length are aligned on the window start and cut to the session's shortest window before averaging per target; the paper does not say how they were averaged (a standard choice).fig3b-pmd-higher-dimensional
All units of each area are used; the paper does not say whether a rate threshold applies here.fig3b-pmd-higher-dimensional
Assumed that the noise eigenvalues are scaled to the noise of a trial average (the difference of two trials divided by the square root of 2K, K the mean number of trials per target). STAR Methods, Dimensionality reduction, says only that the activity of random pairs of trials was subtracted; taken as it is, or divided by the square root of two, the noise exceeds every component of the trial-averaged data and every estimate is 0, which cannot give the paper's medians. All three scalings are shown.figs4d-dimensionality-robust-to-size
100 subsamples per size (the claims file's standard choice) with 100 noise repetitions per subsample, to fit the time limit; the paper's 1000 noise repetitions are kept for all neurons.figs4d-dimensionality-robust-to-size
Run with both kernel readings (standard deviation and full width at half maximum of 100 ms) in windows (a) and (b).figs4d-dimensionality-robust-to-size
The readings of magnitude disagree in one of eight cases; the sum of absolute values, the claims file's first reading of "the magnitude ... summed across all dimensions" (STAR Methods, Output-potent and output-null subspaces), is adopted.figs4c-null-potent-index-centered
The one failing reading (sums of squares, window a) uses the full-width kernel. Under the standard-deviation kernel, the reading adopted for every claim, all four window and magnitude readings meet the criterion.figs4c-null-potent-index-centered
The t-test is run over all 21 candidate sessions, a superset of the paper's 16, which the outcome rests on; the same test on every choice of 16 sessions is shown.fig6c-null-leads-potent
The outcome depends on the kernel width, which the claims file expected not to matter (O4). Significance holds with a standard deviation of 100 ms and not with a full width at half maximum of 100 ms (p 0.18 to 0.38), which would make it not reproduced. The claims file's header says a claim whose open-detail readings disagree is blocked as methods_underspecified; the protocol (author, step 1.6) lets the author adopt the best-supported reading as an assumption, and that rule is followed. The standard deviation is adopted for every claim, as the usual way a Gaussian kernel's width is stated and the form the same group used later on Monkey C recordings ("s.d. 50 ms", Gallego, Perich et al. 2020). Under it fig3b-pmd-higher-dimensional is partially reproduced.fig6c-null-leads-potent
The lag sign follows the Figure 6B legend (a peak left of zero means output-null leads).fig6c-null-leads-potent
A maximum within one step of the edge of the lag search (340 or 350 ms) is flagged and the test is shown with such sessions kept and dropped; the paper does not say how edge maxima were treated.fig6c-null-leads-potent

Provenance

Attempt record: 000688-perich-2018-01
Paper: A Neural Population Mechanism for Rapid Learning
Date: 2026-10-31
Made by: An agent, run by CatalystNeuro
Dataset: DANDI:000688
Version: 0.250122.1735 (doi:10.48324/dandi.000688/0.250122.1735)
Run on: GitHub Actions
Harness: Claude Code
Model: claude-opus-5-5
Protocol: pilot-12

Transcripts of the agent runs

Each run’s prompt, messages and tool calls, with long output cut and the text of the paper left out, and its final report.

← DANDI:000686 DANDI:000689 →