DANDI:000252

participant N: 100 trials / session for 10 sessions; participant J: 100 trials / session for 2 sessions). We recorded 111 neurons per session (min: 102; max: 119) from N-PPC, 160 neurons per session (min: 159; max: 160) from J-PPC, and 130 neurons per session (min: 120; max: 130) from J-MC

Results, Decoding individual finger presses of either hand, paragraph 2

The dandiset has 10 sessions of participant N and 2 of J, each of 100 trials (10 per condition). Before the 0.5 Hz exclusion N-PPC has 111.4 units per session (102 to 119), J-PPC 159.5 (159 to 160) and J-MC 129.5 (120 to 139); the J-MC maximum of 139 is above the paper's 130 but inside the 10% widening of that bound (143). After the exclusion the means are 94.3 and 93.0 for N-PPC under O1 (a) and (b), outside tolerance, and 157.5 or 156.0 for J-PPC and 126.0 or 125.0 for J-MC. The paper's counts are those before the exclusion.

How the test differed from the paper’s method

• The counts that match are those before the 0.5 Hz exclusion, which the Methods (Feature Extraction) say applies to all analyses; the criterion names this count among its readings.

1111 N-PPC neurons were recorded across 10 sessions

Supplementary Figure 11 legend

1114 N-PPC units over the 10 sessions before the exclusion (paper 1111), 943 under O1 (a) and 930 under O1 (b), both outside tolerance.

How the test differed from the paper’s method

• As for tenfinger-recording-counts, the matching count is the one before the exclusion.

A linear classifier (Methods) was able to discriminate between all ten fingers (cross-validated classification accuracy, N-PPC: 70%; J-PPC: 66%; J-MC: 75%; chance: 10%)

Results, Decoding individual finger presses of either hand, paragraph 3; Figure 7c to 7e

Diagonal LDA per session, stratified 8-fold, predictions pooled over sessions. Under O1 (a) and O2 (a), 72.2% (N-PPC, 1000 trials), 69.5% (J-PPC, 200) and 78.5% (J-MC, 200); over all readings of O1 (including no exclusion) and O2 the values are 71.4 to 73.0%, 68.0 to 70.0% and 78.0 to 78.5%, and single seeds of O2 (b) 69.9 to 73.6%, 66.0 to 72.5% and 76.5 to 80.0%, all within tolerance. Chance is 10%; with shuffled labels per-session accuracies are 0.04 to 0.17.

How the test differed from the paper’s method

• The variance floor of the diagonal LDA (1e-9 of the largest variance, as scikit-learn's GaussianNB) is a standard choice that only prevents division by zero.

The majority (N-PPC: 76%; J-PPC: 66%; J-MC: 68%) of classification errors were adjacent-finger-confusion or matching-across-hand-confusion

Results, Decoding individual finger presses of either hand, paragraph 3; Figure 7c to 7e

Under O1 (a) and O2 (a), 213 of 278 N-PPC errors (119 adjacent, 94 matching), 44 of 61 J-PPC errors and 27 of 43 J-MC errors are adjacent or matching; over all readings the shares are 73.9 to 76.6%, 66.2 to 72.1% and 62.8 to 65.3%, within tolerance of 76%, 66% and 68%.

On average, matching right/left finger pairs were 1.56 distance-units (95% CI: [1.33, 1.78], Figure 8b ) closer to each other than non-matching fingers were. Matching fingers were also represented more similarly than non-matching fingers in J-MC (mean difference: 4.30, 95% CI: [2.74, 5.46]), but this result was not conclusive in J-PPC (mean difference: 0.27, 95% CI: [-0.17, 0.64])

Results, Compositional representation of finger and laterality, paragraph 1; Figure 8b; Supplementary Figures 9b and 10b

What the paper reports

• N-PPC, non-matching minus matching distance: 95% CI 1.33 to 1.78 (n = RDM elements of 10 sessions)
• J-MC, non-matching minus matching distance: 95% CI 2.74 to 5.46 (n = RDM elements of 2 sessions)
• J-PPC, non-matching minus matching distance: 95% CI -0.17 to 0.64 (n = RDM elements of 2 sessions)

Under O4 (a), with each partition's precision scaled to that of the partition mean, the non-matching minus matching difference is 1.42 (BCa 95% CI 1.19 to 1.64) in N-PPC, 3.06 (1.73 to 4.13) in J-MC and 0.22 (-0.10 to 0.67) in J-PPC for O1 (a) and O3 (a); over O1 and O3 the values are 1.42 to 1.71, 2.99 to 3.38 and 0.19 to 0.22, all within tolerance, with N-PPC and J-MC CIs above 0 and the J-PPC CI including 0 (also by the percentile CI, in all 18 readings). Under O4 (b), and the unscaled variant O4 (a1), the sign pattern is the same but the values are about a fifth as large (N-PPC 0.28 to 0.36, J-MC 0.51 to 0.68), outside tolerance, which would be partially reproduced. On synthetic data with a known distance, O4 (a) returns 5.17 times the distance in single-trial noise units and O4 (a1) and (b) 1.03 times; the N-PPC mean RDM spans 0.57 to 3.90 under O4 (a), 0.11 to 0.78 under O4 (a1) and 0.12 to 0.79 under O4 (b). The paper's text gives no colour-bar range for Figure 8a to tell the scales apart. One N-PPC unit with zero residual variance in a partition is dropped.

How the test differed from the paper’s method

• Assumed that the per-partition noise passed to rsatoolbox is the precision of the partition mean, the single-trial precision times the 5 trials per condition in each partition. The Methods say the toolbox (ref 57) incorporates the noise covariances of both folds, which is its list-of-noises branch, and the toolbox's calc_rdm_crossnobis documents that this list must hold the precision of the mean of each fold. The paper does not say how it computed the noise, and the scaling changes the values about fivefold; the unscaled result is shown.
• Units with zero residual variance in a partition make the noise covariance singular and are left out of that session's distance (one N-PPC unit); a standard choice.

We found that inter-finger vectors were similar across hands, with the index finger relatively distinct from the middle and ring fingers

Results, Compositional representation of finger and laterality, paragraph 4; Figure 8a and 8d

In the N-PPC distances averaged over 10 sessions, under every reading of O1, O3 and O4, the index-middle and index-ring distances exceed the middle-ring distance in both hands (O1 a, O3 a, O4 a, right 2.16 and 2.96 against 0.62, left 1.30 and 2.23 against 0.57). In single sessions the pattern holds in both hands in 7 to 10 of 10 sessions. The outcome does not depend on the noise scaling assumed for fig8b-matching-closer, as all readings agree.

How the test differed from the paper’s method

• The MDS of the figure is drawn by classical MDS of the clipped squared distances and aligned by Generalized Procrustes with scaling; it is shown only and does not decide the outcome.

Consistent with this hypothesis, cross-condition hand-decoding generalization performance (hand CCGP) was significantly above chance (p < 0.001, permutation test). Furthermore, hand CCGP was close to the standard decoding accuracy

Results, Compositional representation of finger and laterality, paragraphs 5 and 6; Figure 8e; Supplementary Figures 8, 9e, 10e and 11

What the paper reports

• N-PPC hand CCGP against chance: permutation test, p < 0.001 (n = pseudo-population of N-PPC neurons pooled over 10 sessions, 1001 shuffles)
• J-MC hand CCGP against chance: permutation test, p < 0.001 (n = pseudo-population of J-MC neurons pooled over 2 sessions)
• J-PPC hand CCGP against chance: permutation test, p < 0.001 (n = pseudo-population of J-PPC neurons pooled over 2 sessions)

In N-PPC (943 neurons under O1 a), J-PPC (315) and J-MC (252), under every reading of O1, O6 and O7, no one of 1001 shuffles reaches the hand CCGP (p = 1/1002). Hand CCGP is 0.970 to 0.990 in N-PPC and J-PPC and 0.860 to 0.870 in J-MC, against shuffle means near 0.50, and at least 0.913 of the hand accuracy on the same pseudo-population, so close by both O9 (a) and (b).

How the test differed from the paper’s method

• The shuffle permutes the ten condition labels across pseudo-trials and recomputes CCGP; the paper says only that labels were shuffled 1001 times. For O6 (b) each shuffle uses a new random matching of trials.
• Standard accuracy is computed on the same pseudo-population (assumption, see jppc-finger-ccgp-low); the per-session accuracy, run under every O1 x O7, does not change this outcome, as hand CCGP is at least 0.940 of it.

The finger-type dimension also generalized across hands (p < 0.001), and finger-type CCGP was close to the standard decoding accuracy

Results, Compositional representation of finger and laterality, paragraphs 5 and 6; Figure 8e; Supplementary Figure 9e

What the paper reports

• N-PPC finger-type CCGP against chance: permutation test, p < 0.001 (n = pseudo-population of N-PPC neurons pooled over 10 sessions, 1001 shuffles)
• J-MC finger-type CCGP against chance: permutation test, p < 0.001 (n = pseudo-population of J-MC neurons pooled over 2 sessions)

In N-PPC and J-MC, under every reading, no shuffle reaches the finger-type CCGP (p = 1/1002). N-PPC CCGP is 0.918 to 0.954, at least 0.925 of its standard accuracy. J-MC CCGP is 0.735 to 0.790, 0.770 to 0.823 of the within-hand accuracy (O8 a) and 0.775 to 0.857 of the accuracy over both hands (O8 b). Closeness holds for both arrays under O9 (b), a ratio above 0.55, and fails for J-MC under the 0.9 margin of O9 (a). The criterion asks for one reading, so the claim is reproduced. With the per-session standard accuracy both arrays also pass O9 (a) in some readings (J-MC 0.908 of the within-hand accuracy for O1 a, O6 a, SVM), and N-PPC CCGP exceeds the per-session accuracy (ratio 1.15 to 1.31); the outcome is the same under both populations.

How the test differed from the paper’s method

• As for fig8e-hand-ccgp, the shuffle scheme and the population of the standard accuracy are choices the paper does not state.

Even when accounting for differences in neural population size, finger-type CCGP for J-PPC was lower than finger-type CCGP for N-PPC and J-MC

Results, Compositional representation of finger and laterality, paragraph 6; Supplementary Figure 11

With 100 random subsets per size, at 50, 100, 150, 200 and 250 neurons and under every reading of O1, O6 and O7, the mean J-PPC finger-type CCGP (0.253 to 0.338) is below the means of N-PPC (0.472 to 0.867) and J-MC (0.559 to 0.790) at the same size.

Interestingly, J-PPC finger representations were less compositional. While above chance (p < 0.001), the finger-type CCGP (36%) was much lower than the within-hand accuracy (65%)

Results, Compositional representation of finger and laterality, paragraph 6; Supplementary Figure 10e

What the paper reports

• permutation test, p < 0.001 (n = pseudo-population of J-PPC neurons pooled over 2 sessions)

J-PPC finger-type CCGP is 0.290 to 0.380 over the readings (shuffle mean 0.20), far below the within-hand accuracy on the pseudo-population (0.828 to 0.890). The permutation p is below 0.001 in one of eight readings (O1 b, O6 a, linear SVM, no null value of 1001 reaching the CCGP of 0.380, p = 1/1002, within-hand 0.850); in the others 1 to 37 null values reach it (p 0.002 to 0.038), so the paper's p below 0.001 holds only in that reading, at the floor of the test. There the CCGP is within tolerance of 36% and the within-hand accuracy is outside the tolerance of 65%, with CCGP below 0.9 of it, which the criterion calls partially reproduced. With the within-hand accuracy computed per session, run under every O1 x O7, it is 0.680 to 0.690, within tolerance in all eight readings, and the same reading (O1 b, O6 a, SVM) is reproduced. The two populations give different outcomes; the pseudo-population is kept for the reasons in the deviation below, a judgement call, as the round-2 verifier notes.

How the test differed from the paper’s method

• Assumed that the standard (within-hand) accuracy is computed on the same pseudo-population as CCGP, since the Methods (Compositionality of finger representations) call it an upper bound on CCGP, which holds only for the same neurons, and the CCGP literature it cites (ref 60) computes both on one pseudo-population; with the per-session accuracy N-PPC finger CCGP exceeds the bound (ratio up to 1.31). The Methods' offline classification trains a separate classifier per session, and that reading, now run under every O1 x O7 next to each O6 row, gives 0.680 to 0.690, near the paper's 65%, and would change the outcome to reproduced.
• The shuffle scheme of the permutation test is not stated; condition labels are permuted across pseudo-trials. The J-PPC p-values lie near the 0.001 bound, so the scheme can matter. For O6 (b) the observed CCGP is a mean over 100 matchings and each null value uses one matching, which makes p conservative; it does not decide the outcome, which rests on an O6 (a) reading.

most neurons (N-PPC: 66%; J-PPC: 57%; J-MC: 78%) discriminated firing rates across fingers (q < 0.05)

Results, Decoding individual finger presses of either hand, paragraph 2

One-way ANOVA of Go-window rates across the ten conditions, Benjamini-Hochberg over the neurons of each array pooled over sessions. Under O1 (a) 47.8% of 943 N-PPC, 38.1% of 315 J-PPC and 65.1% of 252 J-MC neurons have q below 0.05; under O1 (b) 48.9%, 38.1% and 65.6%; before the exclusion 40.4%, 37.6% and 63.7%. All are below the lower tolerance bounds (59.4%, 51.3%, 70.2%). As variants, uncorrected p below 0.05 gives at most 56.1%, 46.5% and 68.8%, FDR per session, averaged over sessions, gives at most 0.3 points more than pooled FDR (40.7% against 40.4% for N-PPC before the exclusion, 38.4% against 38.1% for J-PPC under O1 a; cell 13), and the Delay, Cue and ITI windows give fewer (0 to 49%). Moving the 500 ms window's start from 0.5 s before to 0.5 s after the Go cue gives at most 47.8%, 43.5% and 72.2%. A 1 s window from the Go cue gives 60.8%, 56.8% and 82.9%, all within tolerance; it is a variant only, as the Methods use 500 ms windows unless stated and the claims file fixes the 200 to 700 ms Go window (O11). Ten-finger decoding on the same rates matches the paper, so the gap is in the single-neuron test or its window, not in the data's information.

How the test differed from the paper’s method

• The Go window is used, as claims-file O11 says the text supports no other.
• Units with the same rate in every trial (5 N-PPC and 2 J-MC before the exclusion, none after it) have no ANOVA p; they are counted in the FDR's m and as not discriminative.