Claims 52
From the paper Sampling representational plasticity of simple imagined movements across days enables long-term neuroprosthetic control · claims file · checked 2026-10-21: ready
- fig1g-somatotopy-b1
In B1, high-gamma spatial maps of right-hand actions are more similar to each other (Jaccard coefficient) than to maps of left-hand actions and than to maps of all other actions.
maps of right-hand movements were somatotopically more similar to each other 20 , 27 , 28 than to left-handed or other whole-body movements (B1, Fig. 1G , Mann Whitney U tests between Jaccard similarity coefficients: within maps of Rt. Hand movements vs. between maps of Rt. and Lt. Hand movements, z = 3.724 , p = 1.964 × 10 – 4 , within maps of Rt. Hand movements vs. between maps of Rt. Hand and All other movements z = 4.435 , p = 9.2 × 10 – 6
Results, Mesoscale representation of body parts, paragraph 2; Figure 1GHow it will be judged
Methods (Analyses to identify significant neural responses; Representational structure analyses). Per action, high-gamma (70 to 150 Hz, Hilbert envelope averaged over the eight sub-bands, 4th order Butterworth filters) is z-scored per session to the first 2 min baseline, epoched from 1000 ms before the get-ready cue to the end of the go period, and z-scored per trial to the 1000 ms before the get-ready cue; time points whose z-scored value across trials exceeds 3 are removed. At every channel and time point a one-sample t-test (threshold 0.05) is cluster corrected with 750 bootstrap samples of mean-centered trials, cluster mass the sum of squared t within clusters contiguous in time and channel neighbours, tested at 0.05. A channel belongs to an action's map when it is part of any significant cluster. Maps are binary vectors; the Jaccard coefficient is taken for every pair of right-hand maps (within), every right-hand and left-hand pair, and every right-hand and other-action pair; Mann Whitney U tests compare the within set with each between set. Open details, named as readings. (1) Channel neighbourhood for clustering, which the paper does not define; readings (a) 4-connected grid neighbours, (b) 8-connected grid neighbours. (2) Which actions count as right hand, left hand and other; the paper names them only in Figure 1E-F, so the reading is the action names' effector, as the dataset labels them. Reproduced when both tests give p < 0.05 with the within-right-hand similarity higher. Reading (1) is a method detail, so the outcome is the one both readings give; if they differ, the claim is blocked as methods_underspecified.
- fig1h-somatotopy-b3
In B3, high-gamma maps of right-hand actions are also more similar to each other than to left-hand maps and than to all other actions' maps.
B3, Fig. 1H : within maps of Rt. Hand movements vs. between maps of Rt. and Lt. Hand movements, z = 2.11 , p = 0.0345 , within maps of Rt. Hand movements vs. between maps of Rt. Hand and All other movements z = 3.7119 , p = 2.057 × 10 – 4
Results, Mesoscale representation of body parts, paragraph 2; Figure 1HHow it will be judged
As fig1g-somatotopy-b1, on B3's 32 actions and 253 channels. Reproduced when both tests give p < 0.05 with the within-right-hand similarity higher, under both readings of the channel neighbourhood; if the readings differ, blocked as methods_underspecified.
- fig1ij-within-effector-levels
Actions of one effector differ in activation level at a shared channel, with a hand-knob electrode showing different high-gamma responses to three hand actions and a ventral precentral channel showing different responses to orofacial actions.
an electrode in the ‘hand-knob’ area had significantly different γ H between three hand actions ( Fig. 1I , Fig. S1C ). Similarly, a ventral channel along the precentral gyrus 29 showed differential γ H response levels to orofacial movements ( Fig. 1J )
Results, Mesoscale representation of body parts, paragraph 2; Figure 1I-JHow it will be judged
Methods (Analyses to identify significant neural responses, paragraph on Fig. 1I-J). At each time point a repeated measures ANOVA over trials compares the within-effector actions' high-gamma ERPs at one channel; clusters of contiguous significant time points are summed in F and compared with 750 bootstrap samples of mean-centered data (maximum cluster mass), at 0.05. Open details. The two channels and the three hand and the orofacial actions are shown only in the figure; readings (a) the channel the paper identifies as hand knob (Figure S1C) and the channel in ventral precentral gyrus, if the dataset labels channels, (b) every channel significant for the effector in fig1g, reported as the fraction with a significant cluster. How an rmANOVA is formed over trials of different actions is not stated (trials are not paired across actions); readings (a) one-way ANOVA across actions with trials as observations, (b) trials paired by repetition index within block. Reproduced when under every reading at least one cluster for the hand actions and one for the orofacial actions is significant at the channels of reading (a), or a majority of channels under reading (b).
- figs2-pairwise-distances-significant
Almost all pairwise Mahalanobis distances between actions are greater than chance (99.7% of pairs in B1, 100% in B3, FDR corrected).
almost all the pairwise neural distances between movements were significantly greater than chance ( Fig. S2 , B1: 99.7% of all pairwise comparisons; B3: 100% of all comparisons, α = 0.05 level, FDR corrected)
Results, Mesoscale representation of body parts, paragraph 3; Figure S2E-FHow it will be judged
Methods (Representational structure analyses). Data are 200 ms non-overlapping bins of delta, beta and high-gamma power from 600 to 3400 ms after the go cue on all channels (384 features in B1, 759 in B3), temporally smoothed, z-scored per feature across all trials of a session, and concatenated across sessions per action. The distance is (mu_a - mu_b)' ((C_a + C_b)/2)^-1 (mu_a - mu_b). The null for each pair (Figure S2) fits one Gaussian to the pooled samples of both actions, draws as many samples, splits them in two with k-means and takes the distance between the clusters; the pair is significant when its distance exceeds the 95th percentile, after FDR (Benjamini-Hochberg) over all pairs at 0.05. Open details. The number of null iterations and the temporal smoothing kernel are not given; a standard choice (1000 iterations, no extra smoothing beyond binning) is used and reported. The tolerance of 0.3 points is one pair in B1's 435 pairs (0.23 points) rounded up, since 10% of a value near 100% would admit 90%. Reproduced when both fractions are within tolerance; partially reproduced when both are at least 90%.
- fig1kl-b1-b3-correlation
For the 30 actions common to both participants, the pairwise distances of B1 and B3 are correlated (r = 0.626), indicating a common representational structure.
For the 30 common actions, the pairwise distances were significantly correlated between B1 and B3 ( r = 0.626 , p = 9.724 × 10 – 49 , on the off diagonals of matrices in Fig. 1K and 1L )
Results, Mesoscale representation of body parts, paragraph 3; Figure 1K-LHow it will be judged
Methods (Representational structure analyses, last paragraph). Distances as in figs2-pairwise-distances-significant; upper-diagonal entries for the 30 common actions are correlated. The correlation type is not named; the Methods say "correlated" and the Results give r, so the readings are (a) Pearson, (b) Spearman. Reproduced when r is within tolerance and p < 0.05; the correlation type is a method detail, so the outcome is the one both readings give, and if they differ the claim is blocked as methods_underspecified.
- fig1mn-hierarchical-clustering
Hierarchical clustering of each participant's pairwise distance matrix shows the relations of actions within and across effectors.
Relative comparisons within and across effectors are shown in Fig 1M – N for both participants via hierarchical clustering of the pairwise distance matrix
Results, Mesoscale representation of body parts, paragraph 3; Figure 1M-NHow it will be judged
Methods (Representational structure analyses). Complete-linkage hierarchical clustering of each participant's distance matrix. The paper gives no measure or test of how clusters follow effectors, and states no grouping in words. The attempt shows both dendrograms. As a variant only, it reports the fraction of actions whose nearest neighbour in the dendrogram (first merge) is an action of the same effector, against 1000 permutations of effector labels; this measure is not the paper's and cannot decide the outcome. The claim is recorded as methods_underspecified, with the dendrograms and the variant shown.
- fig2-session-trial-counts
Across the daily sessions (10 in B1, 11 in B3) the average number of trials per session was 89 for OL, 62 for CL1 and 48 for CL2.
In B1 and B3, we respectively performed 10 and 11 sessions. For each session, the average number of OL trials was: 89 [95% C.I. 83 – 94], CL1 trials: 62 [95% C.I. 50 – 75] and CL2: 48 [95% C.I. 38 – 59]
Results, Design of the BCI experiments and decoders, paragraph 2How it will be judged
The counts are over the 21 daily-initialization sessions pooled across B1 and B3, as the sentence reads; a reading per participant is also shown. Reproduced when the three pooled means are within tolerance.
- fig2e-above-chance
Decoding accuracy was above chance in every OL, CL1 and CL2 session in both participants.
decoding accuracies across trials ( Fig. 2E ) were significantly above chance (binomial distribution, FDR corrected, α = 0.05 level, Fig. S4A – B ) for every single session in B1 and B3
Results, Discriminability of representations with closed-loop control, paragraph 1; Figure 2E; Figure S4A-BHow it will be judged
Methods (Action decoding in the perspective of a virtual 3D Jaco arm). OL accuracy is cross-validated, 10 random splits of the day's OL trials into 68% training, 12% validation and 20% test, with an MLP of one hidden layer of 120 units (leaky ReLU, dropout 0.4, L2 0.0001, Adam, batch 32, early stopping after six epochs without improvement in validation loss); each test trial is classified by the most frequent decode over the bins of its active period (max vote), and accuracies are averaged over splits. CL1 and CL2 accuracy is the max vote of the online decodes per trial. The p-value follows the Methods, the binomial distribution "for a decoder given the total number of trials (N) and number of trials successfully decoded (K)", that is Binomial(N, K/N), and p is its area from 0 to N/m successes (chance or worse), with m = 7; FDR (Benjamini-Hochberg) at 0.05 over all decoders (63, from 21 sessions and three contexts). The usual upper tail P(X >= K) under success probability 1/7 is shown as a variant. Reproduced when every session's three decoders pass FDR at 0.05.
- fig2e-cl2-vs-ol-accuracy
Decoding accuracy is higher in CL2 than in OL, by 11.04 percentage points on average.
decoding accuracies were higher during CL1 and CL2 as compared to OL ( Fig. 2E , non-parametric linear mixed effect models, LMM, CL2 vs. OL: difference of 11.04%, t ( 40 ) = 4.59 , p = 0.001 , and Wilcoxon signed rank test, WSRT, z = 3.945 , p = 7.98 × 10 – 5
Results, Discriminability of representations with closed-loop control, paragraph 1; Figure 2EHow it will be judged
Methods (Action decoding, paragraph on tests). Session accuracies as in fig2e-above-chance. A linear mixed model of accuracy on experiment type (OL or CL2, fixed effect) with a random intercept per participant, over the 42 session values (t with 40 degrees of freedom); its p is the fraction of 2000 permutations, labels shuffled within each participant, whose |t| is at least the observed |t|. A Wilcoxon signed rank test over the 21 paired sessions. Reproduced when the difference is within tolerance and both tests give p < 0.05; partially reproduced when the tests hold and the difference is positive but outside tolerance.
- fig2e-cl1-vs-ol-accuracy
Decoding accuracy is higher in CL1 than in OL, by 6.325 percentage points on average.
CL1 vs. OL: difference of 6.325%, LMM, t ( 40 ) = 2.4026 , p = 0.022 , and WSRT, z = 2.93 , p = 0.0033
Results, Discriminability of representations with closed-loop control, paragraph 1; Figure 2EHow it will be judged
As fig2e-cl2-vs-ol-accuracy, for CL1 against OL.
- fig2e-cl2-vs-cl1-accuracy
CL2 accuracy is higher than CL1 accuracy, by 4.713 percentage points on average.
CL2 had significantly greater mean decoding accuracies when compared to CL1 (difference of 4.713%, LMM, t ( 40 ) = 2.06 , p = 0.044 , and WSRT, z = 2.209 , p = 0.027 )
Results, Discriminability of representations with closed-loop control, paragraph 1; Figure 2EHow it will be judged
As fig2e-cl2-vs-ol-accuracy, for CL2 against CL1.
- figs4ef-ol-accuracy-stable
OL decoding accuracy shows no linear trend across days in either participant.
Notably, OL accuracies remained stable and consistently lower across-sessions even with training ( Fig. S4E – F , linear fit of chronological OL decoding accuracy over days, B1: p = 0.354 ; B3:
Results, Discriminability of representations with closed-loop control, paragraph 1; Figure S4E-FHow it will be judged
A claim of absence, judged on the paper's test (Methods, Action decoding, last sentence). Ordinary least squares of each session's cross-validated OL accuracy on session order, per participant; reproduced when the slope's p >= 0.05 in both participants, not reproduced when it is below 0.05 in either. The slopes are reported either way.
- figs4g-percent-change-accuracy
Expressed as percentage change per session, CL1 improves accuracy over OL by 9.4%, CL2 over OL by 16.3%, and CL2 over CL1 by 6.6%.
CL1 and CL2 had positively improved decoding performance relative to OL (CL1 increase relative to OL by 9.427 ± 3.89 % , t ( 20 ) = 2.423 , p = 0.029 ; CL2 increase relative to OL by 16.275 ± 5.47 % , t ( 20 ) = 2.975 , p = 0.011 ). CL2 in turn had greater improvements in decoding accuracy as compared as CL1 ( by 6.62 ± 2.637 % , t ( 20 ) = 2.519 , p = 0.014 )
Figure S4G legendHow it will be judged
Methods (Action decoding, paragraph on controls). Per session, 100 (CL - OL) / OL and 100 (CL2 - CL1) / CL1. A mixed model of the 21 values with an intercept and a random intercept per participant; its p is the fraction of 1000 bootstrap samples of each participant's mean-centered data whose |t| is at least the observed. The tolerance is the reported spread (the paper does not say whether it is SD or SE). Reproduced when all three means are within tolerance and all three p < 0.05.
- figs4hi-feature-importance
All three bands contribute to decoding, with delta and high gamma the most informative.
Offline analyses revealed that while all three neural features ( δ , β , γ H ) were important for decoding, δ and γ H were most informative
Results, Discriminability of representations with closed-loop control, paragraph 1; Figure S4H-IHow it will be judged
Methods (Importance of neural features). Trials from all daily sessions per participant, 15 random splits (the Methods print "70% of trials for testing, 15% validation and 15% testing"; the reading is 70% training), decoders on all three bands, each pair of bands and each band alone; held-out trial accuracy per split. The paper gives no test. Read as (i) all three bands together reach at least the median accuracy of every single band, and (ii) delta alone and high gamma alone each have a higher median accuracy than beta alone, in both participants. Reproduced when both hold in both participants.
- fig2i-distance-cl-vs-ol
Average pairwise Mahalanobis distance between actions in latent space is greater in CL1 and in CL2 than in OL, in both participants.
CL demonstrated greater average Mahalanobis distances (all pairwise comparisons) than OL ( Fig 2I , LMM, combining B1 and B3, CL1 vs. OL: t ( 40 ) = 10.74 , p = 1 × 10 - 3 , CL2 vs. OL: t ( 40 ) = 11.73 , p = 1 × 10 - 3 ). This main effect was also present within B 1 and B3 individually
Results, Changes in representational statistics with closed-loop control, paragraph 1; Figure 2IHow it will be judged
Methods (Autoencoder design; Analyses of representational statistics in latent space). Per session, one autoencoder on that day's active-period bins of OL, CL1 and CL2, all actions, labelled by the cued action; layers input, 48, 16, 2 (bottleneck), 16, 48, output, ELU, dropout 0.3, plus a softmax classifier on the bottleneck, loss the sum of reconstruction error and cross-entropy, Adam with learning rate 0.001, batch 32, split 80/10/10, early stopping after six epochs, Gaussian augmentation per band (Data augmentation). All bins of each context are projected to the bottleneck; for each pair of the 7 actions the Mahalanobis distance with the pooled covariance (C_a + C_b)/2; the session value is the mean over the 21 pairs. Tests as in fig2e-cl2-vs-ol-accuracy with 1000 permutations, and Wilcoxon signed rank per participant. Open details. (1) The reconstruction loss is not named; readings (a) mean squared error, (b) mean absolute error. (2) The amount of augmentation (M) is not given; readings (a) none, (b) training data doubled. Random seeds are a standard choice. Reproduced when the LMM gives p < 0.05 for both contrasts and the signed rank test gives p < 0.05 for both contrasts in each participant, with CL greater. (1) and (2) are method details, so the outcome is the one all four combinations of readings give; if they differ, the claim is blocked as methods_underspecified. The same rule holds for the claims that use these autoencoders.
- fig2i-distance-cl2-vs-cl1
Latent-space distances between actions are greater in CL2 than in CL1, in both participants.
Between the two closed-loop experiments, CL2 was characterized by greater mean distance separation between movements as compared to CL1 (LMM, Fig. 2I , t ( 40 ) = 3.7186 , p = 0.001 ). This effect was present in both B 1 and B 3 individually as well (WSRT, B1: z = 2.701 , p =
Results, Changes in representational statistics with closed-loop control, paragraph 1; Figure 2IHow it will be judged
As fig2i-distance-cl-vs-ol, for CL2 against CL1.
- fig2j-variance-cl-vs-ol
Mean neural variance of actions in latent space is lower in CL1 and CL2 than in OL, in both participants.
the mean neural variance (over movements) during both CL experiments were lower in comparison to OL ( Fig. 2J , LMM, CL2 vs. OL: t ( 40 ) = 9.427 , p = 0.001 , CL1 vs. OL: t ( 40 ) = 8.365 , p = 0.001 ), an effect that was present within B1 and B3 individually as well
Results, Changes in representational statistics with closed-loop control, paragraph 1; Figure 2JHow it will be judged
Methods (Analyses of representational statistics in latent space, variance paragraph). On the session autoencoders of fig2i-distance-cl-vs-ol, each action's variance is the product of the eigenvalues of its 2x2 latent covariance (its determinant), averaged over the 7 actions per session and context; tests as in fig2i-distance-cl-vs-ol. Open detail. Figure 2J says variances were scaled per subject, without saying how; this matters for the pooled LMM, not the per-participant signed rank tests. Readings (a) no scaling, (b) divided by the participant's maximum, (c) z-scored within participant. Reproduced when the LMM gives p < 0.05 for both contrasts and the signed rank tests give p < 0.05 in each participant, with CL lower. The scaling is a method detail, so the outcome is the one all three readings give; if they differ, the claim is blocked as methods_underspecified.
- fig2j-variance-cl2-vs-cl1
CL2 has lower latent-space variance than CL1, in both participants.
Comparison of the two CL experiments revealed that CL2 had lower neural variances across-sessions on average than CL1 in B1 and B3 (LMM, t ( 40 ) = 2.517 , p = 0.015 , WSRT, B1: z = 2.803 p = 0.0051 ; B3: z = 2.8451 , p = 0.0044 )
Results, Changes in representational statistics with closed-loop control, paragraph 1; Figure 2JHow it will be judged
As fig2j-variance-cl-vs-ol, for CL2 against CL1, with the same readings of scaling.
- figs5-channel-variance-reduction
At the channel level, the standard deviation of each band's activity falls from OL to CL1, from OL to CL2 and from CL1 to CL2 (high gamma by 32.6%, 38.4% and 10.5%; beta by 29.0%, 33.9% and 7.6%; delta by 31.0%, 36.1% and 7.2%).
we found that there were significantly greater reductions in variance when going from OL to CL1 across sessions (in terms of S.D., by - 32.59 % , t ( 20 ) = – 6.44 , p = 0.0001 ) , OL to CL2 (by – 38.39 % , t ( 20 ) = – 8.496 , p = 0.0001 ) and when going from CL1 to CL2 (by – 10.49 % , t ( 20 ) = – 3.969 , p = 0.002 )
Figure S5E-G legend; Results, Changes in representational statistics with closed-loop control, paragraph 1How it will be judged
Methods (Analyses of neural variance in high-dimensional channel-feature space). Latent samples of each session's autoencoder are passed through its decoder layers to reconstruct the features; for each channel, band and action, the SD over samples in each context; the percent change of SD for each transition; the session value is the median over channels and actions; a mixed model on the 21 session values (intercept, random intercept per participant) with p from 1000 bootstrap samples of centered data. Open detail. The Methods first describe the reconstruction, then "examined the variance of feature activity at each spatial channel", without saying whether the SD is taken on the reconstructed or the input (L2-normalized, Figure S5D) features; readings (a) reconstructed features, (b) input features. The two are different measures, so the outcome is the one both give; if they differ, the claim is blocked as methods_underspecified. Reproduced when the nine session-mean changes are negative with p < 0.05 and all nine are within tolerance; partially reproduced when all nine are negative and significant but some fall outside tolerance.
- figs5bc-somatotopy-preserved
Each action's average spatial map is preserved (highly correlated) between OL, CL1 and CL2 within a session.
Our result showed preserved somatotopy for each movement on average when transitioning from open-loop to closed-loop control
Methods, Analyses of neural variance in high-dimensional channel-feature space; Figure S5A-CHow it will be judged
Per session and action, the average reconstructed spatial map of each band is flattened and correlated (Pearson, the paper's "linear correlation") between contexts. The paper gives no threshold or test here. Its one number for a map it calls preserved is the average pairwise r = 0.88 of Figure 4B; with the 10% tolerance of the protocol this gives 0.79. Read as preserved when the median correlation over actions and sessions is at least 0.79 for each band and context pair in both participants; the full distributions are reported.
- figs6-nine-actions
With nine actions (two wrist rotations added) on two days in B1, accuracy is higher in CL1 (71.30%) and CL2 (80.86%) than in OL (65.60%), latent distances are larger and latent variances smaller in CL than in OL and in CL2 than in CL1.
decoding accuracies tended to be higher in closed-loop control blocks (CL1, mean decoding accuracy of 71.30%, CL2 mean decoding accuracy of 80.86%) as compared to open-loop control blocks (OL, cross-validated mean decoding accuracy of 65.60%)
Figure S6D-F legend; Results, Changes in representational statistics with closed-loop control, paragraph 1How it will be judged
Methods (Replicating results with a larger action set). Accuracies as in fig2e-above-chance with m = 9; the mean is over the two days. Distances are the 36 pairwise latent distances per day, pooled over both days (the legend says 36 pairs across the two days; the Methods say 36 pairs per day collated over both days, 72 values; readings (a) 72 values, (b) 36 values averaged over days) compared with signed rank tests; variances are the 9 actions' latent variances, pooled over the two days (18 values), compared with signed rank tests. Reproduced when the three accuracies are within tolerance and, under one reading, all six signed rank tests give p < 0.05 in the stated directions.
- fig2kl-ol-distance-stable
OL latent distances between actions show no trend across days in either participant.
while the average distances between movements in OL remained stable across-days (B1: β = 0.11 ± 0.137 , t ( 8 ) = 0.797 , p = 0.448 , Fig. 2 K ; B3: β = 0.65 ± 0.425 , t ( 9 ) = 1.536 , p = 0.158 , Fig. 2L )
Results, Changes in representational statistics with closed-loop control, paragraph 2; Figure 2K-LHow it will be judged
A claim of absence, judged on the paper's test (Methods, Analyses of representational statistics, trend paragraph). OLS of each session's mean OL latent distance on session order, per participant. Reproduced when p >= 0.05 in both participants; not reproduced when p < 0.05 in either. Slopes are reported either way.
- fig2kl-cl-distance-increase
CL latent distances between actions increase across days in both participants (slope 1.655 in B1, 2.371 in B3).
CL was characterized by a striking increase in separability (via a random intercept LMM over CL1 and CL2, B1: β = 1.655 ± 0.563 , t ( 18 ) = 2.937, p = 0.0087 , Fig. 2 K ; B3: β = 2.371 ± 0.495 , t ( 20 ) = 4.786 , p = 1 × 10 – 3 , Fig. 2L )
Results, Changes in representational statistics with closed-loop control, paragraph 2; Figure 2K-LHow it will be judged
Methods (trend paragraph). Per participant, a mixed model of session mean latent distance on session order over CL1 and CL2 together, with CL1 and CL2 as random intercepts; p is the fraction of 1000 permutations (day order shuffled within CL1 and within CL2) whose |t| is at least the observed. The slope is in the latent space's own units, which depend on each autoencoder; the values are read as approximate. Reproduced when both slopes are positive with p < 0.05 and within tolerance; partially reproduced when both are positive and significant but outside tolerance.
- figs6gh-highdim-trends
In the full feature space, too, CL distances increase across days while OL distances stay stable, in both participants.
A linear mixed effect (LME) model combining CL1 and CL2 showed a significant increase in the separation between movements across-days (magenta line with slope indicated, t ( 18 ) = 2.1213 , p = 0.048 )
Figure S6G-H legend; Results, Changes in representational statistics with closed-loop control, paragraph 2How it will be judged
As fig2kl-ol-distance-stable and fig2kl-cl-distance-increase, on the mean pairwise Mahalanobis distance of each session's features instead of latent values. Open detail. "Full high dimensional neural data" is not defined for B1; readings (a) the 96 decoder features (2x2 pooled, L2-normalized), (b) the 384 unpooled channel features. B3 uses its 759 features. The outcome is the one both readings give; if they differ, the claim is blocked as methods_underspecified. Reproduced when the CL slope is positive with p < 0.05 and the OL slope has p >= 0.05, in both participants.
- fig2mn-distance-predicts-accuracy
Across sessions, the mean latent distance between actions predicts the session's decoding accuracy (leave-one-out logistic regression, better than permuted data), in both participants.
These across-days increases in average distances were significantly predictive of improved decoding accuracies in B1 and B3 ( Fig. 2M and 2N , leave-one-out cross validated logistic regression with permutation B1: CV log loss of 0.51, p = 0.002 , B3: CV log loss of 0.392, p = 0.002 )
Results, Changes in representational statistics with closed-loop control, paragraph 2; Figure 2M-NHow it will be judged
Methods (paragraph on logistic regression). Per participant, the points are the sessions' OL, CL1 and CL2 values (the Methods say accuracy "from each individual session", and the figure plots each session's distance against its accuracy); logistic regression of accuracy (a proportion) on mean latent distance; leave-one-out log loss averaged over points; p is the fraction of 1000 permutations of the accuracies whose log loss is smaller. Open details. (1) Whether points are 30 and 33 (all three contexts) or 10 and 11 (one per session); readings (a) all contexts, (b) CL2 only, (c) mean of the contexts. (2) The fit with a fractional outcome; reading a binomial GLM with the proportion as response, unweighted. Reproduced when under one reading of (1) both p < 0.05 and both losses are within tolerance; partially reproduced when both p < 0.05 only.
- fig3c-manifold-stable
Daily autoencoder manifolds are similar across days (CKA); in B1, 80% of day pairs have all five layers significantly similar and the rest at least four; in B3, 83.64% have all five, 12.72% four and 3.64% three.
In B1, 80% of all pairwise comparisons had all 5 layers significantly similar; the remaining 20% had at least 4 of 5 similar layers. In B3, 83.64% of all pairwise comparisons had all layers similar; 12.72% had at least 4 of 5 similar layers and the remaining 3.64% had at least 3 of the 5 similar layers
Results, Stability of action representations, paragraph 1; Figure 3CHow it will be judged
Methods (Representational stability analyses). For each pair of days, linear CKA between the two autoencoders' activations at each of the five hidden layers (48, 16, 2, 16, 48 units), activations mean-centered; the null is 1000 CKA values after randomly shuffling the weights and biases of the two layers compared; p is the fraction of null values at least the observed; FDR at 0.05 over all layers and pairs. Open details. (1) Which samples are fed to both networks, which CKA needs to match; readings (a) the pooled samples of both days, (b) day X's samples through both networks and day Y's through both, averaging the two CKAs. (2) What "shuffled" permutes; readings (a) entries of the layer's weight matrix and bias vector are permuted, (b) the layer's units (rows) are permuted. There are 45 day pairs in B1 and 55 in B3, five layers each. Reproduced when both "all 5" fractions are within tolerance and no B1 pair has fewer than 4 similar layers; partially reproduced when at least 75% of pairs in each participant have at least 4 similar layers. (1) and (2) are method details, so the outcome is the one all four combinations give; if they differ, the claim is blocked as methods_underspecified.
- figs7b-refit-less-discernible
In B1, daily initialized ReFit-KF velocity commands are less discernible between targets in latent space than the discrete actions; feedforward commands are more discernible than all commands within 45 degrees.
daily initialized discrete representations had greater median Mahalanobis distances between targets across sessions as compared to daily initialized early feedforward velocity commands (first ~ 2.5 s , Mann Whitney U test p = 1.2 × 10 - 5 ) or all intended velocity commands to the target within a 45 degree tolerance
Figure S7B legend; Results, Stability of action representations, paragraph 2How it will be judged
Methods (Representational analyses of denovo BCI velocity commands). Five days of ReFit-KF data in B1, six target classes, daily autoencoders on OL, CL1 and CL2 bins, either the first 2.5 s of each trial or all bins whose decoded velocity is within 45 degrees of the target; the session value is the mean pairwise latent distance; CL1 and CL2 values are pooled and compared with the discrete sessions' CL values by Mann Whitney U. Reproduced when all three tests give p < 0.05 in the stated directions.
- figs7cd-refit-less-stable
In B1, the daily manifolds of ReFit-KF velocity commands are much less stable across days than those of discrete actions; most day pairs (65%) share no similar layer, feedforward commands never do, and commands within 45 degrees share 1.5 layers on average.
Specifically, the majority of the pairwise comparisons (65%, averaged for both types of commands) between individual days’ autoencoders had no significantly similar layers between them
Figure S7C-D legend; Results, Stability of action representations, paragraph 2How it will be judged
As fig3c-manifold-stable, with its readings, on the ReFit-KF daily autoencoders; the layer-wise CKA values of all day pairs are compared between command types by Mann Whitney U. Reproduced when the 65% and 1.5 values are within tolerance, no feedforward day pair has a similar layer, and the three tests give p < 0.05 with discrete highest. As in fig3c-manifold-stable, the outcome is the one all combinations of its readings give; if they differ, the claim is blocked as methods_underspecified.
- figs7i-ibid-straighter-than-refit
Trajectories to cardinal targets deviate less from the straight line under IBID than under the ReFit-KF.
Results revealed greater median error in trajectories when using ReFit-KF after closed-loop decoder adaptation as compared to IBID after closed-loop decoder adaptation (Mann Whitney U test, z = 6.967 , p = 3.2442 × 10 - 12 )
Figure S7I legend; Results, PnP control of a virtual robotic arm, paragraph 2How it will be judged
Methods (Robot kinematics analyses). Per trial, the square root of the summed squared differences between decoded and ideal straight-line positions over all time points, divided by the number of points; Mann Whitney U between IBID and ReFit-KF CL2 trials. Open detail. The Results name 48 IBID and 18 ReFit-KF trials on one testing day, but with 48 and 18 values a Mann Whitney z cannot exceed about 6.2, below the reported 6.967; the reported test therefore used more values (more trials or sessions, or per-time-point errors), which the paper does not say. The test is run on per-trial errors as the Methods define them, and the z and p are not expected to match. Reproduced when p < 0.05 with the ReFit-KF median higher.
- fig4b-day-differences
For right thumb OL trials in B1, the mean high-gamma map is preserved across three days (average pairwise r = 0.88), but channel-level high-gamma means over M1/S1 differ between days.
Even with a preserved neural map and somatotopy (average pairwise r = 0.88 ), there was little distributional overlap between days ( Fig. 4B , permutation t – tests show significant differences in γ H mean over trials, Day 1 vs. Day 4, t ( 11 ) = 17.87 , p = 1 × 10 – 4
Results, Plug-and-play control in presence of distributional drift, paragraph 1; Figure 4A-BHow it will be judged
Open details, as the paper names neither which three days (called "three random days"; the tests name days 1, 4 and 8), nor which channels form the M1/S1 rectangle, nor what the 12 observations behind t(11) are. Readings (a) days 1, 4 and 8 of the daily sessions, channels the dataset marks as M1/S1 hand area, a paired permutation t-test over 12 channels of trial-mean high gamma; (b) the same days, a two-sample permutation t-test over trials of the mean over those channels. Reproduced when under one reading the mean pairwise map correlation is within tolerance and day 1 differs from days 4 and 8 with p < 0.05; partially reproduced when only the differences hold. The day 4 against day 8 result is reported.
- fig4cd-day-decodable
The recording day can be decoded from held-out OL samples on a common multi-day manifold (90.96% in B1, chance 10%; 98.4% in B3, chance 9.09%).
In general, the recording day could be classified from held-out samples on the common manifold ( Fig. 4C – D B1: cross-validated decoding accuracy of 90.96% [90.68% - 91.24%], chance = 10%, B3: cross-validated decoding accuracy of 98.4% [98.24% - 98.54%], chance = 9.09%)
Results, Plug-and-play control in presence of distributional drift, paragraph 1; Figure 4C-DHow it will be judged
Methods (Across-day drifts in centroids). OL samples of all days (about 22k), 40 random splits of 68% training, 12% validation and 20% test; an autoencoder as in fig2i-distance-cl-vs-ol trained with day labels as additional class memberships; accuracy of the classification layer for the day of recording on test samples, averaged over splits. Open details. (1) How day labels enter; readings (a) the classifier layer has one unit per day, (b) one unit per day and action combination, scored on day. (2) Whether samples are split by sample or by trial; the Methods say samples, so (a) by sample, with (b) by trial shown as a variant. The interval tolerances are the paper's own and very narrow; the claim is reproduced when both accuracies are within their tolerance, and partially reproduced when both are within 10% of the reported values and far above chance. (1) is a method detail, so the outcome is the one both readings give; if they differ, the claim is blocked as methods_underspecified.
- fig4cd-demeaned-chance
After mean-centering each day's data, the recording day can no longer be decoded (B1 10.03%, B3 8.77%, not different from chance).
In contrast, centering and demeaning each day’ data before building a common manifold eliminated discernibility ( Fig. 4C , B1 decoding accuracy: 10.03% [9.71% - 10.37%], chance = 10%, WSR: p = 0.282 , Fig. 4D : decoding accuracy of 8.77% [8.617% - 9.138%], chance = 9.09%, WSR: p = 0.3 )
Results, Plug-and-play control in presence of distributional drift, paragraph 1; Figure 4C-DHow it will be judged
A claim of absence, judged on the paper's test. As fig4cd-day-decodable after subtracting each day's mean from its samples; a Wilcoxon signed rank test of the 40 split accuracies against chance (10% in B1, 9.09% in B3). Reproduced when p >= 0.05 in both participants; not reproduced when p < 0.05 in either. The accuracies are reported.
- figs8ab-day-decodable-trials
A two-layer MLP on the full feature space also decodes the recording day from held-out trials, significantly above chance in both participants.
A MLP in high-dimensional channel-feature space revealed similar significant trial-level decoding accuracies ( Fig. S8A – B )
Results, Plug-and-play control in presence of distributional drift, paragraph 1; Figure S8A-BHow it will be judged
Methods (Across-day drifts, paragraph on the MLP). Trials of all days split 70/15/15 by trial, 10 random splits; MLP with two layers of 64 units, leaky ReLU, dropout 0.4, trained on samples; trial-level day accuracy averaged over splits; p from the binomial (as in fig2e-above-chance) with the number of held-out trials and chance 1/10 (B1) or 1/11 (B3). Open detail. How a trial's samples give its label is not stated for this decoder; readings (a) max vote over the trial's samples, as for the other trial-level decoders (Methods, Action decoding), (b) the fraction of the trial's samples decoded correctly, averaged over trials. The outcome is the one both readings give; if they differ, blocked as methods_underspecified. Reproduced when p < 0.05 in both participants.
- fig4fg-sampling-days-increases-distance
Distances between actions on held-out days increase linearly with the number of prior days used to build the common manifold, in both participants.
we found that the pairwise distances between actions on a given day linearly increased as a function of the number of prior days used to build the common manifold (robust linear regression on median Mahalanobis distances, B1: slope b = 0.258 , z = 6.474 , p = 1.847 × 10 – 10 , Fig. 4F
Results, Plug-and-play control in presence of distributional drift, paragraph 2; Figure 4F-GHow it will be judged
Methods (Across-day drifts, paragraph on cumulative days). For k = 1 to n - 1 days, an OL autoencoder is built on days 1 to k (as in fig2i-distance-cl-vs-ol); each held-out day k+1 to n is projected and its mean pairwise latent distance taken; the median over held-out days is regressed robustly on k. Open details. (1) The robust weighting is not named; readings (a) bisquare (MATLAB robustfit default), (b) Huber. (2) The statistic is called z, which suggests a normal approximation on 9 or 10 points; the p is taken from the fit as reported by the software. (3) The printed slope is identical for B1 and B3, so it belongs to one of them and the paper does not say which; readings (a) it is B1's, (b) it is B3's. Reproduced when both slopes are positive with p < 0.05 and the slope of the participant named by reading (a) or (b) is within tolerance (either reading may match); partially reproduced when both are positive and significant only. (1) is a method detail, so the outcome is the one both weightings give; if they differ, the claim is blocked as methods_underspecified.
- fig4i-boundary-convergence
In B1, a boundary between two hand actions trained on more prior days generalizes better to a new day, with decoding accuracy rising by 5.5 points per added day.
in B1 trends in decoding performance assessed via linear regression, slope, b = 5.541 ± 0.966 , t ( 8 ) = 5.735 , p = 4.365 × 10 – 4
Results, Plug-and-play control in presence of distributional drift, paragraph 2; Figure 4IHow it will be judged
Methods (Across-day drifts, last paragraph). A two-layer MLP in the common latent space classifies left thumb against both middle fingers, trained on days 1 to k (k = 1 to 10) and tested on a day distinct from the 10 daily sessions; OLS of accuracy on k. Open details. The held-out day is not identified ("a new day distinct from the 10 daily experimental days"); the MLP width is not given; 64 units per layer, as for the day-decoding MLP, is a standard choice, not the paper's, and widths 32 and 128 are shown as variants. Reproduced when the slope is positive with p < 0.05 and within tolerance; if the widths give different outcomes, the claim is blocked as methods_underspecified.
- fig5abc-pnp1-b1
In B1's first PnP experiment (fixed MLP, 28 days) the average accuracy was 87.43%, time to target 1.195 s and bit rate 1.72 bps.
The average accuracy across the entire PnP period was 87.43% [84.0%–90.8%] and the average time to target was 1.195 s [1.087s – 1.33s] ( Fig. 5A – B ). This resulted in a mean bitrate of 1.72 bps [1.51 – 1.94]
Results, Plug-and-play control in presence of distributional drift, paragraph 3; Figure 5A-CHow it will be judged
Methods (Discrete PnP experiments; PnP experiments, performance metrics). A trial's output is the first non-zero output of the 5-bin mode filter (the action decoded in more than 50% of the accumulated above-threshold bins, ties giving none, OSF code runningMode.m); a trial is accurate when that output is the cued action, and its time is from trial start to that output; trials time out at 5 s. Bit rate per block is log2(C - 1) max(0, N_acc - N_fail) / T with C = 7, T the block's total time. Open details. (1) Whether the averages are over blocks or pooled over trials; readings (a) mean of block values, (b) pooled trials, (c) mean of daily values. (2) Timed-out trials count as failures with 5 s, the reading the bit-rate formula implies. Reproduced when under one reading of (1) all three values are within tolerance.
- fig5def-pnp2-b1
In B1's second PnP experiment (38 days, decoder updated with the first experiment's data) the average accuracy was 80.71%, time to target 0.875 s and bit rate 1.848 bps.
The average accuracy was 80.71% [77.6% – 83.5%] and the average time to target was 0.875 s [0.842 s – 0.917 s ] resulting in an effective mean bit rate of 1.848 bps [1.68 – 2.02]
Results, Plug-and-play control in presence of distributional drift, paragraph 3; Figure 5D-FHow it will be judged
As fig5abc-pnp1-b1, with its readings.
- fig5-pnp-experiments-equal
Bit rates do not differ between B1's two PnP experiments.
There were no significant differences in performance between the two PnP experiments (Mann-Whitney U test on bitrates z = – 0.9574 , p = 0.338 )
Results, Plug-and-play control in presence of distributional drift, paragraph 3How it will be judged
A claim of absence, judged on the paper's test. Mann-Whitney U on the block bit rates of the two experiments; reproduced when p >= 0.05, not reproduced when p < 0.05. The difference in medians is reported.
- figs8cde-pnp-stable
Daily median PnP accuracy shows no linear trend across days in either B1 experiment or in B3.
a linear fit to the daily median revealed stable performance across-days for both ( Fig. S8C – D , p = 0.897 for PnP Exp 1 and p = 0.11 for PnP Exp 2)
Results, Plug-and-play control in presence of distributional drift, paragraph 3; Figure S8C-EHow it will be judged
A claim of absence, judged on the paper's test. OLS of each PnP day's median accuracy (median over blocks) on day, per experiment. Reproduced when p >= 0.05 in all three experiments; not reproduced when p < 0.05 in any. Slopes are reported.
- fig5-pnp-b1-overall-bitrate
Over both B1 PnP experiments the mean effective bit rate was 1.8 bps, with a maximum of 3.351 bps.
Across both experiments, the mean effective bit rate was 1.8 bps [1.66 – 1.93] and the highest was 3.351 bps
Results, Plug-and-play control in presence of distributional drift, paragraph 3How it will be judged
As fig5abc-pnp1-b1, pooling the blocks of both experiments; the maximum is over blocks. Reproduced when both values are within tolerance under one reading.
- fig5ghi-pnp-b3
In B3's PnP experiment (39 days) the average accuracy was 84.9%, time to target 0.818 s and bit rate 2.22 bps (maximum 3.508 bps), and accuracy stayed stable across days.
The average accuracy was 84.9% [82.14% – 87.43%], and the average time to target was 0.818 s [0.789 s – 0.848 s ] with a mean bit rate of 2.22 bps [2.046 bps – 2.394 bps ] (maximum of 3.508 bps )
Results, Plug-and-play control in presence of distributional drift, paragraph 4; Figure 5G-IHow it will be judged
As fig5abc-pnp1-b1, with its readings. Stability is claim figs8cde-pnp-stable.
- fig6g-ibid-angular-error
Under IBID, the angular error between integrated and ideal diagonal velocities is centred on zero (mean 1.885 degrees, not different from zero), showing integrated off-axis control.
The mean of this circular distribution was 1.885 degrees [95% C.I. −1.44 degrees to 5.206 degrees] and not significantly different from zero (circular median, p = 0.348 )
Results, PnP control of a virtual robotic arm, paragraph 3; Figure 6GHow it will be judged
A claim of absence, judged on the paper's test (Methods, Robot kinematics analyses), the CircStat circular median test (circ_medtest) of the instantaneous angular errors between IBID velocities and the ideal diagonal velocity, over all time points and trials to diagonal XY targets; reproduced when p >= 0.05, not reproduced when p < 0.05. The circular mean and its interval are reported (expected value with the paper's interval). The Methods take the errors "across time and trials", so the test is run on all time points; excluding time points with zero velocity is shown as a variant only, since a claim of absence is judged on the paper's test as written.
- fig6g-velocity-covariation
PCA of velocities during diagonal reaches shows covariation between X and Y velocity, unlike the non-integrated null model whose components recapitulate the cardinal axes.
In contrast, data revealed significant velocity covariation due to the integration in IBID ( Fig. 6G , left, middle)
Results, PnP control of a virtual robotic arm, paragraph 3; Figure 6F-GHow it will be judged
Methods (Robot kinematics analyses). PCA of 2D (X, Y) velocities over the diagonal trials. The paper calls the covariation "significant" but gives no test, threshold or number for it, in the Methods or the legend. The attempt reports the PC loadings and angles. As a variant only, it reports whether the first component lies more than 20 degrees from both cardinal axes with both loadings above 0.3 in absolute value; this threshold is not the paper's and cannot decide the outcome. The claim is recorded as methods_underspecified, with the loadings shown.
- figs8hi-perspective-perturbation
B1 used the PnP decoder proficiently under changed visual perspectives, with 86% (allocentric) and 94% (left-handed) accuracy over 42 trials each.
We found that B1 could immediately and proficiently use the decoder even with significant visual perturbations, with 86% and 94% decoding accuracies when operating the PnP decoder from either an allocentric or left-handed reference frame respectively (42 trials each
Results, PnP control of a virtual robotic arm, paragraph 4; Figure S8H-IHow it will be judged
Trial accuracy as for the PnP experiments (fig5abc-pnp1-b1), over the 42 trials of each perturbation. Reproduced when both are within tolerance.
- figs8g-coactivation
Imagined coactivations of two actions are discernible from the single actions.
B1 was also able to access off-axis directions by coactivating multiple representations simultaneously
Results, PnP control of a virtual robotic arm, paragraph 3; Figure S8GHow it will be judged
Methods (Coactivations of representations). A decoder (3-layer MLP of 64 units, as for the PnP decoder) classifies right thumb or right hand, left leg, head and the coactivations; trial accuracy by max vote. The legend reports "significant discernability" without naming the test; the paper's test of decoder significance elsewhere (Methods, Action decoding) is the binomial test of fig2e-above-chance, used here with m the number of classes (five, right hand, left leg, head and the two coactivations, per the Methods). Reproduced when it gives p < 0.05.
- fig7de-mode-switch-learned
B1 reached 100% accuracy with the mode switch in the virtual reach-to-grasp and transport task after about one week of practice, with a median completion time of 45 s.
B1 gained rapid proficiency in using the mode-switch with 100% accuracy after ~1 week of practice ( Fig. 7D – E , median task completion time of 45s)
Results, Flexible control of reaching and grasping, paragraph 2; Figure 7D-EHow it will be judged
Over the 10 sessions of the task. Open detail. Which sessions count as "after ~1 week" and which trials the median time is over; readings (a) sessions from day 7 on, (b) the last session, (c) all sessions for the median time. Reproduced when under one reading the success rate is within tolerance and the median time is within tolerance.
- fig7ij-reach-rotate
With the fixed RNN, B1 completed the real-world reach and rotate task with a median success rate of 90% and completion times of 60.8 s.
The median success rate was 90% with task completion times of 60.8 s [54.76 s – 76.95 s ]
Results, Long-term control of grasping and object manipulation using the Kinova Jaco robot, paragraph 1; Figure 7I-JHow it will be judged
Methods (Robot reach-to-grasp metrics). A trial succeeds only when all phases are completed and the reoriented object is placed at the target; completion time runs from the start of robot control to gripper release. Open details. (1) The median success rate is over days or blocks; readings (a) per day, (b) per block. (2) Whether 60.8 s is a median or a mean of successful trials; readings (a) median, (b) mean. Reproduced when under one reading both values are within tolerance.
- fig7k-wall-first-half
In the wall task, B1 reached and grasped the object off the wall proficiently across the whole PnP window (median success 100%, mean 94.6%, median time 38.053 s).
B1 was able to perform the first part (reach and grasp the target off the wall center) proficiently across the entire PnP testing window and for all levels of approach difficulty ( Fig. 7K top, median success rate of 100%, mean of 94.6%, median time to reach and grasp object 38.053 s [26.825 s – 45.458 s ])
Results, Long-term control of grasping and object manipulation using the Kinova Jaco robot, paragraph 1; Figure 7K topHow it will be judged
As fig7ij-reach-rotate, for the first part of the wall task (object grasped off the wall), over all PnP days before recalibration, with the same readings of the median.
- fig7k-wall-full-decline
Full wall-task success declined after PnP day 35 (median 88.19% over the first 35 days against 69.05% over days 35 to 210), with longer completion times.
Here, we observed a drop in performance after PnP day 35 (first 35 days median accuracy of 88.19%, total task completion time 75.52 s [70.22 s – 85.326 s ]; days 35–210 median accuracy of 69.05%, total task completion time 104.854 s [94.648 s – 126.639 s ])
Results, Long-term control of grasping and object manipulation using the Kinova Jaco robot, paragraph 1; Figure 7K bottom, 7LHow it will be judged
As fig7ij-reach-rotate, for the full wall task. Open detail. Day 35 falls in both ranges as printed; readings (a) day 35 in the first period, (b) in the second. Reproduced when under one reading the four values are within tolerance and the later median success is lower; partially reproduced when the later success is lower but values fall outside tolerance.
- fig7kl-recalibration-restores
Recalibrating the RNN with about 8 min of data immediately restored full wall-task success to 100%, with a median completion time of 80.73 s.
Notably, recalibration of the RNN (data required for decoder update was ~8 min, Methods ), immediately improved the success rate to 100% with a median completion time of 80.73 s [51.277 s – 108.247 s ]
Results, Long-term control of grasping and object manipulation using the Kinova Jaco robot, paragraph 2; Figure 7K-LHow it will be judged
As fig7ij-reach-rotate, over the wall-task trials after recalibration (day 0 in Figure 7K). Reproduced when both values are within tolerance.
- rnn-training-data-amount
The biLSTM decoder was trained on about 14 hours and 48 minutes of open-loop and closed-loop data from B1's virtual and real-world hDoF experiments, before augmentation and segmentation.
By the point we got to training a LSTM, we had already collected roughly 14 hours and 48 minutes of experimental open-loop and closed-loop data (prior to data augmentation and segmentation) in the virtual 3D environment and with practice with the Kinova Jaco in the real-world using the MLP
Methods, RNN decoders based on spatiotemporal activity patterns, paragraph 2How it will be judged
The total duration of the trials used to train the RNN, before segmentation. Open detail. Which part of a trial counts; the RNN segments come from "the active portions of the task" (Methods), and the OSF code (get_lstm_features.m) takes task states 2 and 3 and pads from state 4; readings (a) the whole stored trial, (b) the active (cue and control) portion only. The dataset may also hold the later real-world fine-tuning trials, which the 14 h 48 min precede; when it labels contexts, the duration is shown with and without real-robot fine-tuning trials. Reproduced when under one reading the duration is within tolerance.
- figs3-spatial-pooling
In B1, 2x2 spatial pooling of the grid makes single-trial high-gamma activity more consistent and improves online decoding in pilot blocks.
A Mann Whitney U test revealed significant differences in median variance between pooled and unpooled channels ( z = 8.721 , p = 2.748 × 10 - 18 )
Figure S3C and S3F legendHow it will be judged
Methods (Spatial feature extraction). Variance over time and trials of binned (200 ms) high gamma during right thumb trials, per channel, for the 128 raw channels against the 32 channels after a 1 s running average and 2x2 pooling with stride 2; Mann Whitney U. The block accuracy comparison uses online pilot blocks, which only the online decodes can give. Reproduced when the variance test gives p < 0.05 with pooled lower; the accuracy part is reported when testable.
Pipeline
- Paper
- done (1 of 1)
- Claims
- done
- Claims check
- done (ready)
- Attempt
- todo
- Questions
- none
- Verified
- none