acceptodds
Under review as a conference paper at ICLR 2027

FILM: A Dataset of Flexible-Electrode Intracortical Recordings for Longitudinal Motor Decoding

Abstract

Reliable long-term use of intracortical brain–computer interfaces (iBCIs) depends on sustained recording quality and decoding performance across days. Flexible intracortical electrodes support chronic recordings, but the long-term stability of motor decoding from these signals remains insufficiently characterized. To address this gap, we introduce FILM (Flexible-Electrode Intracortical Recordings for Longitudinal Motor Decoding), a longitudinal dataset for evaluating motor decoding from chronic flexible-electrode recordings. FILM pairs neural features derived from 512-channel recordings with synchronized two-dimensional cursor kinematics from two non-human primates performing Center-Out and Random Target reaching tasks. FILM comprises 178 sessions organized into three subject–task subsets spanning up to 277 days. Using FILM, we characterize longitudinal behavioral and neural changes, evaluate within-day and cross-day decoding performance, and assess target-day recalibration. Mean velocity profiles for each movement direction show substantial similarity across days, while movement speed, trial duration, neural activity, and estimated velocity tuning vary. Within-day decoding performance remains generally high, whereas cross-day decoding performance without recalibration tends to decline as the interval between training and testing days increases. Recalibration with limited target-day data improves cross-day decoding. Compared with matched recalibration without historical weights, incorporating these weights improves mean performance at the smallest calibration budget, with gains diminishing as more target-day data become available. Retrospective analyses further associate larger source–target weight mismatch with smaller or negative gains. FILM provides a resource for studying longitudinal neural distribution shifts and data-efficient adaptation in chronic iBCIs.

Primary area:
datasets and benchmarks
Keywords:
dataset, intracortical brain-computer interfaces, neural data, neural decoding, neural non-stationarity

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_7NjfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper describes a dataset, FILM, which provides neural features from chronic flexible-electrode recordings paired with cursor kinematics. The data comprises 178 sessions from two non-human primates, who were implanted with 512 acquisition channels. The data were recorded over 32 to 277 days. The study suggests that decodability remains relatively high within the same day (R^2 ~0.45 to ~0.95, selected to be >=0.3) but decreases drastically across days without recalibration. The authors also discuss how much recalibration helps. The paper is written well, and it is clear that time and effort has been put into creating it. The authors also used various methods to analyze the dataset they collected. Though it also has various weaknesses.

Critical strengths
  • The paper is well written. There is attention to detail. The figures generally make sense.
  • Good introduction overall (though see a few comments in Weaknesses).
  • The dataset it presents appears valuable. It includes 111 sessions over 277 days, which is a long coverage for flexible electrodes in primates. The paper also includes both the spiking-band power and threshold-crossing rate in the standard NWB format, which makes it valuable and easy to use.
  • The authors are careful to do a proper train-test split for the decoding.
  • They also compare recalibration across sessions with and without historical weights.
  • Their limitations section is sincere and useful.
Critical weaknesses
  • The authors need to do a better job explaining exactly what advantages flexible electrodes have over rigid ones. They do have advantages in terms of scarring, single-neuron tracking, and high channel density at depth. But why is decoding stability with flexible electrodes a separate issue from that with rigid electrodes, such as Utah arrays? What kind of result would distinguish the two? For example, the single-neuron tracking is not relevant in this paper, as no spike sorting has been carried out on these data. The authors mention this in the limitations, but it still limits the value of the dataset.
  • Though recording from different brain regions in MB was a good idea, two of the recording sites (outside PMd) were different between MA and MB. This makes interpreting those data more difficult. It would have been better to record from the same regions in MA and MB. The long-term claims rely only on MA, while MB was only recorded for 23 sessions over 32 days.
  • Quite a bit of text is required to explain the dataset and how to interpret it, which does not work with a 9-page limit for an ICLR paper. That means that to understand the dataset and its value the reader needs to go through all 31 pages of text, with 22 of those 31 pages (71%) in appendices. This manuscript is therefore better suited for a journal article than a conference paper.
  • The authors note that sessions were kept only when both WF and GRU had R^2>=0.3, within-session cross validated. However, the authors do not report statistics on how many sessions were discarded, where they took place in the lust of sessions, etc. And the session omission procedure is described only in an appendix (C.1). This is problematic for a dataset meant to characterize long-term recording stability. The excluded sessions should also be released, so that anyone using the dataset could inspect them.
  • ICLR is an ML conference. The authors suggest their dataset as a benchmark for decoding and domain adaptation. But their only baselines are WF and GRU, with recalibration tested only on WF. The authors should do much more to make this database valuable for the ICLR audience. By way of example, the FALCON dataset they cite include many more baselines, from humans to primates to birds.
  • The overall upshot of the paper is not clear. The authors do not have a specific description of their contributions, which is uncommon in ICLR papers. The authors propose it as a dataset with some conclusions about within- and across-day results. But those are neither very surprising nor seem different from results with rigid electrode. Perhaps a comparison with studies of rigid electrodes would make this one more valuable. Put differently, the main findings are well established for Utah arrays, including good within-day decoding, declining cross-day transfer, recalibration helping, and historical weights helping most with little data. The novelty in this piece therefore needs to come from the electrode type. But there is no comparison across electrodes. The paper cites LINK. Running the same frozen-transfer and recalibration protocol there would be valuable in testing whether flexible electrodes behave differently.

Some minor issues:

  • Unless SBP was carried out only in PMd and TCR only in M1, it is confusing to use the same color for PMd and M1 distinction in Fig. 1A and between SBP and TCR in Fig. 1B.
  • In Fig. 1C, Vy at trial end seems to be >200 px/s, while Vx=0 exactly, which is strange. If the figure ends before the trial does, indicate that.
Questions for authors

You can find my questions in the Weaknesses section.

9 comments
AuthorsOpenReview ↗

Response to Weakness 1

Thank you for this comment. We agree that our study does not assess single-neuron tracking because we did not perform spike sorting. We used unsorted channel-level activity because it can retain movement information relevant to motor BCIs. In chronic recordings, changes in spike waveforms can make individual units harder to isolate and track. Events that still cross the detection threshold contribute to the channel's threshold-crossing rate (TCR), even when they cannot be reliably assigned to a single neuron. Prior work has shown that unsorted threshold crossings can support online cursor control. Longitudinal recordings have also shown that threshold-based decoding can remain effective despite declines in spike amplitude.

A practical question for both flexible-electrode and Utah-array recordings is how well a fixed decoder can use these channel-level signals to predict cursor velocity on later days. A larger pool of task-related channels may help preserve decoding performance if some channels become less informative over time. To examine this possibility, we compared FILM and LINK in terms of the number of task-related channels, the stability of their tuning across days, and the performance retained by a fixed decoder (Fig. 4 in the revised manuscript).

FILM (MA_CO) has a larger task-related channel pool, with a median of 167 of 512 recorded channels (32.6%), compared with 16 of 96 (16.7%) in LINK (Fig. 4A). Within FILM, cross-day decoding performance improves as more task-related channels are included. At a fixed input size of 96 channels, selecting channels by task relevance also outperforms random selection (Fig. 4B). These results show that the additional task-related channels provide useful information for cross-day decoding.

To compare tuning stability at the same selected channel count, we examined the 15 most task-relevant channels in each dataset. Preferred direction and modulation depth were more stable in the selected FILM channels than in the selected LINK channels over extended intervals (Fig. 4C).

In the full-channel comparison, we trained a WF decoder on one day and applied it to later days without updating its weights. Using the same protocol and TCR features in both datasets, FILM achieved higher cross-day decoding R2R^2 than LINK and retained a larger fraction of its training-day performance. This fraction also declined more gradually in FILM as the train–test interval increased (Fig. 4D).

For practical BCIs, these results motivate a concrete goal of extending the time over which a fixed decoder maintains useful performance without recalibration. FILM makes it possible to test whether selecting and combining task-related channels with stable movement tuning can extend that period. Its longitudinal recordings and evaluation protocol provide a basis for measuring progress toward long-term “plug-and-play” BCI decoding. We will revise the manuscript to make this connection explicit.

AuthorsOpenReview ↗

Response to Weakness 2

Thank you for this comment. We agree that the different recording sites in MA and MB make direct between-subject comparisons difficult, and that MB’s 23 sessions over 32 days do not support the same long-term conclusions as MA’s 111 sessions over 277 days.

Our long-term cross-day results are based on MA_CO, recorded from PMd, and do not rely on comparisons with MB. Within each subject and recording subset, we evaluate decoding across days using the same implanted array and channel identities. MB provides a shorter-term evaluation in a second subject, over 23 sessions spanning 32 days. Its recordings from PMd, PMv, and M1 also allow exploratory analyses across those regions, but differences in recording sites prevent us from interpreting these as controlled comparisons between MA and MB.

We will revise the manuscript to distinguish MA's long-term evidence from MB's shorter-term and exploratory contributions, and to present cross-subject transfer between MA and MB as a separate analysis rather than as evidence for the long-term findings.

AuthorsOpenReview ↗

Response to Weakness 3

Thank you for this comment. We agree that the main paper should communicate FILM’s contribution without requiring readers to work through a lengthy appendix. We will revise the main text to give a concise description of the dataset, explain the questions it enables, and present the key results that establish its value.

We will also review the appendices for material that can be shortened or consolidated, especially explanations that repeat the main text. The remaining appendices will provide the detail needed to assess and reproduce the work: dataset acquisition and processing methods, along with supplementary analyses and results. We will clarify their contents and point readers to the relevant sections from the main text. This should make the paper easier to follow within the ICLR page limit while retaining the supporting detail.

Reviewer_W77dOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FILM, a longitudinal dataset of flexible-electrode intracortical recordings and synchronized two-dimensional cursor kinematics from two non-human primates performing reaching tasks. It contains 178 sessions across three subject–task subsets, with 512 acquisition channels per subject and a recording span of up to 277 days. The dataset provides two neural feature types, threshold-crossing rate (TCR) and spiking-band power (SBP). The authors characterize changes in behavior, neural activity, and estimated velocity tuning across days, and benchmark within-day decoding, frozen cross-day transfer, and target-day recalibration using Wiener filter and GRU decoders. They find that within-day decoding remains relatively strong, whereas frozen cross-day transfer becomes less reliable over longer intervals. Limited target-day data improves transfer; in matched comparisons, anchoring recalibration to historical weights provides the largest average benefit when the calibration budget is small, although this benefit diminishes with more data and greater retrospective weight mismatch.

Critical strengths

1. A valuable longitudinal recording resource. FILM combines flexible intracortical arrays with multiple recording sites along each shank, high channel count, long recording spans, and synchronized behavior. It provides both TCR and SBP features. Datasets with this combination are difficult and costly to acquire, yet are important for evaluating motor decoders under realistic cross-day changes. The multiple sessions and consistent task data make FILM a useful resource for studying longitudinal neural variability and data-efficient adaptation.

2. The results reveal a useful dependence on calibration budget and retrospective weight mismatch. At the smallest budget, normalization alone can outperform zero-centered recalibration, while source-anchored recalibration achieves the highest mean performance; with more labeled data, both supervised approaches improve beyond normalization.

Critical weaknesses

1. Limited synthesis of the two neural features. Although the appendices extend the analyses to both SBP and TCR, the main text focuses on SBP from MA_CO. A concise, matched comparison of their cross-day stability and recalibration behavior would help readers understand which conclusions generalize across features and when one feature may be preferable.

2. Limited use of the anatomical coverage. M B_CO includes arrays in PMd, PMv, and M1, but the reported results largely aggregate channels. An exploratory array-level analysis could make this aspect of the dataset more informative, while acknowledging that the available data cannot establish a general hierarchy of stability across brain regions.

Questions for authors

1. Based on the analyses already presented in the appendices, how do SBP and TCR compare in cross-day stability, for example in performance loss relative to each feature’s within-day reference? Are the main recalibration conclusions consistent across both features?

2. For M B_CO, is there evidence of differences in tuning stability or cross-day decoding across the PMd, PMv, and M1 arrays? If such a comparison is too limited by the number of arrays or recording days, a brief discussion of that limitation would be helpful.

**Minor editorial note: ** The Ethics Statement appears to repeat the same approval and anonymization text; removing the duplication would improve the presentation.

2 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1

Table 1. Within-day and frozen cross-day decoding with SBP and TCR (mean ± SD of R2R^2).

SubsetDecoderSBP (WD)SBP (CD)TCR (WD)TCR (CD)
MA_COWF0.777 ± 0.069-1.357 ± 1.8000.759 ± 0.0430.135 ± 0.232
MA_COGRU0.758 ± 0.1110.065 ± 0.2120.757 ± 0.1110.211 ± 0.170
MA_RTWF0.748 ± 0.063-0.493 ± 0.5850.702 ± 0.0620.259 ± 0.147
MA_RTGRU0.707 ± 0.0730.127 ± 0.1190.701 ± 0.0700.215 ± 0.139
MB_COWF0.648 ± 0.041-0.862 ± 0.5760.610 ± 0.0500.325 ± 0.158
MB_COGRU0.792 ± 0.0360.053 ± 0.1590.791 ± 0.0400.393 ± 0.158

WD denotes within-day decoding, and CD denotes frozen cross-day decoding. For each subset and decoder, bold marks the higher mean R2R^2 between SBP and TCR, separately for WD and CD. Following the decoding protocol in the manuscript, inputs were z-scored and clipped to [−5,5][-5, 5].

Table 2. WF recalibration across subsets and neural features.

Subset / featureFrozenN (5 s)H (5 s)F (5 s)N (120 s)H (120 s)F (120 s)REF
MA_CO / SBP-1.1830.1580.2850.0510.2970.7530.7660.774
MA_CO / TCR0.1370.2740.3880.1410.3520.7420.7570.759
MA_RT / SBP-0.3970.5030.5400.2050.5420.7560.7520.738
MA_RT / TCR0.2640.4300.4870.2320.4520.7160.7180.702
MB_CO / SBP-0.8670.2880.327-0.0800.4350.6340.6200.640
MB_CO / TCR0.3270.3600.371-0.0180.4590.6080.6040.613

Values are mean held-out R2R^2, with equal weighting across target days. Figure 13 of the revised manuscript shows the corresponding 95% confidence intervals. Frozen denotes transfer without recalibration. N denotes normalization only, H denotes source-anchored recalibration, and F denotes zero-centered full recalibration. REF is the full-training within-day reference.

Thank you for suggesting a matched comparison of SBP and TCR in the main text. Table 1 above shows that SBP achieves slightly higher mean within-day R2R^2, while TCR shows a smaller mean drop in R2R^2 from its own within-day reference under frozen cross-day transfer. This pattern holds across all three subsets and both decoders.

The main WF recalibration findings hold for both features (Fig. 13 in the revised manuscript). Table 2 above summarizes the results at the smallest and largest calibration budgets. Normalization alone partially recovers performance in all six subset–feature combinations. At 5 s, source-anchored recalibration (H) exceeds zero-centered recalibration (F) by 0.234–0.407 in mean held-out R2R^2. At 120 s, the difference (H minus F) ranges from −0.015 to 0.014, and both methods perform close to their respective within-day references.

We will highlight this comparison in the revised Results. Under frozen transfer, TCR shows smaller losses relative to its own within-day reference. With 120 s of target-day calibration data, SBP achieves slightly higher mean held-out R2R^2 than TCR in all three subsets under both H and F.

AuthorsOpenReview ↗

Response to Weakness 2 and Question 2

Table 3. Array-level tuning similarity and velocity decoding in MB_CO (mean ± std).

FeatureArrayTuning cosine (WD)Tuning cosine (CD)WF R² (WD)WF R² (CD)GRU R² (WD)GRU R² (CD)
SBPPMd-10.344 ± 0.2010.130 ± 0.1170.075 ± 0.051-0.803 ± 0.3530.028 ± 0.023-0.001 ± 0.016
SBPPMd-20.956 ± 0.0190.870 ± 0.0660.627 ± 0.052-0.520 ± 0.4740.619 ± 0.0450.311 ± 0.137
SBPPMv0.192 ± 0.1370.110 ± 0.054-0.454 ± 1.990-1.127 ± 0.7680.011 ± 0.015-0.004 ± 0.004
SBPM10.678 ± 0.1190.422 ± 0.1500.438 ± 0.112-3.805 ± 3.0360.544 ± 0.0760.200 ± 0.098
TCRPMd-10.551 ± 0.1450.347 ± 0.1060.058 ± 0.039-0.064 ± 0.0590.036 ± 0.035-0.058 ± 0.022
TCRPMd-20.893 ± 0.0270.707 ± 0.1060.531 ± 0.0600.225 ± 0.1480.551 ± 0.0370.312 ± 0.099
TCRPMv0.625 ± 0.1300.335 ± 0.1460.061 ± 0.049-0.082 ± 0.0870.013 ± 0.019-0.017 ± 0.023
TCRM10.821 ± 0.0590.678 ± 0.1110.379 ± 0.0380.156 ± 0.1030.363 ± 0.0660.129 ± 0.059

WD denotes within-day and CD denotes cross-day. Within each feature, bold marks the highest mean across arrays in each column. Acquisition channels (1-based) are PMd-1 (1–128), PMd-2 (129–256), PMv (257–384), and M1 (385–512). Each array contributes 124 electrode channels after excluding four reference channels.

Thank you for this question. We compared the four MB_CO arrays using 124 electrode channels per array across 23 recording days spanning 32 days (Table 3). With two PMd arrays and one array each in PMv and M1, we interpret the observed differences at the level of the recorded arrays.

Among the four arrays, PMd–2 had the highest mean cross-day tuning cosine and decoding R2R^2 for both SBP and TCR. For SBP, the mean cross-day tuning cosine was 0.870 for PMd–2, 0.422 for M1, 0.130 for PMd–1, and 0.110 for PMv. In frozen cross-day decoding with SBP, mean GRU R2R^2 was 0.311 for PMd–2 and 0.200 for M1, and was near zero for PMd–1 and PMv. The contrast between PMd–1 and PMd–2 further shows that cross-day tuning similarity and decoding performance vary even between arrays within the same cortical area. We will include this exploratory array-level comparison in the revised manuscript.

Reviewer_a15gOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces FILM, a longitudinal dataset for studying cross-day motor decoding from flexible intracortical electrodes. It contains recordings from two non-human primates, 512 acquisition channels per subject. The dataset comprises 178 retained sessions with a maximum recording span of 277 days. It provides threshold-crossing and spiking-band features aligned with two-dimensional cursor kinematics during joystick-controlled center-out and random-target tasks. Baseline experiments using Wiener filter and GRU decoders show strong within-day decoding in the retained sessions but reduced cross-day generalization. Limited target-day recalibration improves performance, while the benefit of incorporating historical decoder weights depends on calibration budget and source–target mismatch.

Critical strengths

The dataset includes recordings from two non-human primates, with 512 acquisition channels per subject, including 16 reference channels. Across subjects, the arrays cover multiple motor-related cortical areas—PMd, PMv, and M1—providing a useful resource for studying longitudinal motor decoding across different subjects and recording configurations.

Critical weaknesses

The dataset’s incremental contribution over LINK is insufficiently established. LINK already provides longitudinal intracortical recordings paired with motor behavior over 1,242 days, substantially longer than FILM’s maximum span of 277 days. Although FILM adds flexible-electrode recordings, higher channel counts, and a second subject, the current analyses do not clearly establish what distinct scientific questions or decoding challenges these differences enable. A systematic comparison with LINK would strengthen the justification for the new dataset.

Baseline methods are limited to WF and GRU, it would be important to show how other advanced methods perform on this dataset.

Questions for authors
  1. session selection was based on WF and GRU decoding criteria, what if model is not well trained, won't this hurt the diversity of the dataset?
  2. How the across day performance change on the filtered samples based on WF and GRU decoding criteria?
  3. How differnet brain regions decoding performance changes across day?
6 comments
AuthorsOpenReview ↗

Response to Weakness 1

Thank you for pointing out that FILM's distinction from LINK needs to be clearer. We agree that LINK provides longer longitudinal coverage, so duration alone does not establish FILM's contribution. We have added a direct comparison of FILM (MA_CO) and LINK (Fig. 4 in the revised manuscript) to examine a different question: how do the size of the task-related channel pool and tuning stability relate to the preservation of behavior decoding across days?

FILM has a larger task-related channel pool: a median of 167 of 512 recorded channels (32.6%), compared with 16 of 96 (16.7%) in LINK (Fig. 4A). Within FILM, cross-day decoding performance (R2R^2) increases as more task-related channels are included (Fig. 4B), and selecting 96 channels by task relevance outperforms selecting 96 channels at random. These analyses test whether the additional task-related channels provide useful information for cross-day decoding, beyond the raw number of recorded channels.

To examine tuning stability while holding channel count fixed, we compared the top 15 task-relevant channels in each dataset. FILM showed greater preferred-direction and modulation-depth stability over extended intervals (Fig. 4C). Under the same WF protocol and TCR features, FILM also achieved higher cross-day R² than LINK, retained a larger fraction of source-day performance, and showed a more gradual decline as the train–test interval increased (Fig. 4D).

Together, these analyses examine two aspects of longitudinal decoding robustness: how performance scales with the number of task-related channels, and how well tuning and decoding are preserved within a fixed-size channel set. While LINK’s multi-year recordings are well suited for studying long-term drift and recalibration under a smaller channel pool, FILM pairs high channel redundancy with greater single-channel tuning stability. It therefore complements LINK by enabling systematic investigation of how channel redundancy and tuning stability interact to buffer against neural turnover across days.

Crucially, the combination of these two properties makes a qualitative shift possible: moving from frequent recalibration toward long-term, calibration-free ("plug-and-play") decoding with a fixed decoder. This establishes a distinct decoding challenge on FILM: how to design algorithms that select and integrate information from a large task-related channel pool so that a fixed decoder can sustain zero-shot accuracy over weeks to months without target-day calibration. We will revise the manuscript to clarify how FILM complements LINK and to make both this scientific question and the plug-and-play decoding challenge explicit.

AuthorsOpenReview ↗

Response to Weakness 2

Table 1. FILM baselines. Performance comparison of methods on FILM using TCR features without z-score normalization. Values are reported as mean ± standard deviation of R².

MethodClassMA_CO (≤ 30 d)MA_CO (All)MA_RT (≤ 30 d)MA_RT (All)MB_CO (All)
WFWD0.7403 ± 0.07350.6987 ± 0.23490.7069 ± 0.07870.6755 ± 0.10790.5584 ± 0.0685
GRUWD0.6654 ± 0.06120.6419 ± 0.13200.5728 ± 0.06170.5673 ± 0.06850.6371 ± 0.0547
WFZS0.3190 ± 0.2390-0.0673 ± 0.50980.0089 ± 0.5350-0.2237 ± 0.62570.3578 ± 0.1451
GRUZS0.5906 ± 0.19040.1043 ± 0.30650.6032 ± 0.14940.3840 ± 0.37700.6834 ± 0.1124
NDT2ZS0.7165 ± 0.11610.3194 ± 0.19680.4814 ± 0.18650.3879 ± 0.18940.7097 ± 0.1607
RAT-SNNZS0.4594 ± 0.28550.0519 ± 0.3333-0.0109 ± 1.2511-0.0424 ± 0.96700.5961 ± 0.1665
CycleGANFSU0.5966 ± 0.09560.3495 ± 0.19260.5177 ± 0.16820.4560 ± 0.17140.4934 ± 0.0835
NoMADFSU0.6736 ± 0.28280.2481 ± 0.41730.3024 ± 0.27010.2124 ± 0.27400.4918 ± 0.2548
SPINTFSU0.5493 ± 0.18340.2586 ± 0.25350.5127 ± 0.09850.3822 ± 0.18610.6872 ± 0.1172
GRUFSS0.8970 ± 0.04650.7641 ± 0.17670.8146 ± 0.06640.7636 ± 0.11040.7866 ± 0.0741
NDT2FSS0.8791 ± 0.04610.8381 ± 0.05860.7947 ± 0.04950.7597 ± 0.07460.8186 ± 0.0511
RAT-SNNFSS0.7189 ± 0.45140.6518 ± 0.38060.7882 ± 0.07340.7375 ± 0.14600.7749 ± 0.0724

WD denotes within-day decoding. ZS denotes zero-shot decoding (frozen cross-day transfer). FSU denotes few-shot unsupervised decoding. FSS denotes few-shot supervised decoding. FSU and FSS use the same one-minute calibration segment from each target session. In each column, bold indicates the highest mean R² within each of the ZS, FSU, and FSS classes.

Thank you for this suggestion. We have expanded the FILM benchmark from WF and GRU to seven methods by adding NDT2, RAT-SNN, CycleGAN, NoMAD, and SPINT. Table 1 summarizes their performance using TCR features. NDT2 and RAT-SNN are evaluated under zero-shot (ZS) and few-shot supervised (FSS) settings, while CycleGAN, NoMAD, and SPINT are evaluated under the few-shot unsupervised (FSU) setting. FSU and FSS use the same one-minute calibration segment from each target session, with different access to behavioral labels.

The added baselines provide several informative comparisons. Across the “All” columns, NDT2 achieves the highest mean zero-shot (R2R^2) on all three datasets. For example, on MA_CO, it achieves 0.3194, compared with 0.1043 for GRU and −0.0673 for WF. Under FSS, NDT2 achieves 0.8381 on MA_CO and 0.8186 on MB_CO, compared with 0.7641 and 0.7866 for GRU, respectively. Among the evaluated FSU methods, CycleGAN achieves the highest mean (R2R^2) on MA_CO and MA_RT, whereas SPINT performs best on MB_CO. These results broaden the baseline coverage and show how method performance varies across datasets and adaptation settings.

AuthorsOpenReview ↗

Response to Question 1

Thank you for raising this concern. We agree that a poorly fitted decoder can yield a low R2R^2 even when a recording contains useful movement-related information. Screening on WF and GRU performance could therefore reduce the diversity of sessions represented in the benchmark. In FILM, however, only two of 180 candidate sessions (1.11%) fell below the within-session cross-validated R2R^2 threshold of 0.3 for at least one decoder. Both were MA_CO recordings from December 5 and 6, 2023, during online testing and task-protocol adjustments. The first MA_CO longitudinal recording block began later, on December 25, 2023.

Our primary benchmark uses 178 sessions to evaluate cross-day decoding and recalibration. It includes every candidate session within each of the three MA_CO recording blocks, as well as all 44 MA_RT and all 23 MB_CO sessions. Screening removed no intervening session from any reported longitudinal recording period. Omitting the WF/GRU criterion within those periods would therefore yield exactly the same set of sessions for the reported analyses.

We will revise the main text to describe the screening criterion, report the exclusion count, and give the dates and recording context of the two excluded sessions. We will publicly release all 180 candidate sessions, with the two excluded from the primary benchmark clearly labeled, so users can inspect those recordings and apply their own inclusion criteria.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.