Pith. sign in

REVIEW 4 major objections 5 minor 64 references

SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SPINT claims that a decoder trained once on labeled sessions can transfer to new intracortical recording sessions by treating neurons as an unordered set and inferring each neuron's identity from a few unlabeled calibration trials, with…

desk verdict SPINT's identity embedding is a real idea, but the headline comparison to zero-shot baselines is not controlled—SPINT trains on multiple sessions while those baselines get one. read the letter →

arxiv 2507.08402 v1 pith:4BHYGFYT submitted 2025-07-11 q-bio.NC cs.LG

Trung Le , Hao Fang , Jingyuan Li , Tung Nguyen , Lu Mi , Amy Orsborn , Uygar Sümbül , Eli Shlizerman This is my paper · ORCID
classification q-bio.NCcs.LG
keywords intracorticalbrain-computerinterfacemotordecodingpermutationinvariancecross-sessiongeneralizationfew-shotunsupervisedadaptationtransformerneuralidentityembeddingpopulationnonstationarity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that an intracortical decoder can be made to work across recording sessions without the usual alignment or recalibration steps. SPINT treats each session's recorded neurons as an unordered set and learns a short identity vector for each neuron from a few unlabeled calibration trials; that vector is added to the neuron's activity as a context-dependent positional embedding, so a transformer can decode behavior while remaining indifferent to how units are ordered. The authors test this on three multi-session motor decoding datasets from human and non-human primates and report that it outperforms all zero-shot and few-shot unsupervised baselines without any test-time gradient updates. A sympathetic reader would care because long-term iBCI use is currently limited by the need to re-align decoders as electrode arrays drift and unit compositions change; this is a direct attempt to remove that burden.

What carries the argument

The load-bearing piece is a context-dependent positional embedding. A shared IDEncoder computes $E_i = \psi(\mathrm{pool}(\varphi(X_i^C)))$ from a unit's $M$ calibration trials, so the embedding is a function of that unit's recent spiking signature rather than a fixed learned slot. Adding $E_i$ to every activity window of unit $i$ produces identity-informed tokens, and cross-attention with learnable behavior queries decodes the last-step behavior. Because row permutations of units permute $E$ in the same way, the whole pipeline is invariant to unit order while still retaining unit-specific information; dynamic channel dropout, which varies the dropout rate each training iteration, keeps the model from overfitting to the exact population composition seen in training.

What would settle it

Swap the inferred identity vectors of two units in a held-out session before decoding and measure the cross-session $R^2$. If performance barely changes, identity inference is not the mechanism driving transfer and the central claim is falsified; if performance collapses to the no-identity baseline, the mechanism is doing the work.

Watch

Extended reading notes

Core claim

The paper's central claim is that cross-session iBCI decoding does not require aligning neural populations to a fixed template: a decoder can instead treat every session's recorded units as an unordered set and infer each unit's identity on the fly. Its identity encoder maps a few unlabeled calibration trials from a unit to a vector $E_i$, and adding that vector to the unit's activity window yields identity-informed tokens $Z = X + E$. Cross-attention over these tokens is permutation-invariant because the identity vectors permute with the units, and this is what lets a single trained model generalize to unseen sessions with different unit counts and membership. The authors report that SPINT outperforms all zero-shot and few-shot unsupervised baselines on three multi-session motor decoding datasets with no test-time gradient updates, and matches or beats some supervised baselines that had access to labeled held-out data.

Load-bearing premise

The load-bearing premise is that an identity encoder trained on earlier sessions can infer a useful identity vector for a unit it has never seen, from only a short unlabeled calibration trial. If that cross-session inference fails, the decoder loses all unit-specific information and degrades to a set-average aggregator.

Editorial extensions

If this is right

  • A decoder trained once can be deployed on a new session using only a handful of unlabeled calibration trials, removing the need for test-time labels or parameter updates.
  • Because input size and order are unconstrained, the same trained model can serve sessions with different numbers of recorded units, and dynamic channel dropout makes it tolerate shrinking populations.
  • The method's cross-session accuracy improves with more training days, so it is positioned to scale into a pretrained general-purpose decoder as larger multi-session datasets become available.
  • Latency measurements at about 0.13 to 0.14 times real time indicate the architecture is light enough for online iBCI use.
  • Ablations show that fixed positional embeddings break the transfer and that no identity information also breaks it, so both the permutation-invariance and the inferred identity are load-bearing design choices.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If identity inference transfers between sessions, a natural extension is cross-subject or cross-region decoding, because the architecture never assumes a fixed electrode layout; the paper only tests within single subjects.
  • The observed correlation between attention weights and firing-rate statistics suggests SPINT is effectively learning to attend to behaviorally informative units, which could be turned into an explicit electrode-selection or signal-quality tool.
  • The identity encoder is trained with behavior labels; a self-supervised version trained only on spike statistics could remove that labeling requirement, and SPINT's one-trial success on the largest dataset hints this is feasible.
  • A direct stress test would be to record the same physical neurons across sessions and check whether their inferred identity vectors stay stable; that would separate genuine unit-identity learning from mere session-specific adaptation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes SPINT, a permutation-invariant transformer for continuous intracortical motor decoding. Neural units are treated as an unordered set; a context-dependent IDEncoder computes unit identity embeddings from a short unlabeled calibration period, which are added to unit activity before cross-attention decoding, and a dynamic channel dropout regularizer is introduced. The authors evaluate on the FALCON M1, M2, and H1 datasets and report that SPINT outperforms zero-shot and few-shot unsupervised baselines in cross-session R2 while requiring no test-time gradient updates, together with ablations, scaling analyses, and latency measurements.

Significance. If the cross-session identity inference works as intended, SPINT's gradient-free few-shot adaptation and low inference latency would be practically valuable for long-term iBCI, and evaluation on EvalAI private splits is a strength. The permutation-invariance proof in Appendix A.2 is straightforward and appears correct. However, the headline comparative claim is currently supported only by an uncontrolled comparison: baselines receive a single held-in training session while SPINT receives all held-in sessions, and per-dataset output scaling is applied only to SPINT. These issues make the quantitative contribution uncertain pending re-analysis.

major comments (4)
  1. [Section 4.2 vs. Section 4.1/A.4.5] The comparison in Table 1 is not controlled for training-data volume. The ZS Wiener Filter and ZS RNN are explicitly 'fitted using a single held-in session' (Section 4.2), and the FSU baselines CycleGAN+WF and NoMAD+WF also start from a day-0 decoder, whereas SPINT is trained on all available held-in sessions (4, 4, and 6 sessions for M1, M2, and H1; Section 4.1 and A.4.5). The observed margin over these baselines may therefore reflect 3–6x more labeled training data rather than the permutation-invariant identity mechanism. Please retrain the baselines on the same held-in sessions as SPINT, or train SPINT on a single held-in session, and report the comparison; without this, the headline claim in Section 4.3 is not established.
  2. [Appendix A.4.2 and Table A3] SPINT's training uses per-dataset behavior output scaling factors (1, 0.2, and 0.05 for M1, M2, and H1, respectively) that are hand-tuned, while no equivalent scaling is reported for the Wiener Filter, RNN, or alignment baselines. Because the reported R2 is computed between the scaled predictions and the original targets, this factor changes the effective prediction scale and can influence cross-session R2. Please report how the factors were selected, apply the same output transformation to all baselines, and show that the Table 1 conclusions are invariant to this choice.
  3. [Section 4.3, Table 1] No significance tests are provided, and in several datasets the reported standard deviations overlap substantially (e.g., M2: SPINT 0.26 ± 0.13 vs. CycleGAN+WF 0.22 ± 0.06; H1: SPINT 0.29 ± 0.15 vs. WF ZS 0.16 ± 0.03). With only 3–7 held-out sessions per dataset, the claim 'outperforms all zero-shot and few-shot unsupervised baselines' requires paired or bootstrap significance testing across held-out sessions, or the conclusion should be softened.
  4. [Section 3.2 and 3.5] The central assumption that IDEncoder learned on held-in sessions generalizes to unseen neural units is not directly validated. The analysis in Appendix A.3 shows that attention scores correlate with firing-rate statistics, but this is also what one would expect if the network used firing rates as a proxy rather than stable unit identities. Please provide a direct test of identity transfer, for example same-unit consistency across sessions where such data are available, or an ablation with shuffled or random identities on held-out sessions. Without such evidence, it is unclear whether the model is performing context-dependent identification or simply set-pooling with firing-rate-based weighting.
minor comments (5)
  1. [Section 3.1] The notation X_{i,[C]} ∈ R^{T'} and the phrase 'M-shot variable-length trials lasting for T' timesteps' is ambiguous; please clarify whether T' is the total calibration length or the length after interpolation.
  2. [Figures 3 and 4] The figures report 'standard error of the mean' while Table 1 reports 'standard deviation'; please state the error-bar convention consistently in the captions and main text.
  3. [Section 4.4] 'with only one single trial' should be 'with only a single trial'; please also specify whether this trial is one of the M calibration trials used by IDEncoder.
  4. [Section 4.6] The evaluation protocol for 'randomly sample a subset' in Figure 3C is underspecified; please state the number of sampled subsets, whether sampling is across units or trials, and whether identity embeddings are recomputed for the subset.
  5. [Appendix A.3] The statement 'All p-values are less than 0.05' is reported without specifying the statistical test or any multiple-comparison correction; please provide details.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: SPINT's identity embedding is a trained encoder applied to calibration trials, and the held-out evaluation is self-contained; the main caveat is an experimental confound in baseline training data, not circularity.

full rationale

SPINT's derivation chain is self-contained. Equation 1 defines the identity vector as E_i = IDEncoder(X_i^C) with a learned encoder trained end-to-end on held-in labeled data; Equation 2 adds E to the window and Equation 3 decodes via cross-attention. At test time the encoder is frozen and E is computed from unlabeled calibration trials, so the predicted held-out behavior is not an input to the model at training time. The permutation-invariance claim (Proposition 1) follows from row-wise ID computation and softmax row normalization in Appendix A.2, not from a fitted value or from a prior result. The two self-citations ([40] STNDT and [63] closed-loop adaptation) are background references and are not load-bearing for the central claim. The Discussion candidly notes that end-to-end training of the unit identifier ties identity to behavior decoding; that is a generalization caveat, not a circular step. The main weakness is experimental: the ZS Wiener Filter and ZS RNN baselines were fitted using a single held-in session (Section 4.2), while SPINT trains on all held-in sessions (Sections 4.1 and A.4.5), so the comparative margin in Table 1 may be confounded by training-data volume. That is a correctness risk for the headline comparison, not a circularity. Behavior scaling factors and dynamic channel dropout are hyperparameters or regularization choices, not fitted predictions. No circular step was found.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper does not introduce new physical entities. The main free parameters are per-dataset hand-tuned scaling factors and hyperparameters. The load-bearing domain assumptions are that identity inference transfers across sessions and that permutation invariance captures the dominant nonstationarity.

free parameters (4)
  • Behavior output scaling factor = 1 (M1), 0.2 (M2), 0.05 (H1)
    Hand-set per dataset in Appendix A.4.2 to scale network outputs; affects the training target and is not derived from data or theory.
  • Observation window size W = 100 (M1), 50 (M2), 700 (H1)
    Hyperparameter from Appendix A.4.5, Table A3, chosen per dataset with no described selection procedure.
  • Hidden dimension = 1024 (M1), 512 (M2), 1024 (H1)
    Hyperparameter from Appendix A.4.5, Table A3, chosen per dataset with no described selection procedure.
  • Dynamic channel dropout rate = Uniform in [0,1] (distribution not explicitly stated)
    Section 3.4 says a rate is randomly sampled between 0 and 1 each iteration, but the distribution is not stated; this design choice is central to the regularization claim.
assumptions (5)
  • standard math Softmax row normalization is invariant to column permutation
    Appendix A.2 relies on softmax applied row-wise, so permuting columns before softmax and applying the same permutation to V leaves the output unchanged.
  • domain assumption Within a session, a neural unit has a stable functional identity inferable from M unlabeled calibration trials
    Section 3.2 defines E_i = IDEncoder(trialized calibration trials); the method assumes this ID is stable across the session and representative of the test period.
  • domain assumption The IDEncoder trained on held-in sessions generalizes to held-out units without fine-tuning
    Section 3.5 states the trained IDEncoder is reused at test time with no adaptation, so cross-session generalization of the identity mapping is load-bearing.
  • domain assumption Cross-session nonstationarity can be captured by treating populations as unordered sets
    The core design premise, stated in Sections 1 and 3.1, is that permutation and composition changes are the dominant failure mode and that permutation invariance is sufficient for transfer.
  • domain assumption FALCON benchmark private held-out splits and EvalAI evaluation provide unbiased performance estimates
    All cross-session numbers come from EvalAI private splits (Section 4.1); the review cannot independently verify these numbers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding." pith.science (2026). https://pith.science/paper/4BHYGFYT

@misc{pith2026250708402,
  author       = {Pith},
  title        = {Pith review of: SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4BHYGFYT}},
  note         = {Machine review of arXiv:2507.08402}
}
read the original abstract

Intracortical Brain-Computer Interfaces (iBCI) aim to decode behavior from neural population activity, enabling individuals with motor impairments to regain motor functions and communication abilities. A key challenge in long-term iBCI is the nonstationarity of neural recordings, where the composition and tuning profiles of the recorded populations are unstable across recording sessions. Existing methods attempt to address this issue by explicit alignment techniques; however, they rely on fixed neural identities and require test-time labels or parameter updates, limiting their generalization across sessions and imposing additional computational burden during deployment. In this work, we introduce SPINT - a Spatial Permutation-Invariant Neural Transformer framework for behavioral decoding that operates directly on unordered sets of neural units. Central to our approach is a novel context-dependent positional embedding scheme that dynamically infers unit-specific identities, enabling flexible generalization across recording sessions. SPINT supports inference on variable-size populations and allows few-shot, gradient-free adaptation using a small amount of unlabeled data from the test session. To further promote model robustness to population variability, we introduce dynamic channel dropout, a regularization method for iBCI that simulates shifts in population composition during training. We evaluate SPINT on three multi-session datasets from the FALCON Benchmark, covering continuous motor decoding tasks in human and non-human primates. SPINT demonstrates robust cross-session generalization, outperforming existing zero-shot and few-shot unsupervised baselines while eliminating the need for test-time alignment and fine-tuning. Our work contributes an initial step toward a robust and scalable neural decoding framework for long-term iBCI applications.

Figures

Figures reproduced from arXiv: 2507.08402 by the authors.

Figure 1
Figure 1. Nonstationarities in long-term iBCI. (A) Examples of iBCI systems in human and non-human primates. Spiking activity is recorded from multichannel electrode arrays together with behavior covariates, e.g., 7 degree-of-freedom robotic arm control or electromyography from the upper limb. Neural activity exhibits nonstationarities over recording sessions. (B) Systematic changes in neuron positions, including the introduc… view at source ↗
Figure 2
Figure 2. SPINT architecture. The model performs continuous behavioral decoding by predicting behavior covariates at the last timestep given a past window of activity from an unordered set of neural units. The universal Neural ID Encoder infers identities of the units using few-shot unlabeled calibration trials, while the cross-attention mechanism selectively aggregates information from the units to decode behavior. 3 Approac… view at source ↗
Figure 3
Figure 3. Scaling analyses. Cross-session performance of SPINT against number of calibration trials (A), training days (B), and population sizes (C) across M1, M2, and H1 datasets. Bars represent mean R2 across held-out sessions, whiskers represent standard error of the mean of R2 across held-out sessions. Class M1 M2 H1 Wiener Filter (WF) OR 0.06 0.08 0.14 RNN OR 0.04 0.04 0.08 NDT2 Multi OR 0.15 0.10 2.29 NDT2 Multi FSS 0.1… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Ablation Study. Analyses showing the critical roles of our proposed context-dependent ID against fixed positional embeddings (PE) and no positional embeddings (A), our dynamic channel dropout against no dynamic channel dropout (B). Results are shown across M1, M2, and …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 57 canonical work pages

  1. [1]

    Neuronal population coding of movement direction

    Apostolos P Georgopoulos, Andrew B Schwartz, and Ronald E Kettner. Neuronal population coding of movement direction. Science, 233(4771):1416–1419, 1986

    work page 1986

  2. [2]

    High-performance neuroprosthetic control by an individual with tetraplegia

    Jennifer L Collinger, Brian Wodlinger, John E Downey, Wei Wang, Elizabeth C Tyler-Kabara, Douglas J Weber, Angus JC McMorland, Meel Velliste, Michael L Boninger, and Andrew B Schwartz. High-performance neuroprosthetic control by an individual with tetraplegia. The Lancet, 381(9866):557–564, 2013

    work page 2013

  3. [3]

    High perfor- mance communication by people with paralysis using an intracortical brain-computer interface

    Chethan Pandarinath, Paul Nuyujukian, Christine H Blabe, Brittany L Sorice, Jad Saab, Fran- cis R Willett, Leigh R Hochberg, Krishna V Shenoy, and Jaimie M Henderson. High perfor- mance communication by people with paralysis using an intracortical brain-computer interface. elife, 6:e18554, 2017

    work page 2017

  4. [4]

    High-performance brain-to-text communication via handwriting

    Francis R Willett, Donald T Avansino, Leigh R Hochberg, Jaimie M Henderson, and Kr- ishna V Shenoy. High-performance brain-to-text communication via handwriting. Nature, 593(7858):249–254, 2021

    work page 2021

  5. [5]

    A high-performance speech neuroprosthesis

    Francis R Willett, Erin M Kunz, Chaofei Fan, Donald T Avansino, Guy H Wilson, Eun Young Choi, Foram Kamdar, Matthew F Glasser, Leigh R Hochberg, Shaul Druckmann, et al. A high-performance speech neuroprosthesis. Nature, 620(7976):1031–1036, 2023

    work page 2023

  6. [6]

    Long-term stability of neural prosthetic control signals from silicon cortical arrays in rhesus macaque motor cortex

    Cynthia A Chestek, Vikash Gilja, Paul Nuyujukian, Justin D Foster, Joline M Fan, Matthew T Kaufman, Mark M Churchland, Zuley Rivera-Alvidrez, John P Cunningham, Stephen I Ryu, et al. Long-term stability of neural prosthetic control signals from silicon cortical arrays in rhesus macaque motor cortex. Journal of neural engineering, 8(4):045005, 2011

    work page 2011

  7. [7]

    Long-term stability of cortical population dynamics underlying consistent behavior

    Juan A Gallego, Matthew G Perich, Raeed H Chowdhury, Sara A Solla, and Lee E Miller. Long-term stability of cortical population dynamics underlying consistent behavior. Nature neuroscience, 23(2):260–270, 2020

    work page 2020

  8. [8]

    Intra-day signal instabilities affect decoding perfor- mance in an intracortical neural interface system

    János A Perge, Mark L Homer, Wasim Q Malik, Sydney Cash, Emad Eskandar, Gerhard Friehs, John P Donoghue, and Leigh R Hochberg. Intra-day signal instabilities affect decoding perfor- mance in an intracortical neural interface system. Journal of neural engineering, 10(3):036004, 2013

    work page 2013

Show all 64 references
  1. [9]

    Intracortical recording stability in human brain–computer interface users

    John E Downey, Nathaniel Schwed, Steven M Chase, Andrew B Schwartz, and Jennifer L Collinger. Intracortical recording stability in human brain–computer interface users. Journal of neural engineering, 15(4):046016, 2018. 10

    2018

  2. [10]

    Making brain–machine interfaces robust to future neural variability

    David Sussillo, Sergey D Stavisky, Jonathan C Kao, Stephen I Ryu, and Krishna V Shenoy. Making brain–machine interfaces robust to future neural variability. Nature communications, 7(1):13749, 2016

    2016

  3. [11]

    Months-long high-performance fixed lstm decoder for cursor control in human intracortical brain-computer interfaces

    Thomas Hosman, Tsam Kiu Pun, Anastasia Kapitonava, John D Simeral, and Leigh R Hochberg. Months-long high-performance fixed lstm decoder for cursor control in human intracortical brain-computer interfaces. In 2023 11th International IEEE/EMBS Conference on Neural Engineering (...

    2023

  4. [12]

    From monkeys to humans: observation-based emg brain–computer interface decoders for humans with paralysis

    Fabio Rizzoglio, Ege Altan, Xuan Ma, Kevin L Bodkin, Brian M Dekleva, Sara A Solla, Ann Kennedy, and Lee E Miller. From monkeys to humans: observation-based emg brain–computer interface decoders for humans with paralysis. Journal of Neural Engineering, 20(5):056040, 2023

    2023

  5. [13]

    Preserved neural dynamics across animals performing similar behaviour

    Mostafa Safaie, Joanna C Chang, Junchol Park, Lee E Miller, Joshua T Dudman, Matthew G Perich, and Juan A Gallego. Preserved neural dynamics across animals performing similar behaviour. Nature, 623(7988):765–771, 2023

    2023

  6. [14]

    Neural data transformer 2: multi- context pretraining for neural spiking activity

    Joel Ye, Jennifer Collinger, Leila Wehbe, and Robert Gaunt. Neural data transformer 2: multi- context pretraining for neural spiking activity. Advances in Neural Information Processing Systems, 36:80352–80374, 2023

    2023

  7. [15]

    A unified, scalable framework for neural population decoding

    Mehdi Azabou, Vinam Arora, Venkataramana Ganesh, Ximeng Mao, Santosh Nachimuthu, Michael Mendelson, Blake Richards, Matthew Perich, Guillaume Lajoie, and Eva Dyer. A unified, scalable framework for neural population decoding. Advances in Neural Information Processing Systems, ...

    2023

  8. [16]

    Few-shot algorithms for consistent neural decoding (falcon) benchmark

    Brianna Karpowicz, Joel Ye, Chaofei Fan, Pablo Tostado-Marcos, Fabio Rizzoglio, Clayton Washington, Thiago Scodeler, Diogo de Lucena, Samuel Nason-Tomaszewski, Matthew Mender, et al. Few-shot algorithms for consistent neural decoding (falcon) benchmark. Advances in Neural Info...

    2024

  9. [17]

    A cryptography-based approach for movement decoding

    Eva L Dyer, Mohammad Gheshlaghi Azar, Matthew G Perich, Hugo L Fernandes, Stephanie Naufel, Lee E Miller, and Konrad P Körding. A cryptography-based approach for movement decoding. Nature biomedical engineering, 1(12):967–976, 2017

    2017

  10. [18]

    Stabilization of a brain–computer interface via the alignment of low-dimensional spaces of neural activity

    Alan D Degenhart, William E Bishop, Emily R Oby, Elizabeth C Tyler-Kabara, Steven M Chase, Aaron P Batista, and Byron M Yu. Stabilization of a brain–computer interface via the alignment of low-dimensional spaces of neural activity. Nature biomedical engineering, 4(7):672–685, 2020

    2020

  11. [19]

    Stabilizing brain-computer interfaces through alignment of latent dynamics

    Brianna M Karpowicz, Yahia H Ali, Lahiru N Wimalasena, Andrew R Sedler, Mohammad Reza Keshtkaran, Kevin Bodkin, Xuan Ma, Lee E Miller, and Chethan Pandarinath. Stabilizing brain-computer interfaces through alignment of latent dynamics. BioRxiv, pages 2022–04, 2022

    2022

  12. [20]

    Inferring single-trial neural population dynamics using sequential auto-encoders

    Chethan Pandarinath, Daniel J O’Shea, Jasmine Collins, Rafal Jozefowicz, Sergey D Stavisky, Jonathan C Kao, Eric M Trautmann, Matthew T Kaufman, Stephen I Ryu, Leigh R Hochberg, et al. Inferring single-trial neural population dynamics using sequential auto-encoders. Nature met...

    2018

  13. [21]

    A generalist intracorti- cal motor decoder

    Joel Ye, Fabio Rizzoglio, Adam Smoulder, Hongwei Mao, Xuan Ma, Patrick Marino, Raeed Chowdhury, Dalton Moore, Gary Blumenthal, William Hockeimer, et al. A generalist intracorti- cal motor decoder. bioRxiv, pages 2025–02, 2025

    2025

  14. [22]

    Plug-and-play stability for intracortical brain-computer interfaces: a one-year demonstration of seamless brain-to-text communication

    Chaofei Fan, Nick Hahn, Foram Kamdar, Donald Avansino, Guy Wilson, Leigh Hochberg, Kr- ishna V Shenoy, Jaimie Henderson, and Francis Willett. Plug-and-play stability for intracortical brain-computer interfaces: a one-year demonstration of seamless brain-to-text communication. ...

    2023

  15. [23]

    Using adversarial networks to extend brain computer interface decoding accuracy over time

    Xuan Ma, Fabio Rizzoglio, Kevin L Bodkin, Eric Perreault, Lee E Miller, and Ann Kennedy. Using adversarial networks to extend brain computer interface decoding accuracy over time. elife, 12:e84296, 2023. 11

    2023

  16. [24]

    Maximum likelihood estimation of cascade point-process neural encoding models

    Liam Paninski. Maximum likelihood estimation of cascade point-process neural encoding models. Network: Computation in Neural Systems , 15(4):243, 2004

    2004

  17. [25]

    Bayesian population decoding of motor cortical activity using a kalman filter

    Wei Wu, Yun Gao, Elie Bienenstock, John P Donoghue, and Michael J Black. Bayesian population decoding of motor cortical activity using a kalman filter. Neural computation , 18(1):80–118, 2006

    2006

  18. [26]

    Pre- diction of arm movement trajectories from ecog-recordings in humans

    Tobias Pistohl, Tonio Ball, Andreas Schulze-Bonhage, Ad Aertsen, and Carsten Mehring. Pre- diction of arm movement trajectories from ecog-recordings in humans. Journal of neuroscience methods, 167(1):105–114, 2008

    2008

  19. [27]

    Neural decoding of hand motion using a linear state-space model with hidden states

    Wei Wu, Jayant E Kulkarni, Nicholas G Hatsopoulos, and Liam Paninski. Neural decoding of hand motion using a linear state-space model with hidden states. IEEE Transactions on neural systems and rehabilitation engineering, 17(4):370–378, 2009

    2009

  20. [28]

    A point process framework for relating neural spiking activity to spiking history, neural ensemble, and extrinsic covariate effects

    Wilson Truccolo, Uri T Eden, Matthew R Fellows, John P Donoghue, and Emery N Brown. A point process framework for relating neural spiking activity to spiking history, neural ensemble, and extrinsic covariate effects. Journal of neurophysiology, 93(2):1074–1089, 2005

    2005

  21. [29]

    Extracting dynamical structure embedded in neural activity

    Byron M Yu, Afsheen Afshar, Gopal Santhanam, Stephen Ryu, Krishna V Shenoy, and Maneesh Sahani. Extracting dynamical structure embedded in neural activity. Advances in neural information processing systems, 18, 2005

    2005

  22. [30]

    Spatio-temporal correlations and visual signalling in a complete neuronal population

    Jonathan W Pillow, Jonathon Shlens, Liam Paninski, Alexander Sher, Alan M Litke, EJ Chichilnisky, and Eero P Simoncelli. Spatio-temporal correlations and visual signalling in a complete neuronal population. Nature, 454(7207):995–999, 2008

    2008

  23. [31]

    Population decoding of motor cortical activity using a generalized linear model with hidden states

    Vernon Lawhern, Wei Wu, Nicholas Hatsopoulos, and Liam Paninski. Population decoding of motor cortical activity using a generalized linear model with hidden states. Journal of neuroscience methods, 189(2):267–280, 2010

    2010

  24. [32]

    Gaussian process based non- linear latent structure discovery in multivariate spike train data

    Anqi Wu, Nicholas A Roy, Stephen Keeley, and Jonathan W Pillow. Gaussian process based non- linear latent structure discovery in multivariate spike train data. Advances in neural information processing systems, 30, 2017

    2017

  25. [33]

    Gaussian-process factor analysis for low-dimensional single-trial analysis of neural population activity

    Byron M Yu, John P Cunningham, Gopal Santhanam, Stephen Ryu, Krishna V Shenoy, and Maneesh Sahani. Gaussian-process factor analysis for low-dimensional single-trial analysis of neural population activity. Advances in neural information processing systems , 21, 2008

    2008

  26. [34]

    Inferring latent dynamics underlying neural population activity via neural differential equations

    Timothy D Kim, Thomas Z Luo, Jonathan W Pillow, and Carlos D Brody. Inferring latent dynamics underlying neural population activity via neural differential equations. InInternational Conference on Machine Learning, pages 5551–5561. PMLR, 2021

    2021

  27. [35]

    Learnable latent embeddings for joint behavioural and neural analysis

    Steffen Schneider, Jin Hwa Lee, and Mackenzie Weygandt Mathis. Learnable latent embeddings for joint behavioural and neural analysis. Nature, 617(7960):360–368, 2023

    2023

  28. [36]

    Identifying distinct neural features between the initial and corrective phases of precise reaching using autolfads

    Wei-Hsien Lee, Brianna M Karpowicz, Chethan Pandarinath, and Adam G Rouse. Identifying distinct neural features between the initial and corrective phases of precise reaching using autolfads. Journal of Neuroscience, 44(20), 2024

    2024

  29. [37]

    Motor cortex activity across movement speeds is predicted by network-level strategies for generating muscle activity

    Shreya Saxena, Abigail A Russo, John Cunningham, and Mark M Churchland. Motor cortex activity across movement speeds is predicted by network-level strategies for generating muscle activity. Elife, 11:e67620, 2022

    2022

  30. [38]

    Representation learning for neural population activity with neural data transformers

    Joel Ye and Chethan Pandarinath. Representation learning for neural population activity with neural data transformers. Neurons, Behavior , Data analysis, and Theory, 2021

    2021

  31. [39]

    Seeing the forest and the tree: Building representations of both individual and collective dynamics with transformers

    Ran Liu, Mehdi Azabou, Max Dabagia, Jingyun Xiao, and Eva Dyer. Seeing the forest and the tree: Building representations of both individual and collective dynamics with transformers. Advances in neural information processing systems , 35:2377–2391, 2022

    2022

  32. [40]

    Stndt: Modeling neural population activity with spatiotemporal transformers

    Trung Le and Eli Shlizerman. Stndt: Modeling neural population activity with spatiotemporal transformers. Advances in Neural Information Processing Systems , 35:17926–17939, 2022. 12

    2022

  33. [41]

    universal translator

    Yizi Zhang, Yanchen Wang, Donato Jiménez-Benetó, Zixuan Wang, Mehdi Azabou, Blake Richards, Renee Tung, Olivier Winter, Eva Dyer, Liam Paninski, et al. Towards a" universal translator" for neural dynamics at single-cell, single-spike resolution. Advances in Neural Information ...

    2024

  34. [42]

    Neural encoding and decoding at scale

    Yizi Zhang, Yanchen Wang, Mehdi Azabou, Alexandre Andre, Zixuan Wang, Hanrui Lyu, The International Brain Laboratory, Eva Dyer, Liam Paninski, and Cole Hurwitz. Neural encoding and decoding at scale. arXiv preprint arXiv:2504.08201, 2025

    2025 arXiv

  35. [43]

    Adversarial domain adaptation for stable brain-machine interfaces

    Ali Farshchian, Juan A Gallego, Joseph P Cohen, Yoshua Bengio, Lee E Miller, and Sara A Solla. Adversarial domain adaptation for stable brain-machine interfaces. arXiv preprint arXiv:1810.00045, 2018

    2018 arXiv

  36. [44]

    Leveraging generative models for unsupervised alignment of neural time series data

    Ayesha Vermani, Il Memming Park, and Josue Nassar. Leveraging generative models for unsupervised alignment of neural time series data. In The Twelfth International Conference on Learning Representations, 2023

    2023

  37. [45]

    Extraction and recovery of spatio-temporal structure in latent dynamics alignment with diffusion models

    Yule Wang, Zijing Wu, Chengrui Li, and Anqi Wu. Extraction and recovery of spatio-temporal structure in latent dynamics alignment with diffusion models. Advances in Neural Information Processing Systems, 36:38988–39005, 2023

    2023

  38. [46]

    Deep sets

    Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Russ R Salakhutdinov, and Alexander J Smola. Deep sets. Advances in neural information processing systems , 30, 2017

    2017

  39. [47]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660, 2017

    2017

  40. [48]

    Set transformer: A framework for attention-based permutation-invariant neural networks

    Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In International conference on machine learning , pages 3744–3753. PMLR, 2019

    2019

  41. [49]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  42. [50]

    Permutation equivariance of transformers and its applications

    Hengyuan Xu, Liyao Xiang, Hangyu Ye, Dixi Yao, Pengzhi Chu, and Baochun Li. Permutation equivariance of transformers and its applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5987–5996, 2024

    2024

  43. [51]

    Enabling hyperparameter optimization in sequential autoencoders for spiking neural data

    Mohammad Reza Keshtkaran and Chethan Pandarinath. Enabling hyperparameter optimization in sequential autoencoders for spiking neural data. Advances in neural information processing systems, 32, 2019

    2019

  44. [52]

    A large-scale neural network training framework for generalized estimation of single-trial population dynamics

    Mohammad Reza Keshtkaran, Andrew R Sedler, Raeed H Chowdhury, Raghav Tandon, Diya Basrai, Sarah L Nguyen, Hansem Sohn, Mehrdad Jazayeri, Lee E Miller, and Chethan Pandari- nath. A large-scale neural network training framework for generalized estimation of single-trial populati...

    2022

  45. [53]

    Spatiotemporal distribution of location and object effects in reach-to-grasp kinematics

    Adam G Rouse and Marc H Schieber. Spatiotemporal distribution of location and object effects in reach-to-grasp kinematics. Journal of neurophysiology, 114(6):3268–3282, 2015

    2015

  46. [54]

    Spatiotemporal distribution of location and object effects in the electromyographic activity of upper extremity muscles during reach-to-grasp

    Adam G Rouse and Marc H Schieber. Spatiotemporal distribution of location and object effects in the electromyographic activity of upper extremity muscles during reach-to-grasp. Journal of neurophysiology, 115(6):3238–3248, 2016

    2016

  47. [55]

    Spatiotemporal distribution of location and object effects in primary motor cortex neurons during reach-to-grasp

    Adam G Rouse and Marc H Schieber. Spatiotemporal distribution of location and object effects in primary motor cortex neurons during reach-to-grasp. Journal of Neuroscience, 36(41):10640– 10653, 2016

    2016

  48. [56]

    Condition-dependent neural dimensions progressively shift during reach to grasp

    Adam G Rouse and Marc H Schieber. Condition-dependent neural dimensions progressively shift during reach to grasp. Cell reports, 25(11):3158–3168, 2018. 13

    2018

  49. [57]

    Real-time linear prediction of simultaneous and independent movements of two finger groups using an intracortical brain- machine interface

    Samuel R Nason, Matthew J Mender, Alex K Vaskov, Matthew S Willsey, Nishant Ganesh Kumar, Theodore A Kung, Parag G Patil, and Cynthia A Chestek. Real-time linear prediction of simultaneous and independent movements of two finger groups using an intracortical brain- machine int...

    2021

  50. [58]

    Ten-dimensional anthropomorphic arm control in a human brain- machine interface: difficulties, solutions, and limitations

    B Wodlinger, JE Downey, EC Tyler-Kabara, AB Schwartz, ML Boninger, and JL Collinger. Ten-dimensional anthropomorphic arm control in a human brain- machine interface: difficulties, solutions, and limitations. Journal of neural engineering, 12(1):016011, 2014

    2014

  51. [59]

    A brain-computer interface that evokes tactile sensations improves robotic arm control

    Sharlene N Flesher, John E Downey, Jeffrey M Weiss, Christopher L Hughes, Angelica J Herrera, Elizabeth C Tyler-Kabara, Michael L Boninger, Jennifer L Collinger, and Robert A Gaunt. A brain-computer interface that evokes tactile sensations improves robotic arm control. Science...

    2021

  52. [60]

    Falcon benchmark challenge

    EvalAI. Falcon benchmark challenge. https://eval.ai/web/challenges/ challenge-page/2319/evaluation, 2025. Accessed: 2025-05-12

    2025

  53. [61]

    Extrapolation, interpolation, and smoothing of stationary time series

    Norbert Wiener. Extrapolation, interpolation, and smoothing of stationary time series . The MIT press, 1964

    1964

  54. [62]

    Long short-term memory

    Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997

    1997

  55. [63]

    Closed-loop decoder adaptation shapes neural plasticity for skillful neuroprosthetic control

    Amy L Orsborn, Helene G Moorman, Simon A Overduin, Maryam M Shanechi, Dragan F Dimitrov, and Jose M Carmena. Closed-loop decoder adaptation shapes neural plasticity for skillful neuroprosthetic control. Neuron, 82(6):1380–1393, 2014

    2014

  56. [64]

    Adam: A method for stochastic optimization

    Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 14 A Appendix A.1 Within-session performance comparison We include the within-session performance comparison between SPINT and baselines in Table A1. This table is similar to T...

    2014 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.