Skip to content

BACKGROUND.md

Latest commit

 

History

History
412 lines (354 loc) · 32.4 KB

BACKGROUND.md

File metadata and controls

412 lines (354 loc) · 32.4 KB
Raw

B2SS — Background & Literature

The scientific foundation for a conduction-velocity-modulated BCI decoder, with the proposal's own references verified against the primary sources and enriched with recent related work (2019–2026).

How this was built. Each citation in the proposal (v1.1, §11) was looked up, its actual finding extracted, and the proposal's claim about it adversarially fact-checked against the source. Newer work was then searched per theme. Every paper below was confirmed to exist; DOIs are given. Where the proposal overstates or mis-attributes a source, it is flagged in Citation accuracy notes — those are corrections to the proposal text, not to the science, which is sound.

1. Conduction velocity, the g-ratio, and the biophysics

The premise: action-potential speed on a myelinated axon is set by axon calibre and the g-ratio (inner axon diameter ÷ total fibre diameter). Rushton's classic cable analysis established that conduction velocity scales with fibre diameter and is maximised at an optimal g-ratio near 0.6 (Rushton, 1951) — the biophysical anchor for treating g-ratio as a velocity proxy. (Our b2ss/cv.py reproduces this optimum at g = e^(−½) ≈ 0.607 as a sanity check.)

Turning this into an in-vivo, per-tract number requires MRI. Stikov et al. (2015) gave the reference recipe for estimating the myelin g-ratio non-invasively by combining a myelin-volume map (e.g. magnetization transfer) with a fibre/neurite-density map from diffusion MRI. Drakesmith et al. (2019) then closed the gap to velocity: a 14-parameter sensitivity analysis of a myelinated-axon electrophysiology model showed that ~85% of the variance in conduction velocity is captured by just two MRI-measurable quantities — axon diameter and g-ratio — and that a simplified Rushton relation (v ≈ d·√(−ln g)) captures the dependence. Their whole-brain estimates put corpus-callosum CV at ~8–10 m/s, aligned with primate electrophysiology, but with a critical caveat: estimates are accurate (<5% error) only for large axons (diameter > 4 µm, g-ratio 0.6–0.85) and become unreliable for the sub-micron axons that dominate the CNS. This bounds where a decoder should trust its CV prior.

Berman, Filo & Mezer (2019, Modelling conduction delays…) is the specific source for modelling conduction delay from an MRI-measured g-ratio — the lineage of the proposal's CV = k·v(g) relation. (The proposal's reference list cites a different Berman 2019 paper by mistake; see the notes.)

The most direct validation to date arrived in 2025: Asadi et al. (bioRxiv) built a multivariate model predicting neurophysiological CV from 7T MRI microstructure (axon radius, axonal water fraction, extra-axonal diffusivity, T1) and validated it against ground-truth intracranial cortico-cortical evoked potential (CCEP) latencies — exactly the sEEG validation the B2SS proposal plans (Experiment 2b). Their model explained ~29% of the variance in measured CV. That number is a realistic ceiling for how much a structural CV prior can contribute, and a useful prior for B2SS's own expectations. Supporting the microstructure inputs, volume-electron-microscopy work (Abdollahzadeh et al., 2024) validates the diffusion-MRI "Standard Model" parameters against histology, and Barakovic-style cross-species analyses (2024) characterise how axon diameter and g-ratio co-vary across tracts — relevant because CV depends on both jointly, so they should not be treated as independent decoder inputs.

2. CV carries behaviourally-relevant information

The load-bearing empirical result for B2SS is Clark et al. (2022, eLife): in 217 healthy adults, the MR g-ratio of the parahippocampal cingulum (hence inferred CV) was associated with autobiographical memory recall — specifically the number of internal/episodic details. The effect was tract- and task-specific (present for real-life recall, absent for lab memory tests) and appeared driven by inner axon diameter rather than myelin. This is the first demonstration, at BCI-relevant sample size, that an MRI-derived CV of a specific tract is a real, individually-varying predictor of behaviour — the strongest existing support for feeding a per-subject, per-tract CV estimate into a decoder as a prior.

Magnitude, stated accurately: the association is a small correlation, r(211) = 0.18, p = 0.008 (≈ Cohen's d 0.37), and it is correlational, not predictive or causal. The proposal's "d ≈ 0.45" overstates it (see notes) — worth tracking, since B2SS's power analysis leans on this effect size.

3. CV is plastic: activity-dependent myelination

B2SS treats CV not as fixed hardware but as a variable that changes with learning and can be nudged (Experiment 5, ccPAS). The biology backs this:

  • Gibson et al. (2014, Science) — the seminal causal demonstration. Optogenetic stimulation of premotor cortex in awake mice drove OPC proliferation, oligodendrogenesis, and circuit-selective myelination, with an associated improvement in motor function; pharmacological blockade of oligodendrocyte differentiation abolished the gain. (Precisely: the readout was motor function — contralateral forelimb swing speed — via a drug, not a motor-learning task via a genetic knockout; see notes.)
  • Fields (2015, Nat Rev Neurosci) — the framing review establishing activity-dependent myelination as a distinct plasticity mechanism that tunes conduction and timing.
  • Pajevic, Basser & Fields (2014) and follow-ups model how myelin plasticity tunes conduction delays to shape oscillations and synchrony.
  • Recent computational work makes the mechanism concrete and transferable to decoder design: an Activity-Dependent Myelination learning rule where net CV scales with firing rate, acting as homeostatic timing control (Nat. Comput. Sci., 2022); an oligodendrocyte model that synchronises correlated spike trains by selectively speeding lagging axons, using only local signals, optimal at ~10–40 ms glial time-constants (eLife, 2023); and a large-scale model showing conduction delays are a high-gain control variable — a ~1 ms shift can move a gamma-band phase by tens of degrees (PNAS, 2020).
  • Human, in-vivo, during motor learning: longitudinal DTI + myelin mapping (Cerebral Cortex, 2024) shows corticospinal white-matter change during motor- skill training that precedes cerebellar grey-matter adaptation — direct evidence the CV substrate B2SS targets genuinely shifts in users, and that a decoder's inputs will be non-stationary over training.

4. Measuring and modifying corticospinal conduction: TMS-EEG & ccPAS

B2SS estimates CV non-invasively via TMS-EEG MEP latency and modifies it via paired stimulation. The supporting methods literature:

  • What an MEP latency actually indexes (J. Physiol., 2023): MEP latency is the sum of intracortical processing, corticospinal conduction, spinal integration, and neuromuscular transmission — so a CV pipeline must attribute latency shifts to conduction, not cortical excitability. This defines B2SS's target variable and its confounds.
  • ccPAS is gated by conduction delay (Neuromodulation, 2023): paired stimulation of the two motor cortices at asynchronies straddling the ~9 ms transcallosal delay flipped the sign of plasticity (14 ms strengthened, 4 ms weakened, 9 ms no change) — human spike-timing-dependent plasticity tuned to axonal conduction time. This is why B2SS sets its ccPAS inter-pulse interval from each subject's measured CV.
  • ccPAS review (Clin. Neurophysiol., 2023): synthesises how ISI-vs-delay choice determines direction/magnitude of connectivity change, and flags outcome variability and unstandardised ISI selection as the field's open problems — motivating per-subject conduction-time measurement over group-average ISIs.
  • Lazari et al. (2022, Cell Reports) — the proposal's plasticity anchor: dual-site Hebbian TMS produced an increase in an MRI myelin marker whose significant cluster overlapped the tract connecting the stimulated regions, measured 24 h later. (Precisely: the marker was magnetization-transfer saturation that increased, not qT2 that decreased; the effect was a brain–behaviour correlation, not a group-mean change; only a single 24 h timepoint was measured — see notes.)
  • i-TEP (Brain Stimulation, 2024): an immediate TMS-evoked EEG potential starting ~2 ms post-pulse may expose the earliest corticospinal volleys to EEG, a possible route to a real-time, EMG-free conduction feature for closed-loop use.
  • TMS safety (Rossi et al., 2009): the consensus guidelines B2SS's stimulation protocols operate within.

5. BCI decoding: state of the art and where B2SS sits

The decoding lineage B2SS extends and must be benchmarked against:

  • Intracortical motor BCIs — Hochberg et al. (2012, Nature): people with tetraplegia controlling a robotic arm for reach and grasp (BrainGate2), the clinical proof-of-concept B2SS's prosthetic task descends from.
  • Single-trial latent dynamics — LFADS (Pandarinath et al., 2018, Nature Methods): a sequential (variational) autoencoder that infers single-trial neural dynamics and substantially improves behavioural decoding over spike smoothing. The discrete-time RNN baseline every continuous-time method is measured against. (Proposal dates it 2017 with wrong volume/pages; see notes.)
  • Robustness to non-stationarity — Sussillo et al. (2016, Nat Commun): a multiplicative-RNN decoder, trained with data augmentation, stayed usable under recording-condition changes that crippled a Kalman filter — in 2 macaques, with largely simulated variability (electrode dropping, stale training data), and significant in only one animal (see notes). B2SS reframes structural CV variability as a further, previously-unmodelled source of variability.
  • Modern deep decoders — a high-performance speech neuroprosthesis (Nature, 2023) hit 62 words/min from intracortical speech cortex via an RNN + language model; on the EEG side, CNN-Transformer hybrids (CTNet, 2024; TCFormer, 2025) and EEG foundation models (LaBraM, ICLR 2024) define the current motor-decoding recipe; SPINT (2025) and POSSM (NeurIPS 2025) target the exact deployment regime B2SS aims at — causal, real-time, drift-tolerant decoding, with POSSM matching Transformer accuracy at ~9× lower inference cost. These set the accuracy/latency/generalisation bar and show where a CV term could be injected (the temporal/convolutional front-end).

6. Continuous-time models: why Transformer + Neural ODE

B2SS's architecture (encoder → CV gate → Neural-ODE readout) rests on:

  • Transformer (Vaswani et al., 2017) — self-attention sequence modelling; the encoder backbone, and (in B2SS) the surface the CV-derived τ modulates via a temporal attention mask.
  • Neural ODEs (Chen et al., 2018, NeurIPS best paper) — continuous-depth models parameterising the derivative of the hidden state. The natural place to inject a velocity-conditioned term: CV scales the integration time of the vector field.
  • Latent ODEs / ODE-RNNs (Rubanova et al., 2019) — continuous-time latent state between observations; the scaffolding for a decoder whose effective timescale is not constant.
  • NODE vs RNN, head-to-head (Sedler et al., 2023): Neural-ODE sequential autoencoders recover neural population dynamics and fixed-point structure at the true low latent dimensionality where RNNs fail — the strongest single argument for choosing a continuous-time vector field over a discrete recurrence for a velocity-modulated decoder.
  • Deployable latent dynamics — DFINE (Nat. Biomed. Eng., 2024) keeps nonlinear manifolds with a tractable, causal (Kalman-like) dynamics path robust to dropped channels; the template for real-time CV-tunable inference.

7. Closest analogues to B2SS

Two lines are the nearest existing work — worth watching and citing directly:

  1. Structure-constrained conduction-delay brain models. "Mapping Brain Lesions to Conduction Delays" (Human Brain Mapping, 2025) builds a personalised whole-brain oscillator model where conduction delays are set by white-matter structure (τ = distance/velocity + damage term), then inverts the delay parameter from empirical MEG — a published template for exactly B2SS's structure→delay→signal→fit loop, and evidence that delay location matters.
  2. CV-from-MRI validated against electrophysiology. Drakesmith et al. (2019) and Asadi et al. (2025) together are the methodological spine: MRI → CV, with the second validated against CCEP latencies (the B2SS Experiment-2b design) and quantifying the ~29% variance ceiling.

The gap B2SS fills remains real: no existing decoder uses an individually measured white-matter CV as a structural constraint on its temporal integration window. The pieces exist separately (CV-from-MRI, conduction-delay brain models, continuous-time decoders); B2SS is the first to combine them for prosthetic control.

8. Evidence base for the v2 software methods

The publication-grade upgrade (real-EEG benchmark, CV proxy, heterogeneous-CV ablation, uncertainty gate) rests on the following literature.

CV proxy from EEG (mu peak frequency). The proxy uses each subject's sensorimotor mu-rhythm peak frequency, estimated with a lightweight version of the Corcoran et al. (2018) restingIAF method (Welch PSD, 1/f removal, 7–13 Hz peak). The rationale that peak frequency indexes conduction speed comes from Nunez's global standing-wave theory (resonant frequency ∝ cortico-cortical conduction velocity / loop length, f ≈ nν/L) and from Valdés-Hernández et al. (2010), who found peak alpha frequency correlates with white-matter fractional anisotropy (N=222). Honest caveat (encoded in the code and docs): this link is real but weak-to-moderate, correlational, and mechanistically ambiguous — alpha/mu frequency is also governed by thalamocortical membrane dynamics, and IAF↔white-matter correlations are small and inconsistent across studies. The proxy is presented as an indirect surrogate, never a calibrated per-person CV.

CV heterogeneity → a constant window is suboptimal (the "information" case). Conduction velocity varies ~2× between human callosal tracts (prefrontal ~4.9 vs motor ~8.8 m/s; Wang et al. 2008; Caminiti et al. 2013) and up to ~20× within a tract (macaque pyramidal tract 5–94 m/s), so no single temporal window fits the fibre population — the premise of the heterogeneous-CV regime in data.make_heterogeneous. Learnable per-input delay models are the direct engineering precedent that per-context delays beat a global constant: DCLS-Delays (Hammouamri et al. 2023) and learnable axonal delays (Sun et al. 2023) reach SOTA on spiking benchmarks and their ablations show learned per-connection delays beat fixed/uniform ones. Honest caveat: genuinely fast CV modulation (ms–min) is activity/membrane-driven and modest; structural CV (g-ratio) is slow. Arousal does not reliably change axonal CV (Stoelzel et al. 2017 found none) — so B2SS frames its window as set by structural, per-tract/per-context CV, and treats any state-dependent latency as effective end-to-end latency, not axonal CV.

Uncertainty-aware gate. Drakesmith et al. (2019) show MRI-derived CV is trustworthy only for large axons (diameter > 4 µm, g 0.6–0.85) and unreliable for the sub-micron axons dominating the CNS. The gate therefore shrinks τ toward the population-average window when the CV estimate's uncertainty (bootstrap SD) is large — trusting the prior only where it is reliable.

Cropped-window training. Deep EEG decoders need many trials; PhysioNet gives only ~45 per subject. We use cropped-window training with trial-level aggregation (Schirrmeister et al. 2017, Deep learning with CNNs for EEG decoding), standard practice for small-trial EEG and a natural fit to B2SS's window-based design.

Baselines & bars. EEGNet (Lawhern et al. 2018) and CSP+LDA (the FBCSP family, Ang et al. 2012) are the reimplemented baselines. Competitive within-subject accuracy on BCI IV-2a (4-class) is ~77–83% (CTNet 82.5%, EEGNet ~67–77%, FBCSP ~68%); PhysioNet 2-class within-subject is inherently modest and high-variance (be skeptical of published >95%, which usually reflect train/test leakage).

9. Measured-CV data: does a public paired dataset exist?

The decisive experiment — does a measured conduction velocity improve decoding on the same subjects — needs a dataset pairing (A) decodable neural recordings with (B) a per-subject/per-tract CV correlate. A verified scan of OpenNeuro, DANDI, PhysioNet, HCP, and the microstructure archives found: no public dataset cleanly pairs a motor/BCI decoding task with a measured CV on the same subjects. The closest options, each a compromise on one axis:

  • CCEP on iEEG — Utrecht RESPect, OpenNeuro ds004080 (74 patients). Single-pulse stimulation gives direct per-subject cortical conduction velocity on the very electrodes recorded (van Blooijs et al., Nat Neurosci 2023). Strongest CV signal, but the "task" is stimulation, not behavior — tests "does measured CV predict neural responses," not motor decoding.
  • EEG + diffusion MRI — VEPCON ds003505 (20 subjects; visual-discrimination behavior + DWI/FA on the same people) and HCP Young-Adult MEG (95 subjects with a motor task + dMRI). Real behavior, but CV is a diffusion proxy, and the tasks are perceptual/coarse-motor.
  • F-TRACT CCEP atlas (780 patients) gives CV/delay norms but only group-level in common space — usable as a prior on plausible CV ranges, not per-subject data.
  • Microstructure-rich sets (MICA-MICs, LEMON) have g-ratio-grade CV proxies but no decodable task.

Bottom line: the exact pairing B2SS needs does not exist publicly — itself a clean justification for a dedicated acquisition (decoding + CCEP/dMRI CV in one cohort). Cheap prototypes today: ds004080 for "measured CV carries decodable information," VEPCON/HCP for "dMRI-CV proxy helps a behavioral decode." (Verified via the associated Scientific Data / Nature Neuroscience papers and HCP docs; OpenNeuro SPDX licenses default to CC0 but should be confirmed per version page.)

10. Where the shrinkage adapter sits — and why the novelty claim is narrow

CADENCE's surviving mechanism is one line: shrink each per-channel calibration estimate toward the source prior with weight w = n/(n+τ). That is not a new statistical idea, and a paper that presents it as one deserves to be rejected. The honest positioning:

  • Empirical Bayes / James–Stein. Shrinking a noisy per-unit estimate toward a pooled prior in proportion to the evidence behind it is the standard construction (James & Stein 1961; Efron & Morris 1975). n/(n+τ) is the conjugate-normal posterior weight with τ the prior's effective sample size. We use it off the shelf.
  • Normalization-statistic adaptation in TTA. Adapting a frozen network by recomputing normalization statistics on the target is the AdaBN/PTBN family (Li et al. 2017; Nado et al. 2020), and interpolating those statistics between source and target rather than replacing them is already known to help under small test batches — e.g. the α-blend in Schneider et al. (NeurIPS 2020, "Improving robustness against common corruptions by covariate shift adaptation"), and the prior-corrected variants that followed. Our w differs only in being set by n rather than tuned as a constant.
  • Euclidean/Riemannian alignment in BCI. Per-session re-centering of the input distribution is the standard transfer trick in EEG BCI (He & Wu 2020; Zanini et al. 2018), and MPA-style per-channel moment alignment is its intracortical analogue.

So what is left that is ours? Not the estimator. The contributions are (a) the measurement — a calibrated decomposition, with a positive control, showing that the real multi-session intracortical gap is not conduction timing; (b) the observation that on a sparse 96-electrode array the standard per-session standardiser does not merely degrade below ~200 calibration windows but diverges, and that the fix is a scale floor plus evidence-weighted shrinkage; and (c) the continual-stream iBCI protocol with a stability metric (regret vs No-Adapt) that actually separates methods. A reviewer who says "this is just empirical Bayes" is right about the estimator and wrong about the claim — but only if the paper says so first.

Citation accuracy notes

Corrections to the proposal text/reference list (the underlying science holds; these are about precision and attribution — several matter for the power analysis and methods):

# Reference Issue Accurate statement
1 Berman et al. (2019), corpus-callosum age/sex Wrong paper + wrong claim. That paper (actually 2018, NeuroImage 182:304–313) finds the callosal g-ratio is stable with age and shows no sexual dimorphism — the opposite of "varies with age and sex" — and contains no CV = k·v(g) formula. The CV(g) relation belongs to Berman, Filo & Mezer (2019), Modelling conduction delays in the corpus callosum using MRI-measured g-ratio, NeuroImage (doi:10.1016/j.neuroimage.2019.116001). Cite that for the formula.
2 Stikov et al. (2015) Wrong volume/pages: cited as 93:239–251. Actual: NeuroImage 118:397–405 (doi:10.1016/j.neuroimage.2015.05.023).
3 Clark et al. (2022) Effect size "d ≈ 0.45" is not in the paper and is inflated; "predicts" overstates a correlation. Small correlation r(211)=0.18, p=0.008 (≈ d 0.37); associated with, not predictive; CV is inferred from g-ratio, not measured. Matters for the H4/Exp-4 power analysis, which cites this d.
4 Lazari et al. (2022) "Decreased qT2 relaxation" — wrong metric and direction; "strictly within the tract" overstates a correlational result; persistence. Marker was magnetization-transfer saturation (MT), which increased; qT2 was never measured. Effect was a brain–behaviour correlation (peak p_corr=0.013), not a significant group-mean change; the cluster overlapped the connecting tract. Only one 24 h post-timepoint — persistence beyond 24 h untested.
5 Gibson et al. (2014) "Abolishes motor learning" via "blocking oligodendrocyte differentiation." Readout was motor function (contralateral forelimb swing speed on gait analysis), not a motor-learning task; the block was pharmacological/epigenetic (HDAC inhibitor TSA), not genetic. Causal link to motor function stands.
6 Sussillo et al. (2016) Phrasing implies documented long-term human decay reduction. 2 macaques, preclinical; variability largely simulated (electrode dropping, withheld "stale" data); robustness significant in one monkey (p<0.01), not the other (p=0.45); the MRNN is fixed/non-adaptive, framed as complementary to recalibration.
7 Pandarinath et al. (2017) (LFADS) Wrong year/volume/pages (2017, 14(12):1216–1224). Actual: 2018, Nature Methods 15:805–815 (doi:10.1038/s41592-018-0109-9). Science unaffected.
8 Zador et al. (2026) — Verified: arXiv:2604.18637 (NeuroAI and Beyond), submitted Apr 2026; preprint, not yet peer-reviewed.

References

Proposal references (verified).

  • Berman, S., West, K. L., Does, M. D., Yeatman, J. D., & Mezer, A. A. (2018). Evaluating g-ratio weighted changes in the corpus callosum as a function of age and sex. NeuroImage, 182, 304–313. doi:10.1016/j.neuroimage.2017.06.076 (note: 2018, not 2019; and see note 1 — not the source of the CV formula)
  • Chen, R. T. Q., Rubanova, Y., Bettencourt, J., & Duvenaud, D. K. (2018). Neural ordinary differential equations. NeurIPS 31. arXiv:1806.07366
  • Clark, I. A., Mohammadi, S., Callaghan, M. F., & Maguire, E. A. (2022). Conduction velocity along a key white matter tract is associated with autobiographical memory recall ability. eLife, 11, e79303. doi:10.7554/eLife.79303
  • Fields, R. D. (2015). A new mechanism of nervous system plasticity: activity-dependent myelination. Nature Reviews Neuroscience, 16(12), 756–767. doi:10.1038/nrn4023
  • Gibson, E. M., et al. (2014). Neuronal activity promotes oligodendrogenesis and adaptive myelination in the mammalian brain. Science, 344(6183), 1252304. doi:10.1126/science.1252304
  • Hochberg, L. R., et al. (2012). Reach and grasp by people with tetraplegia using a neurally controlled robotic arm. Nature, 485(7398), 372–375. doi:10.1038/nature11076
  • Lazari, A., Salvan, P., Cottaar, M., Papp, D., Rushworth, M. F. S., & Johansen-Berg, H. (2022). Hebbian activity-dependent plasticity in white matter. Cell Reports, 40(3), 110951. doi:10.1016/j.celrep.2022.110951
  • Pajevic, S., Basser, P. J., & Fields, R. D. (2014). Role of myelin plasticity in oscillations and synchrony of neuronal activity. Neuroscience, 276, 135–147. doi:10.1016/j.neuroscience.2013.11.007
  • Pandarinath, C., et al. (2018). Inferring single-trial neural population dynamics using sequential auto-encoders. Nature Methods, 15(10), 805–815. doi:10.1038/s41592-018-0109-9 (proposal dated 2017)
  • Rossi, S., Hallett, M., Rossini, P. M., Pascual-Leone, A., et al. (2009). Safety, ethical considerations, and application guidelines for the use of transcranial magnetic stimulation in clinical practice and research. Clinical Neurophysiology, 120(12), 2008–2039. doi:10.1016/j.clinph.2009.08.016
  • Rushton, W. A. H. (1951). A theory of the effects of fibre size in medullated nerve. The Journal of Physiology, 115(1), 101–122. doi:10.1113/jphysiol.1951.sp004655
  • Stikov, N., et al. (2015). In vivo histology of the myelin g-ratio with magnetic resonance imaging. NeuroImage, 118, 397–405. doi:10.1016/j.neuroimage.2015.05.023 (proposal cites 93:239–251)
  • Sussillo, D., Stavisky, S. D., Kao, J. C., Ryu, S. I., & Shenoy, K. V. (2016). Making brain–machine interfaces robust to future neural variability. Nature Communications, 7, 13749. doi:10.1038/ncomms13749
  • Vaswani, A., et al. (2017). Attention is all you need. NeurIPS 30. arXiv:1706.03762
  • Wandell, B. A., & Yeatman, J. D. (2013). Biological development of reading circuits is related to reading skills. PNAS, 110(36), 14576–14577. (cited in proposal; peripheral to B2SS)
  • Wolpert, D. M., Diedrichsen, J., & Flanagan, J. R. (2011). Principles of sensorimotor learning. Nature Reviews Neuroscience, 12(12), 739–751. doi:10.1038/nrn3112
  • Zador, A., Fellous, J.-M., Sejnowski, T., et al. (2026). NeuroAI and beyond: Bridging between advances in neuroscience and artificial intelligence. arXiv:2604.18637

Additional work surfaced (not in the proposal).

  • Berman, S., Filo, S., & Mezer, A. A. (2019). Modelling conduction delays in the corpus callosum using MRI-measured g-ratio. NeuroImage. doi:10.1016/j.neuroimage.2019.116001 (correct source for CV = k·v(g))
  • Drakesmith, M., et al. (2019). Estimating axon conduction velocity in vivo from microstructural MRI. NeuroImage, 203, 116186. doi:10.1016/j.neuroimage.2019.116186
  • Asadi, A., et al. (2025). Non-invasive prediction of conduction velocities in the human brain from MRI-derived microstructure features at 7 Tesla. bioRxiv. doi:10.1101/2025.10.28.685017
  • Abdollahzadeh, A., et al. (2024). Volume electron microscopy in injured rat brain validates white matter microstructure metrics from diffusion MRI. Imaging Neuroscience, 2, 1–20. arXiv:2310.04608
  • (2024). Interplay between MRI-based axon diameter and myelination estimates in macaque and human brain. arXiv:2407.02227 (Imaging Neuroscience).
  • Sedler, A. R., Versteeg, C., & Pandarinath, C. (2023). Expressive architectures enhance interpretability of dynamics-based neural population models. NBDT. arXiv:2212.03771
  • Rubanova, Y., Chen, R. T. Q., & Duvenaud, D. (2019). Latent ODEs for irregularly-sampled time series. NeurIPS 32. arXiv:1907.03907
  • Abbaspourazad, H., et al. (2024). Dynamical flexible inference of nonlinear latent factors and structures in neural population activity (DFINE). Nature Biomedical Engineering, 8(1), 85–108. doi:10.1038/s41551-023-01106-1
  • POSSM (2025). Generalizable, real-time neural decoding with hybrid state-space models. NeurIPS 2025. arXiv:2506.05320
  • Willett, F. R., et al. (2023). A high-performance speech neuroprosthesis. Nature, 620. PMID:36711591
  • CTNet (2024). A convolutional transformer network for EEG-based motor imagery classification. Scientific Reports, 14, 20237. PMC11364810
  • Jiang, W.-B., et al. (2024). LaBraM: Large Brain Model for learning generic representations with tremendous EEG data in BCI. ICLR 2024.
  • SPINT (2025). Spatial Permutation-Invariant Neural Transformer for consistent intracortical motor decoding. arXiv:2507.08402
  • TCFormer (2025). Temporal convolutional transformer for EEG-based motor imagery decoding. Scientific Reports.
  • (2022). Homeostatic coordination and up-regulation of neural activity by activity-dependent myelination. Nature Computational Science, 2(10), 665–676. doi:10.1038/s43588-022-00315-z
  • (2023). Oligodendrocyte-mediated myelin plasticity and its role in neural synchronization. eLife, 12, e81982. doi:10.7554/eLife.81982
  • (2020). Activity-dependent myelination: a glial mechanism of oscillatory self-organization in large-scale brain networks. PNAS, 117(24), 13227–13237. doi:10.1073/pnas.1916646117
  • (2024). Temporal dynamics of white and gray matter plasticity during motor skill acquisition. Cerebral Cortex, 34(8), bhae344. doi:10.1093/cercor/bhae344
  • (2023). Motor potentials evoked by transcranial magnetic stimulation: interpreting a simple measure of a complex system. The Journal of Physiology, 601(14), 2837–2851. doi:10.1113/JP281885
  • (2023). Targeted modulation of human brain interregional effective connectivity with spike-timing-dependent plasticity. Neuromodulation, 26(4), 745–754. doi:10.1016/j.neurom.2022.10.045
  • (2023). Can we manipulate brain connectivity? A systematic review of cortico-cortical paired associative stimulation effects. Clinical Neurophysiology, 154, 169–193. PMID:37634335
  • (2024). TMS of primary motor cortex elicits an immediate transcranial evoked potential (i-TEP). Brain Stimulation, 17(3). doi:10.1016/j.brs.2024.05.003
  • (2025). Mapping brain lesions to conduction delays: the next step for personalized brain models in multiple sclerosis. Human Brain Mapping, 46(7), e70219. doi:10.1002/hbm.70219

Cited for the v2 methods.

  • Corcoran, A. W., et al. (2018). Toward a reliable, automated method of individual alpha frequency (IAF) quantification. Psychophysiology, 55(7), e13064. doi:10.1111/psyp.13064
  • Nunez, P. L., & Srinivasan, R. (2006). Electric Fields of the Brain (global standing-wave theory; f ≈ nν/L). Oxford University Press.
  • Valdés-Hernández, P. A., et al. (2010). White matter architecture rather than cortical surface area correlates with the EEG alpha rhythm. NeuroImage, 49(3), 2328–2339. doi:10.1016/j.neuroimage.2009.10.030
  • Wang, S.-J., et al. (2008). Functional trade-offs in white matter axonal scaling. Journal of Neuroscience / Caminiti, R., et al. (2013), J. Neurosci. 33(36):14501–14511 (tract-specific callosal conduction velocities).
  • Hammouamri, I., et al. (2023). Learning delays in spiking neural networks using dilated convolutions with learnable spacings (DCLS-Delays). arXiv:2306.17670.
  • Sun, P., et al. (2023). Learnable axonal delay in spiking neural networks. (per-neuron learned delays beat global constants.)
  • Stoelzel, C. R., et al. (2017). Axonal conduction delays, brain state, and corticogeniculate communication. J. Neurosci. (arousal did not change axonal CV). PMC5490068.
  • Schirrmeister, R. T., et al. (2017). Deep learning with convolutional neural networks for EEG decoding and visualization (cropped training). Human Brain Mapping, 38(11), 5391–5420. doi:10.1002/hbm.23730
  • Lawhern, V. J., et al. (2018). EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces. J. Neural Eng., 15(5), 056013. doi:10.1088/1741-2552/aace8c
  • Ang, K. K., et al. (2012). Filter bank common spatial pattern algorithm on BCI Competition IV datasets 2a and 2b. Frontiers in Neuroscience, 6, 39.
  • Schalk, G., et al. (2004). BCI2000: a general-purpose brain–computer interface system (PhysioNet EEGMMI). IEEE TBME, 51(6), 1034–1043.

Author lists abbreviated where long; a few very recent entries are preprints — verify against the final peer-reviewed version before formal citation.