PaperlayerView exact revision on arXiv

Source-linked AI summary

Neural State Prediction: Obstructing Shortcut Learning in EEG Foundation Models

Kieren Yu, Ziyang Liu, Chang Huang, Jintai Chen, Kaishun Wu

arXiv:2609.31167v1cs.AI

TL;DR

Masked EEG prediction can rely on positional cues and local correlations rather than distributed neural context, complicating the goal of transferable representations. Neural State Prediction jointly constrains latent targets and visible context; on EEG-FM-Bench it reaches 63.94 macro balanced accuracy, 2.35 percentage points above the strongest evaluated baseline.

  • Problem

    Masked prediction objectives do not distinguish transferable neural structure from positional cues or local correlations, so prediction can avoid integrating distributed context.

    Source: §1 · pp. 2–3

  • Method

    Neural State Prediction combines EMA-based latent supervision with identity-residualized targets and topology-separated context to constrain target content and available evidence.

    Source: §1 · pp. 2–3

  • Results

    63.94 macro balanced accuracy under EEG-FM-Bench full-parameter multi-task fine-tuning, 2.35 percentage points above the strongest evaluated baseline.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Transfer across 30 datasets and component analyses support jointly designing targets and context for EEG representation learning.

    Source: §1 · p. 3, §4 · p. 9

    Despite broad benchmark coverage, responsible use requires prospective validation across populations and acquisition settings, together with neural-data privacy safeguards.

    Source: §4 · p. 9

Abstract

from arXiv · show

EEG foundation models increasingly use masked prediction to learn from unlabeled recordings, but optimizing this objective does not ensure transferable neural representations. A central challenge is that stable positional cues and local correlations can make masked regions predictable without integrating distributed neural context. To reduce this reliance on low-information prediction paths, we introduce Neural State Prediction (NSP), a latent-predictive framework that constrains both the prediction target and the available context. NSP uses a Target Encoder updated by an exponential moving average (EMA) to define latent supervision. Identity residualization removes additive effects associated with channel identity and relative time from the targets, while topology-separated context excludes their immediate spatial and temporal neighborhood from the visible input. We pretrain NSP on 2.2 million EEG segments from TUEG and evaluate it across 30 downstream datasets spanning clinical diagnosis, sleep staging, emotion recognition, motor imagery, event-related potentials, cognitive-state decoding, and language retrieval. Under full-parameter multi-task fine-tuning on EEG-FM-Bench, NSP achieves 63.94 macro balanced accuracy across 14 datasets, exceeding the strongest evaluated baseline by 2.35 percentage points. Controlled component ablations assess the contribution of each mechanism, while matched context controls and held-out interventions characterize the role of context geometry, signal content, and positional information. Jointly designing latent targets and their context offers a promising direction for EEG foundation models that learn from distributed signal structure.

1 INTRODUCTION

EEG foundation models can learn transferable representations from broad unlabeled data, but masked prediction may rely on positional cues or local correlations instead of distributed neural context. Neural State Prediction jointly constrains latent targets and visible context, and achieves 63.94 macro balanced accuracy on EEG-FM-Bench, 2.35 percentage points above the strongest evaluated baseline.

1 INTRODUCTION

  • EEG foundation models pretrain reusable encoders on broad unlabeled corpora, whereas many task-specific EEG systems use limited labeled cohorts and transfer imperfectly across settings.The introduction situates masked prediction among contrastive, reconstruction, discrete or autoregressive, and latent-prediction paradigms.Source: §1 · pp. 1–2
  • Masked EEG prediction can use stable positional cues or local correlations without integrating distributed neural context, limiting what objective success establishes about transferable representations.The introduction identifies position-conditioned prediction and local interpolation as low-information paths.Source: §1 · pp. 2–3
  • NSP uses an EMA-updated Target Encoder, identity-residualized latent targets, and topology-separated context to constrain both what is predicted and the evidence available.The context excludes the target’s nearest-neighbor channels and adjacent time patches; the residualized target removes additive components associated with channel identity and relative time.Source: §1 · pp. 2–3
  • 63.94 macro balanced accuracy under EEG-FM-Bench full-parameter multi-task fine-tuning puts NSP 2.35 percentage points above the strongest evaluated baseline.The evaluation spans 30 downstream datasets, with matched ablations, signal-content interventions, and identity probes assessing components and representations.Source: §1 · p. 3

2 NEURAL STATE PREDICTION

NSP predicts position-matched latent states for masked EEG regions, jointly constraining the target representation and the context available to infer it. Its design combines identity-residualized targets, topology-separated evidence, and complementary prediction and state-regularization objectives.

2.1 OVERVIEW AND REPRESENTATION ENCODER

  • The encoders combine channel-wise spatial attention and within-channel temporal attention; masking withholds EEG content while retaining channel and coordinate information at masked positions.The Target branch processes the complete grid, while the Context branch masks target and guard positions.Source: §2.1 · p. 3

2.2 TARGET CONSTRUCTION: IDENTITY-RESIDUALIZED LATENT STATES

  • Target states are standardized and residualized by removing additive channel and relative-time components, leaving residual variation for position-matched prediction.Electrode coordinates remain available to the encoders and are not directly regressed out; gradients are stopped through the residualized targets.Source: §2.2 · p. 4
  • The Target Encoder is updated by EMA rather than optimizer gradients, providing a slowly varying target as the Context branch and predictor learn.After each Context Encoder update, the Target Encoder parameters follow a scheduled EMA.Source: §2.2 · p. 4

2.3 CONTEXT CONSTRUCTION: TOPOLOGY-SEPARATED EVIDENCE

  • Connected spatiotemporal target blocks are expanded with temporal and electrode-distance guards to prevent immediate-neighborhood interpolation from supplying their prediction.The guard changes available evidence but does not add prediction targets.Source: §2.3 · p. 5

2.4 MASKED LATENT PREDICTION

  • The prediction loss compares each masked-position prediction with the stopped, normalized residualized Target state at that same position, using context outside targets and guards.The loss is evaluated only over target positions.Source: §2.4 · p. 5

2.5 TRAINING OBJECTIVE AND PARAMETER UPDATES

  • A separate complete-input Context pass supports random-projection moment matching that encourages zero mean and unit variance without enforcing a Gaussian distribution.Gradients update the Context Encoder and predictor; the Target Encoder is then updated by EMA.Source: §2.5 · p. 5

3 EXPERIMENTS

Experiments compare NSP across two EEG benchmarks, adaptation protocols, component controls, training durations, label fractions, and representation visualizations. Results show benchmark performance and reveal how transfer varies with model components, task, and adaptation setting.

3.1 EXPERIMENTAL SETUP

  • The evaluation combines 16 selected NeuralBench datasets with 14 EEG-FM-Bench datasets, using each benchmark’s own protocols and comparison set.NeuralBench includes 14 classification tasks and two language retrieval tasks; EEG-FM-Bench evaluates four adaptation protocols.Source: §3.1 · p. 6

3.2 PERFORMANCE COMPARISON WITH BASELINES

  • 63.94 macro BAC under full-parameter multi-task fine-tuning on EEG-FM-Bench exceeds the strongest evaluated baseline by 2.35 percentage points.Figure 4 summarizes dataset-wise BAC across four adaptation protocols.Source: Figure 4 · p. 7, §3.2 · p. 7
  • 73.14 macro BAC across 14 NeuralBench classification tasks exceeds LaBraM (70.33), CBraMod (68.04), and REVE (67.30).NSP has the highest mean on nine tasks and ranks among the top two on twelve; ZuCo and Nieuwland2018 retrieval scores are reported separately.Source: §3.2 · p. 7, Table 2 · p. 7

3.4 TRAINING DURATION, CAPACITY, AND LABEL EFFICIENCY

  • Frozen transfer is generally strongest at the shorter pretraining budget, while full fine-tuning is maintained or improves on several tasks; macro BAC increases with capacity, but preferred scale varies by task.The duration comparison covers frozen and full-parameter single-task transfer, while the capacity comparison evaluates Tiny, Medium, and Large models.Source: Figure 5 · p. 8, §3.4 · pp. 8–9

3.3 COMPONENT AND INPUT/CONTEXT DEPENDENCE ANALYSIS

  • 55.70 macro BAC for Full NSP follows cumulative gains from block masking (45.75), EMA (51.88), and identity residualization (54.24); removing state regularization lowers it to 46.88.The component ablations use frozen single-task transfer with 10% labels.Source: Table 3 · p. 8, §3.3 · p. 8
  • Matched tests show residualization and topology separation each improve means, with larger gains together; interventions probe signal sensitivity and residual channel-identity information.Residualization adds 0.62 points with random exclusion and 1.52 with topology-separated exclusion; topology separation adds 1.06 without residualization and 1.96 with it.Source: §3.3 · p. 8, Table 3 · p. 8

3.4 TRAINING DURATION, CAPACITY, AND LABEL EFFICIENCY

  • With 10% versus 100% labels, TUAB and TUEV change by less than 1.5 percentage points, while SEED rises from 62.17 to 72.50 and PhysioMI from 31.03 to 58.50.The label-efficiency evaluation also includes 25% and 50% label fractions.Source: Figure 6 · p. 9, §3.4 · p. 9

3.5 REPRESENTATION GEOMETRY AND TASK-CONDITIONED SPATIAL EVIDENCE

  • HMC embeddings show more sleep-stage structure after pretraining and adaptation, though classes still overlap; TUEV Grad-CAM patterns differ across event classes.Figure 7 also depicts selected interchannel correlations.Source: Figure 7 · p. 9, §3.5 · p. 9

4 CONCLUSION

The conclusion presents NSP as a joint design of latent targets and visible context, with transfer results and component analyses supporting that direction. It also situates NSP among EEG foundation-model paradigms and calls for prospective validation and privacy safeguards.

4 CONCLUSION

  • NSP combines EMA-based latent prediction, identity-residualized targets, and topology-separated context to reduce additive positional effects and encourage distributed inference.The conclusion notes that future work could adapt exclusions to signal-dependent correlations.Source: §4 · p. 9
  • Despite broad benchmark coverage, responsible use requires prospective validation across populations and acquisition settings, together with neural-data privacy safeguards.Source: §4 · p. 9

A.1 EEG FOUNDATION MODELS; A.2 LATENT PREDICTION AND JEPA

  • EEG foundation models use reconstruction, contrastive, and autoregressive objectives as well as latent prediction; NSP’s distinguishing combination is target residualization with a guard against nearby spatial and temporal context.The cited latent-predictive EEG methods establish the family but do not combine these two design elements.Source: §A.1 · p. 15, §A.2 · p. 15

B.1 ARCHITECTURE

The appendix specifies NSP’s architecture, additive identity projection, pretraining configuration, and training algorithm. It describes the intended prediction target and context constraints alongside their implementation details and limitations.

B.1 ARCHITECTURE

  • The Large configuration uses 42.83M trainable Context-side parameters, eight CST blocks, 0.2-second patches, and a predictor with hidden width 768 and output width 512.The EMA coefficient follows a cosine schedule from 0.996 to 0.9999.Source: §B.1 · p. 16, Table 4 · p. 18

B.2 IDENTITY RESIDUALIZATION AS AN ADDITIVE PROJECTION

  • Identity residualization is the orthogonal residual after projecting latent features onto additive channel and relative-time effects, whose marginals vanish.It removes best additive main effects under current batch weighting, not nonlinear identity information or channel-time interactions, and does not imply channel invariance.Source: §B.2 · p. 16

C.2 HYPERPARAMETER CONFIGURATION

  • The default configuration uses 50% target masking, temporal spans of 2–10 patches, and spatial spans of 3–7 coordinate-near channels.Pretraining uses AdamW with learning rate 1 × 10^-4, global batch size 384, and 50,000 optimization steps.Source: §B.3 · p. 16, Table 5 · p. 18

B.3 PRETRAINING HYPERPARAMETERS

  • Pretraining maps heterogeneous TUEG recordings to 19 standard electrodes at 200 Hz, then forms non-overlapping 6-second segments; the collection contains 2.2 million examples.The spatial guard includes the two nearest channels, and the temporal guard includes adjacent patches.Source: §B.3 · p. 16

C.1 OBJECTIVE JUSTIFICATION AND TRAINING ALGORITHM

  • NSP changes both target content and visible evidence: residualization modifies the target, topology-separated context withholds nearby evidence, and state regularization guards against degenerate representations.The algorithm masks patch content while retaining positional cues, constructs targets, applies a moment regularizer, updates the Context model, and then updates the Target Encoder by EMA.Source: §C.1 · p. 17

C.3.1 DIAGNOSTIC EXPERIMENTS

Diagnostics track optimization and representation health, while benchmark and adaptation procedures specify task-specific evaluation settings. The appendix distinguishes pretraining input conventions from downstream preprocessing requirements.

C.3.1 DIAGNOSTIC EXPERIMENTS

  • Without state regularization, Context-state mean cosine reaches 0.9979 and effective rank falls to 113.6; Full NSP instead retains rank 470.9 with near-zero mean cosine.The control’s lower prediction loss coincides with highly aligned Context states and reduced effective rank.Source: §C.3.1 · p. 19, Figure 8 · p. 19

C.4 DOWNSTREAM BENCHMARK REQUIREMENTS

  • NeuralBench evaluates 16 selected datasets with single-task adaptation; 14 classification tasks contribute to BAC, while two language tasks use separate Top-5 retrieval accuracy.EEG-FM-Bench evaluates 14 datasets under four adaptation protocols.Source: §C.4 · p. 19

C.5 PREPROCESSING PROTOCOLS

  • Downstream preprocessing follows benchmark-specific task configurations rather than requiring every source recording to use the 19-channel, 200-Hz pretraining montage.The benchmark protocols may specify different channel counts, resampling, filtering, windowing, and event selection.Source: §C.5 · p. 20

C.6 DOWNSTREAM ADAPTATION

  • EEG-FM-Bench adaptation uses 30 epochs with a three-epoch warmup and a one-epoch encoder freeze; NeuralBench uses task-specific YAML settings and validation-based checkpoint selection.Source: §C.6 · p. 20

D DOWNSTREAM EVALUATION DETAILS

The evaluation covers 16 NeuralBench datasets and 14 EEG-FM-Bench datasets, using task-specific metrics and multiple adaptation protocols. NSP's results vary by task, with a 63.94 macro BAC on EEG-FM-Bench under multi-task full fine-tuning.

D.1 EVALUATION METRICS

  • Classification metrics are reported as percentages; BAC weights classes equally, AUROC measures threshold-independent ranking, and Top-5 accuracy is a separate retrieval metric.Median rank is unscaled, and lower is better.Source: §D.1 · pp. 20–21

D.2 BASELINE MODELS; D.3 ADAPTATION PROTOCOLS

  • NeuralBench uses a unified end-to-end recipe with six official baselines, while EEG-FM-Bench uses a distinct comparison set and four adaptation strategies.The protocols include full fine-tuning, frozen evaluation, and LoRA; checkpoint selection uses the corresponding validation metric.Source: §D.2 · p. 21, §D.3 · pp. 21–22

E.1.1 CLINICAL AND NEUROLOGICAL DISEASE

  • NSP leads BAC on three diagnostic datasets and remains competitive on CHB-MIT, where CBraMod leads BAC.NSP ranks second on CHB-MIT AUROC and AUCPR.Source: §E.1.1 · p. 22, Table 6 · p. 22

E.1.2 BCI AND SENSORIMOTOR DECODING

  • NSP leads BAC, weighted F1, and κ on Thielen2021, Scherer2015, and Wang2017, while REVE leads on Srisrisawang2024.NSP ranks second across all three metrics on Srisrisawang2024.Source: §E.1.2 · p. 23, Table 7 · p. 23

E.1.3 EVENT-RELATED AND EVOKED RESPONSES

  • NSP leads all three reported metrics on ERP CORE-N400 and Schreuder2010, ranks second on ERP CORE-N170, and trails leading baselines on ERP CORE-ERN.The reported metrics are BAC, AUROC, and AUCPR.Source: §E.1.3 · p. 23, Table 8 · p. 23

E.1.4 COGNITIVE STATE AND LANGUAGE DECODING

  • NSP ranks second on Zyma2019 BAC, leads all three Hinss2023 metrics, and records the best Top-1, Top-5, and median-rank results on both retrieval datasets.Retrieval metrics are reported separately and are not pooled into the 14-task BAC macro.Source: §E.1.4 · pp. 23–24, Table 9 · p. 24

E.2 EEG-FM-BENCH

  • 63.94 macro BAC across EEG-FM-Bench’s 14 datasets under multi-task full fine-tuning.The benchmark evaluates four adaptation strategies across 10 EEG paradigms.Source: §E.2 · p. 24

E.2.1 EMOTION RECOGNITION

  • Emotion results vary by dataset: multi-task full fine-tuning is NSP’s strongest setting on SEED and SEED-VII, while single-task full fine-tuning is strongest on SEED-V.NSP exceeds the strongest baseline by 1.00 point on SEED under multi-task full fine-tuning and leads the best listed baseline by 5.65 points on SEED-V under single-task full fine-tuning; on SEED-VII it trails the best baseline by 1.28 points under multi-task full fine-tuning.Source: §E.2.1 · pp. 24–25, Table 10 · p. 25, Table 11 · p. 25

E.2.2 MOTOR IMAGERY

Motor-imagery outcomes depend on the dataset and adaptation setting. NSP is below the strongest baseline on PhysioMI, narrowly trails it on Mimul-11, and exceeds it on BCIC-2a under the strongest reported settings.

E.2.2 MOTOR IMAGERY

  • 59.23 ± 00.67 BAC on PhysioMI with single-task full fine-tuning, 0.80 points below the strongest baseline in that setting.Single-task full fine-tuning is NSP’s strongest PhysioMI setting.Source: §E.2.2 · p. 26, Table 13 · p. 26
  • 51.66 ± 00.60 BAC on Mimul-11 with multi-task full fine-tuning, trailing the leading baseline by 0.89 points.Multi-task full fine-tuning is NSP’s strongest Mimul-11 setting.Source: §E.2.2 · pp. 26–27, Table 14 · p. 27
  • 46.58 ± 01.48 BAC on BCIC-2a with multi-task full fine-tuning, exceeding the strongest baseline by 4.08 points.Multi-task full fine-tuning is NSP’s strongest BCIC-2a setting.Source: §E.2.2 · p. 27, Table 15 · p. 27

E.2.3 CLINICAL EEG ANALYSIS

Across the downstream tasks, the strongest NSP adaptation setting varies by dataset. Component studies show cumulative gains, task-dependent effects, and sensitivity to target depth, EMA schedule, and label budget.

E.2.3 CLINICAL EEG ANALYSIS

  • TUEV and TUSL perform best with multi-task full fine-tuning at 74.34 and 81.83 BAC, while TUAB performs best with single-task full fine-tuning at 83.83 BAC.The reported leads over the strongest baseline are 0.07 points on TUEV, 1.27 on TUSL, and 1.11 on TUAB.Source: §E.2.3 · pp. 28–29

E.2.4 SLEEP STAGING

  • HMC reaches 75.66 BAC with multi-task full fine-tuning, exceeding the strongest baseline under that protocol by 3.48 points.HMC uses a sparse four-channel montage and 30-second epochs.Source: §E.2.4 · p. 29, Table 19 · p. 29

E.2.5 SEIZURE DETECTION

  • Siena reaches 86.90 BAC with multi-task full fine-tuning, leading the best baseline by 3.62 points; AUROC and AUCPR are also reported for this rare-event task.The dataset contains rare seizure events, motivating reporting retrieval and ranking metrics in addition to balanced accuracy.Source: §E.2.5 · p. 30, Table 20 · p. 30

E.2.6 MENTAL WORKLOAD

  • Workload reaches 82.53 BAC with single-task full fine-tuning, 8.41 points above the best listed baseline under that protocol.Multi-task full fine-tuning scores 74.28 BAC on Workload.Source: §E.2.6 · p. 30, Table 21 · p. 31

E.2.7 VISUAL TARGET DETECTION; E.2.8 NEURODEGENERATIVE DISEASE IDENTIFICATION

  • Things-EEG-2 reaches 72.77 BAC and leads by 1.68 points with single-task full fine-tuning; ADFTD reaches 56.66 BAC and leads by 0.95 with multi-task full fine-tuning.The strongest adaptation setting differs between these two tasks.Source: §E.2.7 · p. 31, Table 22 · p. 31, §E.2.8 · p. 32

F.1 CUMULATIVE COMPONENT ABLATION

  • Macro BAC rises from 44.44 for the shared-target baseline to 55.70 for Full NSP, while removing regularization lowers it to 46.88.Intermediate configurations score 45.75 with block masking, 51.88 with EMA, and 54.24 with identity residualization.Source: §F.1 · p. 32

F.2 GUARD DESIGN; F.3 RESIDUALIZATION DESIGN

  • Guard comparisons test random, spatial-only, temporal-only, and topology-separated contexts; partial channel or time residualization produces different task-wise changes and neither uniformly reproduces joint residualization.The exact-temporal condition is a separate temporal-only baseline without EMA or residualization, rather than an isolated guard-geometry change.Source: §F.2 · p. 32, Figure 9 · p. 33, §F.3 · p. 33

F.4 STATE REGULARIZATION COEFFICIENT; F.5 TARGET LAYER SELECTION; F.6 EMA SCHEDULE

  • Top-three-layer targets and cosine EMA outperform their stated alternatives, while the observed regularization optimum depends on label budget.Top-three targets score 54.25 macro BAC under Frozen 100%; cosine EMA scores 54.10 versus 51.12 under Frozen 100% and 55.70 versus 50.96 under Frozen 10%. The highest observed regularization macro occurs at λreg = 0.025 with 100% labels and λreg = 0.1 with 10% labels.Source: Figure 11 · p. 33, §F.5 · p. 34, §F.6 · p. 34

F.7 TASK-WISE COMPARISON OF EMA TARGET AND FULL NSP

The task-wise comparison indicates that Full NSP improves some tasks while leaving others similar to the EMA-target configuration. Across six fine-tuning heads, global mean–max has the highest four-task macro, but no head dominates every task.

F.7 TASK-WISE COMPARISON OF EMA TARGET AND FULL NSP

  • Full NSP improves some tasks while leaving others similar to the EMA-target configuration in the frozen single-task comparison.The comparison covers TUAB, TUEV, SEED, and PhysioMI with 10% labels.Source: §F.7 · p. 34, Figure 13 · p. 34

F.8 FINE-TUNING HEAD DESIGN

  • Global mean–max gives the highest four-task macro BAC at 70.82, narrowly ahead of channel mean–max at 70.81; task-wise winners differ, so no head is uniformly dominant.Average pooling leads on TUAB and SEED, global mean–max on TUEV, and channel mean–max on PhysioMI; small macro differences should be read with reported variability.Source: §F.8 · p. 35, Figure 14 · p. 35

G.1 MODEL SCALE

The scale comparison shows higher aggregate BAC with larger NSP encoders but heterogeneous task-level winners. Additional controls indicate transfer depends on learned parameter values and organization, and that recording exposure has some effect on selected downstream results.

G.1 MODEL SCALE

  • Macro BAC rises from 61.76 for Tiny through 63.28 for Medium to 64.51 for Large, while the increase is modest relative to the change in parameter count.The comparison uses independently pretrained models and single-task full fine-tuning.Source: §G.1 · p. 35, Figure 15 · p. 36
  • Best encoder size varies by dataset: Large leads on eight listed tasks, Medium on four, and Tiny on two, so the aggregate increase is not uniform task-wise.Large leads on HMC, TUAB, Workload, PhysioMI, Things-EEG-2, ADFTD, SEED-VII, and TUSL; Medium on TUEV, Siena, SEED-V, and Mimul-11; Tiny on SEED and BCIC-2a.Source: §G.1 · p. 35, Figure 15 · p. 36

G.2 INITIALIZATION CONTROLS

  • Most datasets retain a positive NSP gain over random initialization and weight permutation, though the gain varies and a few tasks do not improve over both controls.The comparison indicates that both learned parameter values and their arrangement can affect transfer.Source: §G.2 · p. 36, Figure 16 · p. 37

G.3 EFFECT OF PRETRAINING RECORDING EXCLUSION

  • Excluding evaluated downstream recordings changes BAC from 83.83 to 82.87 on TUAB, 73.17 to 71.83 on TUEV, and 81.17 to 79.88 on TUSL, with much transfer performance retained.The control defines exposure by recording-level membership in the pretraining manifest.Source: §G.3 · p. 37, Table 24 · p. 37

H MECHANISM DIAGNOSTICS

  • Context interventions and identity probes characterize prediction sensitivity and identity decodability, but do not demonstrate elimination of position-only prediction or local interpolation.These diagnostics complement downstream transfer results.Source: H MECHANISM DIAGNOSTICS · p. 37

H.1 CONTEXT INTERVENTIONS AND IDENTITY PROBES

Context interventions hold Targets fixed while changing Context input, and identity probes measure channel and relative-time information. The diagnostics show reduced, not eliminated, channel identity decodability in Target states.

H.1 CONTEXT INTERVENTIONS AND IDENTITY PROBES

  • Sample-mismatch relative cosine falls from 0.894 to −0.028, while Target channel-probe accuracy falls from 100.00% to 14.43% but remains above the 1/19 chance level.Targets are held fixed during Context interventions; these single-checkpoint diagnostics do not estimate variability across pretraining runs.Source: §H.1 · p. 38, Figure 17 · p. 38

H.2 COMPONENT-TRANSFER OVERVIEW

  • The component-transfer overview compares frozen single-task BAC across five variant groups, four datasets, and 10% versus 100% label budgets; component effects are strongly task-dependent.The figure covers guard, residualization, regularization, target-layer, and EMA variants.Source: §H.2 · p. 38, Figure 18 · p. 39

I.1 T-SNE ON ADDITIONAL TASKS

Additional analyses visualize representation geometry across ten downstream tasks, class-conditioned attribution patterns across 14 datasets, and task-adapted interchannel correlation scores.

I.1 T-SNE ON ADDITIONAL TASKS

  • Figures 19 and 20 compare raw EEG, frozen NSP, and fully adapted NSP representations across ten downstream tasks using independently fit, task-specific t-SNE projections.The projections illustrate task-specific changes in class geometry.Source: §I.1 · p. 39, Figure 19 · p. 40, Figure 20 · p. 40

I.2 GRAD-CAM TOPOLOGY MAPS

  • Class-conditioned Grad-CAM maps show spatially structured, task-conditioned patterns across 14 datasets, but do not identify physiological generators or functional connectivity.Each map sums relevance over time to a per-channel map.Source: §I.2 · p. 40, Figure 21 · p. 41

I.3 INTERCHANNEL CORRELATION STRUCTURE

  • Task-adapted NSP interchannel correlation diagrams encode selected attention-derived scores by ribbon width, with channels colored by anatomical region.Channel labels follow each dataset’s source montage and the benchmark adapter’s channel mapping.Source: §I.3 · p. 41, Figure 22 · p. 41
Loading 2609.31167v1…