PaperlayerView exact revision on arXiv

Source-linked AI summary

Flattening the Connectome Spectrum: A Spectral Filter for FC Induces a Pretraining Target for fMRI Encoders

Giovanni Marraffini, Victoria Shevchenko, Carlo Alberto Barbano, Demian Wassermann

arXiv:2609.37642v1cs.AIcs.LGq-bio.NC

TL;DR

Brain foundation models tested here underperform FC-based regression, despite the promise of transferable representations from unlabelled fMRI. The paper recalibrates FC eigenvalues and pretrains an encoder to match pairwise similarities of short-window embeddings to those of the recalibrated connectomes. The approach matches or exceeds KRR across tested datasets and targets, while the encoder performs comparably to leading tested BFMs with far fewer parameters.

  • Problem

    Although brain foundation models promise transferable representations from unlabelled fMRI, tested models underperform simple linear estimators fitted on FC.

    Source: §1 · p. 1

  • Method

    The paper flattens each subject’s FC eigenvalue spectrum and pretrains an encoder to match pairwise embedding similarities to those of the recalibrated connectomes.

    Source: §1 · p. 2, §3.2 · p. 4

  • Results

    Flattening the spectrum improves prediction across most phenotypes, datasets, and parcellations; the encoder matches or exceeds tested foundation models on almost every dataset and task with roughly an order of magnitude fewer parameters.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Flattened connectomes can improve prediction and serve as a usable pretraining target, while the encoders and transform remain ahead of raw FC with fewer subjects and shorter scans.

    Source: §6 · p. 9

    The exponent was selected on one HCP-YA Schaefer-400 cell, making that cell optimistic; the best exponent may depend on dataset, task, and cohort, and fine-tuning is outside the frozen-model comparison scope.

    Source: §7 · pp. 9–10

Abstract

from arXiv · show

Self-supervised pretraining reshaped prediction in language and vision, and brain foundation models (BFMs) inherited its promise. Representations learned from large unlabelled corpora should capture individual functional dynamics and generalise across cohorts. However, kernel ridge regression (KRR) fitted on functional connectivity (FC) matrices still predicts individual phenotypes more accurately than any BFM we tested. In this paper, we show that KRR is weighted by the eigenvalues of the FC which are miscalibrated for phenotype prediction. We apply an efficient spectral filter to recalibrate the eigenvalues of each subject's FC matrix, enabling the model to exploit more inter-individual variance. Across the 5 datasets, 11 parcellations and 6 prediction targets we tested, we match or exceed the KRR baseline. Based on this finding, we then pretrain a small encoder model on about 4,000 hours of fMRI from 162 open datasets, whereby we align the pairwise similarities between the embeddings of recording snippets with those between the recalibrated connectomes. Our model performs on par with the best of the 6 published BFMs we tested while having an order of magnitude fewer parameters. Our encoder performs better than FC on short scans and in smaller cohorts, especially in fingerprinting. We release the pretrained model weights, the code and the pretraining data, preprocessed and parcellated.

1 INTRODUCTION

Functional connectivity (FC) predicts individual phenotypes, but brain foundation models (BFMs) tested here underperform simple FC-based regression. The paper flattens FC eigenvalue spectra to improve prediction and uses the resulting connectomes as targets for encoder pretraining.

Motivation

  • FC captures subject-specific variance and predicts identity and cognitive measures, while tested BFMs underperform FC-based linear estimators.The paper attributes the promise of BFMs to representations learned from large unlabelled fMRI corpora, but notes that their prediction performance has not surpassed simple FC estimators.Source: §1 · p. 1

Core findings

  • Flattening FC eigenvalues improves prediction across most phenotypes, datasets, and parcellations, with especially large gains for fingerprinting over sex or age prediction.The same spectrum-flattened connectomes serve as pretraining targets for an encoder that performs comparably to the best tested BFMs with roughly an order of magnitude fewer parameters.Source: §1 · p. 2

Spectral transformation

  • The authors identify concentrated FC eigenvalue weighting as a limitation and recalibrate each subject’s eigenvalues using the selected exponent α*=0.35.The recalibration is reported to significantly improve prediction; the selected value is used throughout the paper.Source: §1 · p. 2

Pretraining and release

  • The pretrained encoder matches or exceeds tested foundation models on almost every dataset and task, and beats FC on cognitive composite prediction and fingerprinting with short scans and fewer subjects.The released package includes code, encoders with 1M and 8M parameters, and preprocessed and parcellated data from 162 open datasets.Source: §1 · p. 2

2 RELATED WORK

Prior work has explored alternative connectome representations, behavioural prediction from FC, brain foundation models, and similarity-based distillation. This paper builds on these lines while changing the connectome matrix supplied to an established regression protocol.

Other transforms of the connectome

  • Tangent-space and whitened matrix-log representations outperform raw correlations in connectome-prediction benchmarks, but the optimal power-Euclidean transform for behavioural prediction had not been reported.The power-Euclidean family includes raw FC at α=1 and approaches the log map as α→0; the paper reports that its extremes are worse than mid values.Source: §2 · p. 3

Predicting phenotypes from FC

  • Prior FC phenotype-prediction studies use KRR with correlation kernels and nested cross-validation, while deep networks have not beaten this regression baseline.The authors retain the regression, kernel, and cross-validation protocol and change only the matrix entering it.Source: §2 · p. 3

Brain foundation models

  • Recent rs-fMRI foundation models use Transformer or state-space backbones with differing objectives, but standardised comparisons against FC baselines remain rare.The reviewed objectives include masked reconstruction, masked-patch representation prediction, latent alignment, structural-functional encoding, and semantic token learning.Source: §2 · p. 3

Distillation with a similarity target

  • Similarity-based distillation and centred kernel alignment are established approaches, but their teachers have typically been trained networks or labels rather than a closed-form full-recording statistic.This paper uses the latter as its teacher target.Source: §2 · p. 3

3 METHODS

The methods preserve the established FC prediction protocol while changing the connectome’s eigenvalue weighting and using the transformed connectome as a teacher for encoder pretraining. The exponent is selected by phenotype alignment, and the student is trained to align its pairwise embedding similarities with the teacher’s.

3.1 PREDICTING FROM THE CONNECTOME

  • KRR predicts phenotypes from off-diagonal connectome entries using a Pearson correlation kernel; the BFM comparison instead uses a linear kernel on z-scored features.The correlation kernel normalizes each subject’s feature vector, whereas a linear kernel can retain information in its norm; the two approaches agree within 0.02.Source: §3.1 · p. 3

3.2 THE TRANSFORM

  • The transform raises each subject’s FC eigenvalues to α, changing their weights while preserving the eigenvectors and compressing the spectrum’s dynamic range.It is computed independently per subject, without a group mean or parameters estimated from other subjects; at α=0 it maps each subject to the identity.Source: §3.2 · pp. 3–4, Figure 2 · p. 4

3.3 CHOOSING THE EXPONENT

  • Nested cross-validated prediction of the cognitive composite selects α*=0.35, which is used throughout the paper; kernel-target alignment gives αA=0.387, with α* at 99.55% of its maximum.Kernel-target alignment measures how well the kernel geometry matches the target.Source: §3.3 · pp. 4–5

3.4 PRETRAINING BY DISTILLATION

  • Pretraining minimizes one minus the centered alignment between student embedding similarities and teacher similarities from spectrum-flattened FC of the full recording.The encoder sees only a short recording window while the teacher is computed from the whole recording; the teacher vectors define the correlation kernel used in regression.Source: §3.4 · p. 5

4 EXPERIMENTAL SETUP

The experiments use five named datasets for evaluation and 162 open datasets for pretraining. They assess phenotype and identification targets with nested cross-validation and compare frozen probes and encoder configurations.

4.1 DATASETS

  • Evaluation covers HCP-YA, AOMIC-ID1000, ABIDE-I, ADHD-200, and CoRR, while pretraining uses 162 openly available OpenNeuro datasets.Targets include cognitive composites, sex, age, autism diagnosis, ADHD diagnosis, and identification.Source: §4.1 · p. 5

4.2 EVALUATION

  • Scores are mean Pearson r or AUC with fold standard deviations, and paired representation comparisons use corrected resampled t-tests on matched folds.Nested cross-validation uses 10 outer folds, repeated 20 times unless stated otherwise, with family-aware HCP-YA, subject-level AOMIC, and site-grouped ABIDE-I and ADHD-200 splits.Source: §4.2 · p. 6

4.3 PRETRAINING

  • Two frozen fMRI-BERT encoders are pretrained at Sch-450 for 300,000 steps, using 80-timepoint windows and batches of 1024 recordings.The models have 1M and 8M parameters; the CKA loss improves with batch size, while a KL form of the target scores lower at every tested model size and batch size.Source: §4.3 · p. 6

5 RESULTS

The spectral transform improves FC-based prediction across cohorts and data budgets, and the encoder trained against the transformed connectomes is competitive with published foundation models. The encoder’s advantage over raw FC is clearest on short scans and smaller samples, while full-scan scores remain below the transformed-connectome teacher.

5.1 THE TRANSFORM IMPROVES PREDICTION ON EVERY COHORT TESTED

  • FCα* significantly improves over raw FC on both cohorts at every tested parcellation and also beats standard FC transformations.The α→0 limit scores below the peak, and none of the other tested spectral filters beats λα*.Source: §5.1 · p. 6, Table 1 · p. 7
  • With identical eigenvectors, the top 20 modes score 0.610 with λα* versus 0.544 with raw eigenvalues, showing the gain comes from eigenvalue weighting.Raw FC scores 0.542 with all 400 modes, compared with 0.624 using λα*.Source: §5.1 · p. 6, Table 2 · p. 7

5.2 THE ADVANTAGE HOLDS AT EVERY DATA BUDGET TESTED

  • FCα* stays ahead of raw FC at every tested training-subject count and scan duration.Figure 4 plots full-scan performance by training-set size and performance by duration of the first contiguous scan minutes, using 939 complete-run subjects.Source: §5.2 · p. 7, Figure 4 · p. 7

5.3 AGAINST BRAIN FOUNDATION MODELS

  • On full recordings, the encoder scores above every published model on the cognitive composite and fingerprinting, with significant differences on both; FCα* leads among full-scan representations.The encoder matches or outperforms the best short-window models on the cognitive composite using 2.4 minutes, significantly outperforming raw FC from the same duration.Source: §5.3 · p. 7, Table 3 · p. 8

5.4 DISTILLING THE TRANSFORM INTO A TIMESERIES MODEL

  • Encoder scores rise with scan duration and scale; on full scans they match raw FC but remain below FCα*, while on short scans and with fewer subjects they exceed FC-based KRR.The pretraining objective transfers part of the transform’s structure, passing every published BFM on the cognitive composite and beating raw FC with short scans and fewer subjects; larger batches are hypothesized to narrow the remaining gap.Source: §5.4 · p. 9

6 CONCLUSIONS

Flattening FC eigenvalues improves prediction and fingerprinting, and the flattened connectome can serve as a pretraining target. The transform and encoders stay ahead of raw FC with fewer subjects and shorter scans, and the openly sourced, multi-site pretraining data are released with the models.

6 CONCLUSIONS

  • Flattening each subject’s FC eigenvalues improves prediction and fingerprinting, and the flattened connectome is a usable pretraining target.The transform requires one eigendecomposition per subject and can replace the FC matrix in an existing pipeline.Source: §6 · p. 9
  • FCα* and the encoders stay ahead of raw FC with fewer subjects and shorter scans, a setting especially relevant to clinical cohorts where long scans and large samples are hard to obtain.Neither the transform nor the released encoders needs a GPU.Source: §6 · p. 9
  • The released encoders were pretrained on 162 openly available datasets from many sites and countries, with the data released alongside the models.The authors present this as a starting point for laboratories studying populations under-represented in large biobanks.Source: §6 · p. 9

7 LIMITATIONS

The selected exponent is based on one HCP-YA evaluation cell, and both the preprocessing and participant populations constrain how broadly the findings can be applied. The paper also notes limits in checkpoint selection and model comparisons.

7 LIMITATIONS

  • The exponent was selected on one HCP-YA Schaefer-400 cell, making that result optimistic; the best exponent may vary by dataset, task, or cohort.The same selected value matched or improved prediction on the other tested datasets, parcellations, and targets.Source: §7 · p. 9
  • The 162-dataset pretraining corpus combines sites, countries, protocols, and task types without standardisation, so one preprocessing pipeline probably preprocesses each dataset suboptimally.The cohorts also over-represent Western, healthy, well-educated participants, requiring re-validation before clinical use.Source: §7 · p. 9, ETHICS STATEMENT · p. 10
  • Encoder checkpoints were selected by downstream-probe early stopping, and continued pretraining may eventually hurt downstream performance.Published foundation-model rows also differ in parcellation and window length, not just model.Source: §7 · p. 9
  • Fine-tuning is outside the evaluation scope and could change the reported model ordering.The encoders are compared frozen, for use on small cohorts and short scans without a GPU.Source: §7 · p. 10
  • The study is a secondary analysis of public human neuroimaging cohorts, uses restricted HCP family structure only for fold assignment, and does not redistribute restricted cohorts or images.The authors report no new data collection or re-identification attempt; the proposed transform is a strong fingerprint.Source: ETHICS STATEMENT · p. 10
  • The authors provide code, per-fold results, model weights, and the processed pretraining data, enabling recomputation without imaging data or restricted cohorts.The listed corpus contains 162 datasets, 9,578 subjects, and 27,835 recordings.Source: REPRODUCIBILITY STATEMENT · p. 10

A DATA AND PREPROCESSING

The evaluation combines several public cohorts with different scan designs, preprocessing, parcellations, and targets. The pretraining corpus comprises 162 OpenNeuro datasets processed into short windows and released with the trained encoders.

A DATA AND PREPROCESSING

  • HCP-YA includes 955 adults with resting-state scans; 939 with four complete runs are used when uniform scan length is required, while AOMIC contributes 877 people with cognitive targets.HCP-YA uses four concatenated runs totaling 57.6 minutes; AOMIC participants have fMRIPrep-preprocessed data.Source: A DATA AND PREPROCESSING · p. 14
  • HCP-YA is parcellated at Schaefer resolutions 100–400, Sch-450, and NextBrain-528; AOMIC uses the same parcellations, while regional signals are band-passed and z-scored.HCP-YA uses ICA-FIX-denoised, minimally preprocessed resting-state runs; AOMIC uses fMRIPrep preprocessing.Source: A DATA AND PREPROCESSING · p. 14
  • ABIDE-I contributes 1,035 subjects across 17 sites, but only 158 have every model’s input; ADHD-200 has 298 scans, and CoRR supplies 782 scans from 391 people for test–retest identification.ABIDE-I uses a C-PAC filtered-global derivative; ADHD-200 includes two sites.Source: A DATA AND PREPROCESSING · p. 14
  • The pretraining corpus contains 9,578 subjects and 27,835 recordings from 162 OpenNeuro datasets, processed with fMRIPrep and parcellated at Sch-450.It provides 106,397 windows of 80 timepoints, about 4,000 hours in total; datasets retain native repetition times, so a window spans 1–4 minutes. HCP-YA, ABIDE-I, ADHD-200, and CoRR are excluded.Source: A DATA AND PREPROCESSING · p. 14
  • Targets include cognitive composites, sex and age, autism and ADHD diagnoses, and test–retest identification, with cognitive composites defined separately for HCP-YA and AOMIC.The HCP-YA composite is the first principal component of 58 behavioural measures; AOMIC’s uses four reasoning and memory subscales.Source: A DATA AND PREPROCESSING · p. 14

B EVALUATION PROTOCOLS

The evaluation uses nested, family-aware cross-validation and describes the spectral transform through geometric, heat-kernel, and filtering interpretations. It also presents a reusable eigendecomposition-based procedure for optimizing kernel-target alignment.

B EVALUATION PROTOCOLS

  • Nested evaluation uses 10 family-grouped outer folds, inner five-fold penalty selection, and 20 repetitions; choices are made within training folds, while representation comparisons share folds.The linear-kernel probe preserves embedding norms, unlike the correlation kernel, so scores are compared only within a protocol.Source: B EVALUATION PROTOCOLS · p. 15

C THE SCALED EIGENVECTORS BEFORE AND AFTER THE TRANSFORM

  • The transform scales each eigenvector by its transformed eigenvalue: unlike raw FC, where a few modes extend far beyond the reference sphere, no transformed mode exceeds four sphere radii.The comparison visualizes scaled eigenvectors at a common scale.Source: C THE SCALED EIGENVECTORS BEFORE AND… · p. 15

D.1 THE TRANSFORM AS A GEODESIC

  • Under both stated metrics, Σα is the geodesic from the identity to the raw connectome, with α giving the fraction of path length; eigenvalues move linearly in log space.The α→0 tangent is log Σ, while the transform at an interior α represents a point along the path rather than its tangent direction.Source: §D.1 · p. 16

D.2 THE TRANSFORM AS A HEAT KERNEL

  • In the heat-kernel interpretation, α is diffusion time: the leading mode does not decay, while lower-variance modes decay faster; the correlation-kernel regression has the same kernel under the stated rescaling.This equivalence is exact for correlation-kernel KRR, but not for the z-scored linear probe.Source: §D.2 · p. 16

D.3 THE TRANSFORM AS A SPECTRAL FILTER

  • The power transform is a monotone spectral filter that compresses the eigenvalue dynamic range without reordering modes; relative to raw FC, it amplifies low-variance modes relative to high-variance ones.The filter acts on connectome eigenmodes, not the temporal frequency of the BOLD signal; at α = 0 it becomes whitening.Source: §D.3 · pp. 16–17

D.4 COMPUTING THE ALIGNMENT OPTIMUM

  • Alignment optimization reuses one eigendecomposition per subject, reducing the search to a smooth scalar function of α that is maximized with Brent’s method.The training-fold optimum αA = 0.387 converged to tolerance 10−4 in eight evaluations; the full calculation took about half an hour on 16 CPU cores.Source: §D.4 · p. 18

E THE TRANSFORM SCREEN

The transform screen compares the power family with standard connectome transformations under correlation-kernel KRR. Its reported pattern is that gains over raw FC are associated with reweighting each subject’s own spectrum.

E THE TRANSFORM SCREEN

  • The screen evaluates FCα∗ against raw FC and three standard matrix transformations on HCP-YA Schaefer-400 using cognitive-composite correlation-kernel KRR and family-aware folds.The table caption specifies 10 × 10 cross-validation and mean ± standard deviation over 100 folds.Source: Table 5 · p. 18, E THE TRANSFORM SCREEN · p. 18
  • Group-fit transforms were initially screened on the full cohort; refitting within folds or on an external corpus changed their scores by at most 0.004.Raw FC scores 0.545 in this 10-repetition screen, compared with 0.543 under the 20-repetition protocol.Source: E THE TRANSFORM SCREEN · p. 18
  • Every screened transform scoring above 0.58 reweights each subject’s spectrum, whereas subspace reduction, edge normalisation, and estimators without spectral reweighting score at or below raw FC.Within the power family, scores rise from α = 1 to a maximum between 0.25 and 0.5, then fall toward the log-Euclidean limit.Source: E THE TRANSFORM SCREEN · p. 18

F WHERE THE GAIN COMES FROM: CONTROLS AT MATCHED EIGENVECTORS

The spectral transformation improves cognitive prediction by reweighting each subject’s eigenvalues while retaining the eigenvectors. Controls show that the gain is not explained by subject identity alone, and restricting HCP-YA to complete scans barely changes results.

F WHERE THE GAIN COMES FROM: CONTROLS AT MATCHED EIGENVECTORS

  • FCα∗ scores 0.624 ± 0.077 versus raw FC’s 0.542 ± 0.079; the eigenvectors are unchanged, so the gain comes from eigenvalue reweighting.With 20 modes, reweighting adds +0.066; under raw weights, adding the remaining 380 modes changes the score by −0.002.Source: F WHERE THE GAIN COMES FROM: CONTROLS… · pp. 18–19
  • The prediction signal is distributed across roughly the top 20 modes, while no single mode is strong alone.The top 1, 5, 20, 100 and all 400 modes score 0.321, 0.532, 0.610, 0.617 and 0.624, respectively.Source: F WHERE THE GAIN COMES FROM: CONTROLS… · p. 19
  • A hand-set eigenvalue profile scores 0.586 ± 0.081, clipping scores 0.615 ± 0.087, and neither random nor directly optimized filters exceed the α∗ result of 0.624.The directly optimized 30-parameter filter and 64 × 64 MLP score 0.586 ± 0.090 and 0.579 ± 0.083, respectively.Source: F WHERE THE GAIN COMES FROM: CONTROLS… · p. 19
  • The kernel form has little effect, but the per-subject basis matters: its top 20 eigenvectors score 0.610, while a shared basis needs about 200 components to match that score.A shared basis with 900 components scores 0.589.Source: F WHERE THE GAIN COMES FROM: CONTROLS… · p. 19
  • Identity alone does not predict cognition under family-aware folds: a one-hot identifier scores 0.000 and a random fixed embedding 0.054 ± 0.023.FCα∗ raises split-half prediction reliability from 0.597 ± 0.010 for raw FC to 0.830 ± 0.001; its score of 0.624 is 95% of its disattenuated ceiling of 0.655.Source: F WHERE THE GAIN COMES FROM: CONTROLS… · p. 19

H THE COMPLETE-SCAN SUBSET

I FOUNDATION-MODEL COMPARISON DETAILS

Published models were evaluated using their released checkpoints and feature-extraction code, with model-specific inputs and window lengths. The section also reports paired-test procedures and CPU inference costs.

I FOUNDATION-MODEL COMPARISON DETAILS

  • Published models use their released checkpoints and feature-extraction code, but differ in parcellation, input type and window length; ABIDE-I evaluation uses the 158-subject all-model intersection.Window embeddings are averaged per subject before probing.Source: I FOUNDATION-MODEL COMPARISON DETAILS · p. 20
  • On four CPU threads without a GPU, the 1M model processes a 57.6-minute HCP-YA recording in 0.46 s using under 1 GB; the 8M model takes 2.6 s and 1.4 GB.A 200-timepoint window takes 30 ms.Source: I FOUNDATION-MODEL COMPARISON DETAILS · p. 20
  • Paired tests behind Table 3 use per-fold differences with corrected resampled t-tests for cognitive and demographic targets, and an exact McNemar test for CoRR fingerprint accuracy.The tests use 200 folds for the cognitive composite, 100 for sex and age, and shared scans for CoRR.Source: I FOUNDATION-MODEL COMPARISON DETAILS · p. 20, Table 9 · p. 21

J PRETRAINING ABLATIONS

The ablations compare pretraining objectives, model scale, scan exposure and cohort size, while the distillation details clarify how the teacher similarities and optional augmentations are constructed. Additional analyses qualify cross-parcellation comparisons and the ADHD-200 evaluation.

K DISTILLATION DETAILS

  • The KL objective matches each recording’s student and teacher softmax distributions over other batch recordings, using standardized distances and teacher similarities derived from connectome correlations.The teacher vectors are centered and unit-norm, making squared distance equal to 2 − 2r_ij; the KL match is per-anchor, unlike global Gram-matrix alignment.Source: J PRETRAINING ABLATIONS · p. 21, Figure 9 · p. 24

J PRETRAINING ABLATIONS

  • The best CKA model beats the best KL model by +0.050 on shared folds (pNB = 0.012), ahead in 161 of 200 folds.As batch size grows, KL degrades while CKA improves; Table 13 evaluates the 1M encoder on the cognitive composite with windowed inputs.Source: J PRETRAINING ABLATIONS · p. 23, Table 13 · p. 24
  • CKA gains with a tenfold parameter increase and reaches its best score in fewer steps (12k versus 56k), while KL gains little and the 8M KL model falls below the 1M model.The same objective ordering holds when the whole recording is read as one sequence.Source: J PRETRAINING ABLATIONS · pp. 22–24, Table 14 · p. 24
  • The 8M model is above raw FC up to 180 subjects, while both encoders are above raw FC up to 4.8 minutes of scan and FC leads from 19.2 minutes.At 200 and 250 subjects the 8M model is within 0.01 of FC; FC leads from 300 subjects. For the 1M model, windowed scores remain 0.49 to 0.51 across window lengths, indicating that scan amount determines its score.Source: J PRETRAINING ABLATIONS · p. 24, Figure 10 · p. 25

K DISTILLATION DETAILS

  • For centered, unit-norm teacher vectors, distance-based and correlation-kernel teacher distributions agree to numerical precision; this equivalence does not hold for unnormalized vectors.The reported teacher distributions agree to 4 × 10^-14 in log-probability, with KL = 2 × 10^-17.Source: K DISTILLATION DETAILS · pp. 24–25
  • A positive pair is unnecessary because the teacher weights every other recording; FC-preserving Fourier phase surrogates are admissible but were not used in reported models.With the KL objective, augmentation scores 0.436 ± 0.093 versus 0.433 ± 0.090 without it on the windowed protocol; the matched-step difference is not significant (pNB = 0.15).Source: K DISTILLATION DETAILS · pp. 26–27

K.1 AOMIC PARCELLATION SWEEP

  • The AOMIC parcellation sweep uses six atlases, but its rows are not comparable with Table 1 because the re-parcellation pipeline reproduces Schaefer-400 timeseries and FC edges only to median correlations of 0.60 and 0.77.The Schaefer-400 row serves as a bridge; the sign and size of the gain carry across every tested atlas, including a different atlas family.Source: §K.1 · p. 27, Table 18 · p. 27

K.2 ADHD-200

  • ADHD-200 retains 298 labelled scans across two sites, but several published encoders cover only 119 scans from one site and cannot be scored; all scorable representations are near chance.The results are reported as a null result rather than as a Table 3 column.Source: §K.2 · p. 27
Loading 2609.37642v1…