betaFeedback

WorksComputer Science

Band2Net: Frequency-to-Network Transport Distillation for fMRI-Privileged, EEG-Only Session-Condition Decoding

Abstract

Functional MRI resolves large-scale brain networks but is expensive and stationary, whereas EEG is portable but poorly localised. We treat fMRI as privileged information, available during training and absent at deployment, with EEG and fMRI acquired from the same participants under matched conditions but not simultaneously. Band2Net encodes EEG into five frequency-specific connectivity graphs and fMRI into seven canonical network tokens from a frozen teacher, and solves a balanced entropic transport problem between the two token sets that gives each EEG band its own fMRI target. In subject-grouped cross-validation on the PsiConnect psilocybin data set, Band2Net reaches 0.801 ROC-AUC for EEG-only session-condition decoding, 5.0 AUC points (95% CI 1.8-8.2) above an EEG-only student. Although its margins over a single shared plan, a KL-relaxed plan and label-free network descriptors remain unresolved, secondary comparisons favour band-specific transported targets over logit, global-feature, contrastive and pooled-target distillation, though not after Holm correction.

Introduction

fMRI reveals large-scale functional networks and underlies the canonical parcellations of network neuroscience [1]. EEG resolves millisecond rhythms and is inexpensive, portable and repeatable, yet has no explicit notion of a cortical network and localises poorly in sensor space. The two connectomes are related but far from redundant, and their coupling is present in every frequency band with band-dependent topology [3, 4] [2].

This asymmetry matters when a brain state must be tracked outside a scanner. Psilocybin enhances between-network communication and reduces default-mode integrity [5] while inducing broadband desynchronisation most pronounced in the alpha range [6]; following such network-level effects across a dosing day or repeated trial visits is feasible with EEG and impractical with fMRI. We therefore use pre- vs. post-dose session decoding, whose label the experimenter knows, as a proxy benchmark: if fMRI network structure transfers into an EEG-only model, that model should separate the two states better than one trained on EEG alone.

Current approaches leave three gaps. (i) Fusion models [7] need fMRI at test time, and EEG-to-fMRI synthesis [8] avoids this only by regressing regional BOLD signals from simultaneous recordings. (ii) Privileged learning lets fMRI shape training and then discards it [9, 10], but logit [11] and pooled-feature distillation [12], including from fused fMRI–EEG teachers in CMD-AD [13], reduce the teacher to targets with a fixed correspondence and discard the network structure that makes fMRI worth acquiring; FRIST [14] learns fMRI-informed spectral projections for EEG-only decoding, but from simultaneous EEG–fMRI. (iii) A band-to-network mapping cannot be specified a priori: coupling is band dependent and may differ between individuals [3] [2]; we test whether per-observation routing is needed. Optimal transport compares structured sets; it

has served distillation [15] and privileged multimodal distillation between examples in PKDOT [16, 17], but not routing from EEG bands to fMRI networks.

We therefore ask: given EEG and fMRI from the same participants under matched conditions but not simultaneously, can the network-level organisation captured by fMRI be transferred into a decoder that uses EEG alone, and must that transfer be frequency-band specific? We solve a balanced entropic transport problem between five EEG band representations and seven canonical fMRI network tokens for every paired observation, and use the plan to give each band its own privileged target. Unlike PKDOT (between examples) and CMD-AD (fixed scale-to-scale matching), routing is learned within each observation, over bands against networks.

Our main contributions are as follows:

  • We cast fMRI-privileged EEG learning from separately acquired, state-matched recordings as balanced frequency-to-network transport, giving each band its own fMRI target.

  • We distil these targets into an EEG-only student through an asymmetric stop-gradient that keeps the teacher space fixed.

  • We show under subject-grouped validation that fMRI privilege improves EEG-only decoding, and that label-free network descriptors recover most of the gain without a learned teacher.

Method

Band2Net connects a frozen fMRI teacher to the EEG student through a transport plan estimated individually for each paired observation of a subject (Fig. 1). Training involves both modalities; inference uses only the EEG branch.

Frequency-specific EEG graphs

Each preprocessed recording is band-pass filtered into the five canonical bands b∈{δ,θ,α,β,γ}b \in\{\delta, \theta, \alpha, \beta, \gamma\}. Connectivity in band bb for the NEN_E channels is the weighted phase-lag index [18], which suppresses the zero-lag coupling added by volume conduction, giving AbE∈RNE×NEA_b^E \in\mathbb{R}^{N_E \times N_E}. Each channel takes its row of the unthresholded matrix as node feature, XbE=AbEX_b^E = A_b^E; spectral power is kept aside as a separate baseline. Keeping the strongest 25% of off-diagonal entries per observation as weighted edges, A^bE\hat{A}_b^E, a density fixed before any test and shared with the fMRI graphs, produces one graph GbE=(A^bE,XbE)G_b^E = (\hat{A}_b^E, X_b^E) per band.

One encoder EEE_E serves all five bands, so features share one space. Band identity is restored via the learned embedding pbp_b, giving eb=EE(GbE)+pbe_b = E_E(G_b^E) + p_b and E=[eδ,…,eγ]∈R5×dE = [e_\delta, \ldots, e_\gamma] \in\mathbb{R}^{5 \times d}. An attention mechanism scores each band, sb=⟨q,Web⟩s_b = \langle q, W e_b \rangle with βb=softmax⁡b(sb)\beta_b = \operatorname{softmax}_b(s_b), performs the weighted sum zE=∑bβbebz_E = \sum_b \beta_b e_b, and a classifier outputs y^S=gS(zE)\hat{y}_S = g_S(z_E).

Canonical functional-network teacher

Regional BOLD signals are extracted with a Schaefer parcellation carrying Yeo network labels [19, 1]. Functional connectivity between each pair of the NFN_F ROIs is the Fisher-transformed Pearson correlation of their time series, each region takes its whole-brain profile as a node feature, and edges are thresholded on absolute values as in § 2.1. A compact ROI encoder EFE_F gives ROI embeddings HF∈RNF×dH^F \in\mathbb{R}^{N_F \times d}, where each of the KK canonical networks aggregates the embeddings as

fk=Pool⁡({HiF:i∈Nk}),F=[f1,…,fK]∈RK×d,(1)f_k = \operatorname{Pool}(\{H_i^F : i \in\mathcal{N}_k\}), \quad F = [f_1,\ldots,f_K] \in\mathbb{R}^{K \times d}, \tag*{(1)}

where Nk\mathcal{N}_k represents the ROIs belonging to network kk and Pool is the mean. The teacher classifier gFg_F classifies Pool⁡(f1,…,fK)\operatorname{Pool}(f_1,\ldots,f_K) under a cross-entropy loss; each fMRI run yields its own tokens. One teacher is learned per outer fold and seed (25 in all) on the outer-training participants outside the inner validation set (§ 3.2) and then frozen; it supplies the tokens for all outer-training participants within that fold, so that fkf_k lives in one latent space and is comparable between subjects.

Transport distillation

Only the EEG representation is constrained to be compatible with the teacher. The projector PEP_E, a 64–128–64 MLP, maps each band embedding into the teacher space, ub=PE(eb)u_b = P_E(e_b). With uˉb=ub/∥ub∥\bar{u}_b = u_b/\lVert u_b\rVert and vˉk=fk/∥fk∥\bar{v}_k = f_k/\lVert f_k\rVert for the frozen tokens, the cost for paired observation ss (one participant’s EEG and fMRI from the same session and context; eight per participant, 2 sessions × 4 contexts, never pooled across participants) is Cbk(s)=1−uˉb⊤vˉkC_{bk}^{(s)} = 1 - \bar{u}_b^\top\bar{v}_k. Because the teacher is frozen, the cost can be lowered only by moving both representations together without retaining correspondence.

We state this as a balanced entropic transport problem over U(a,b)={Π≥0:Π15=a, Π⊤15=b}\mathcal{U}(a,b) = \{\Pi\ge0 : \Pi\mathbf{1}_5 = a,\ \Pi^\top\mathbf{1}_5 = b\}, with uniform marginals a=1515a = \frac{1}{5}\mathbf{1}_5 and b=1K1Kb = \frac{1}{K}\mathbf{1}_K. The per-observation plan is

Π∗(s)=arg min⁡Π∈U(a,b)⟨Π,C(s)⟩+ε∑b,kΠbk(log⁡Πbk−1),(2)\Pi^{*(s)} = \underset{\Pi\in\mathcal{U}(a,b)}{\operatorname{arg\,min}}\langle\Pi,C^{(s)}\rangle+ \varepsilon\sum_{b,k}\Pi_{bk}(\log\Pi_{bk} - 1), \tag*{(2)}

solved by batched log-domain Sinkhorn iterations [20] without gradient tracking. With fixed marginals each band sends and each network receives equal mass, so the plan only describes where supervision goes; a band’s influence on the decision is set by the readout weights βb\beta_b.

Rows of Π∗\Pi^* sum to 15\frac{1}{5}, so the transported target for band bb is the row-normalised linear combination tb=5∑kΠbk∗fkt_b = 5\sum_k \Pi^*_{bk}f_k, giving one privileged fMRI target per band rather than one for the whole recording. The student is then trained with

L=LCE+λOT5∑b=15∥PE(eb)−sg⁡(tb)∥22.(3)\mathcal{L} = \mathcal{L}_{\mathrm{CE}} + \frac{\lambda_{\mathrm{OT}}}{5}\sum_{b=1}^{5}\lVert P_E(e_b)-\operatorname{sg}(t_b)\rVert_2^2. \tag*{(3)}

The cosine cost decides only where mass is routed; the loss compares unnormalised vectors, so the student also matches the scale of its target. The stop-gradient covers the whole target path: CC, Π∗\Pi^* and tbt_b are recomputed from the current embeddings but are constants in the backward pass, so gradients reach the student only through PE(eb)P_E(e_b) and LCE\mathcal{L}_{\mathrm{CE}}. Plans can shift as the student learns, but each tbt_b is a convex combination of fixed teacher tokens and the column constraint gives every token equal mass, so targets stay within the convex hull of the tokens and no network can be ignored. At inference the fMRI branch, the projector and the transport solver are discarded.

Experiments

Data and preprocessing

PsiConnect is a public multimodal study of context-dependent brain dynamics under psilocybin, released through OpenNeuro [21, 22, 23]. Each participant was recorded before and after a 19 mg oral dose of psilocybin, in four contexts per session: eyes-closed rest, guided meditation, music listening and movie viewing. EEG and fMRI were acquired separately under matched conditions, not simultaneously, so supervision is state-matched rather than time-locked. Of the 65 participants listed, 56 gave usable EEG and fMRI both pre- and post-administration after quality control, producing 56×2×4=44856 \times2 \times4 = 448 paired observations, balanced by construction across conditions. A session failed if any fMRI run had mean framewise displacement above 0.5 mm, or if RELAX rejected over 20% of channels or left fewer than 120 clean 2-s epochs in any EEG context; a participant was kept only if both sessions passed in both modalities. Brain–MGF [7] fuses both modalities on this cohort but evaluates only rest and meditation on a different sample, so its figures are not comparable.

fMRI was multi-echo EPI (TR = 0.91 s), processed with fMRIPrep [24] using motion correction, normalisation to MNI space, echo combination and band-pass filtering to 0.01–0.1 Hz. Signals were then averaged within the 200 parcels of the Schaefer atlas, each labelled with one of the seven Yeo networks [19], so NF=200N_F = 200 and K=7K = 7; one run covers each context. EEG was recorded from 64 scalp channels at 500 Hz, with separate EOG and ECG channels, and cleaned with RELAX [25]; the four contexts were identified from the continuous recording by stimulus markers. After removing the EOG and ECG channels, interpolating rejected scalp channels (NE=64N_E = 64) and re-referencing to the average, the bands δ\delta 0.5–4, θ\theta 4–8, α\alpha 8–13, β\beta 13–30 and γ\gamma 30–45 Hz were extracted by zero-phase fourth-order Butterworth filtering, with gamma cut off at 45 Hz to avoid the 50 Hz line. wPLI used the Hilbert cross-spectrum over the first 120 clean, non-overlapping 2-s epochs of each context, equalising data length and estimator bias across contexts.

Evaluation protocol

The subject is the unit of observation. Subject-grouped five-fold cross-validation over five fixed seeds ensures that no subject appears in both test and training sets through another session, context or modality, and batches are balanced across contexts. In each outer fold, 20% of the training participants form an inner validation set. The teacher and every graph student are fitted on the rest and early-stopped on inner-validation cross-entropy; the frozen teacher then supplies tokens for all outer-training participants, and outer-test fMRI is never used for training or model selection. For each method separately, ε∈{0.05,0.1,0.2}\varepsilon\in\{0.05, 0.1, 0.2\} (transport variants), the distillation weight λ∈{0.1,0.3,1}\lambda\in\{0.1, 0.3, 1\} (for Band2Net and every KD baseline; λOT\lambda_{\mathrm{OT}} in Eq. (3)), τ∈{0.1,1,10}\tau\in\{0.1, 1, 10\} (unbalanced plan) and the LR/SVM penalty Creg∈{0.01,0.1,1,10}C_{\mathrm{reg}} \in\{0.01, 0.1, 1, 10\} are selected in every outer fold and seed by inner-validation AUC on that fold’s training participants only; each pipeline is fixed before it predicts outer-test EEG. Across the 25 splits, Band2Net’s modal choices were ε=0.1\varepsilon= 0.1 (14) and λOT=0.3\lambda_{\mathrm{OT}} = 0.3 (13).

Each seed yields one out-of-fold prediction per session context; these are averaged over seeds and then over contexts, giving one score for each of the 112 sessions, so no subject is duplicated in metrics or resampling. Metrics are computed once on these averaged probabilities (Platt-scaled for SVMs), never averaged over folds or seeds; balanced accuracy and macro-F1 use a threshold of 0.5, fixed in advance as the classes are balanced. Uncertainty comes from bootstrapping participants with replacement (both sessions together, never context-level rows), giving 95% percentile intervals (10,000 resamples). Paired comparisons apply the same resample to both models, never treating folds as independent observations. Band2Net vs. the EEG-only student is the single primary comparison, fixed in advance; all other paired comparisons are secondary with unadjusted intervals; Holm correction over the 16 secondary comparisons in Table 1, applied post hoc as a robustness check, is marked there, and the four context-shift gains are left unadjusted. The context-shift test reuses these folds, so test participants are unseen, and removes the held-out context from all training, teacher included, and from hyperparameter selection. As condition is partly confounded with acquisition order and day, we speak of session-condition decoding, not a pharmacological biomarker.

MethodROC-AUCGainBAF1
EEG-only training
Band power + LR0.694 [0.640, 0.746]10.7 [6.3, 15.2]∗64.363.9
EEG wPLI FC + SVM0.728 [0.676, 0.778]7.3 [3.2, 11.5]∗66.165.7
Riemannian TS + LR0.739 [0.688, 0.788]6.2 [2.0, 10.5]67.066.6
EEG-only GNN0.751 [0.701, 0.799]5.0 [1.8, 8.2]†68.868.3
fMRI-privileged training, EEG-only inference
Logit KD0.759 [0.709, 0.806]4.2 [1.3, 7.1]69.669.2
Global feature KD0.768 [0.719, 0.814]3.3 [0.6, 6.0]70.570.1
CRD0.771 [0.722, 0.817]3.0 [0.2, 5.8]70.570.1
Pooled-target0.772 [0.723, 0.818]2.9 [0.1, 5.7]70.570.2
Band2Net ablations
Fixed random plan0.774 [0.725, 0.820]2.7 [0.4, 5.1]70.570.2
Row-softmax plan0.776 [0.727, 0.822]2.5 [0.3, 4.7]71.471.0
Shared plan0.787 [0.739, 0.832]1.4 [−0.7, 3.5]72.371.9
Unbalanced plan0.791 [0.744, 0.836]1.0 [−0.9, 2.9]71.471.1
Descriptor targets0.793 [0.746, 0.837]0.8 [−1.4, 3.1]72.372.1
Cross-fitted tokens0.794 [0.747, 0.839]0.7 [−1.1, 2.5]72.371.9
Band2Net (proposed)0.801 [0.755, 0.844]–73.272.8
Correspondence controls (privileged, EEG-only inference)
Cond.-matched shuffle0.775 [0.726, 0.821]2.6 [0.4, 4.9]70.570.0
Unrestricted shuffle0.748 [0.697, 0.796]5.3 [2.2, 8.4]∗69.767.5
Random frozen teacher0.757 [0.707, 0.805]4.4 [1.5, 7.4]∗69.669.3
fMRI teacher (ref.)0.836 [0.793, 0.875]–75.975.6

Table 1. EEG-only decoding of session condition on the 112 session-level predictions (§ 3.2). AUC with 95% subject-level bootstrap CI. Gain: paired AUC difference, Band2Net minus method, in AUC points (AUC×100) with 95% CI; †primary comparison; ∗significant after Holm correction over the 16 secondary comparisons in this table. BA: balanced accuracy, F1: macro-F1 (%). Per-seed AUC SD is at most 0.012 for every method. The fMRI teacher is not an EEG-only model.

Baselines and implementation

Per-band relative power with logistic regression tests how far simple spectral features solve the problem, vectorised wPLI connectivity with a linear SVM whether graph learning helps, and tangent-space features of shrinkage-regularised band-wise covariances with logistic regression [26] give the standard Riemannian EEG baseline. The EEG-only GNN uses the proposed encoder and classifier with the cross-entropy loss alone. Logit [11], global feature [12] and contrastive representation distillation (CRD) [27] to the pooled teacher embedding represent classical privileged learning, while pooled-target matching replaces Π∗\Pi^* with the uniform plan Πlk=1/(5K)\Pi_{lk} = 1/(5K), sending the same pooled teacher vector to every band.

Six ablations change only the plan or the tokens. The fixed random plan solves the balanced problem once for a random cost and keeps it throughout training (four draws), giving band-specific targets without learned routing. The row-softmax plan, Πlk=15softmax⁡(−Cbk/ε)\Pi_{lk} = \frac{1}{5}\operatorname{softmax}(-C_{bk}/\varepsilon), drops the column constraint. The shared plan applies one Sinkhorn plan, computed from the running mean of the cost over training observations, to every observation, so routing cannot adapt to participant, session or context. The unbalanced plan relaxes the column constraint with a KL penalty of weight τ\tau [28]. Descriptor targets replace the learned tokens with label-free descriptors: network kk is described by its mean connectivity to every network, dk∈RKd_k \in\mathbb{R}^{K}, mapped to d=64d=64 by a fixed random orthonormal map. Cross-fitted tokens come from teachers fitted without the participant (four inner folds).

Three variants probe the correspondence. The condition-matched shuffle permutes teacher tokens among training participants within each condition ×\times context cell (one derangement per fold and seed), removing individual correspondence but keeping condition and context; the unrestricted shuffle permutes across all training observations, so it also mismatches condition for about half of them. The randomly initialised frozen teacher is never fitted and carries no learned fMRI content.

All privileged variants share the EEG encoder, classifier and schedule. The encoders are two graph convolutional layers (symmetric normalisation, self-loops) with residual connections, layer normalisation and 0.3 dropout, hidden width 128 and output dimension d=64d=64, with mean node pooling. Training uses AdamW with learning rate 3×10−43 \times10^{-4}, weight decay 10−410^{-4}, batch size 16, at most 200 epochs and early stopping after 20 epochs without inner-validation improvement; Sinkhorn runs for 50 iterations. Withheld bands are dropped before the attention readout, whose softmax renormalises over the rest. Experiments used one NVIDIA A100 (40 GB) GPU. Code will be released upon publication.

Figure 2(a) shows the mean transport plan. It is diffuse rather than near one-to-one: alpha, beta, gamma and delta concentrate on different networks and theta is the least committed band; across participants the cells have a median IQR of 3.3 percentage points of row mass, and per-observation row entropy falls from 1.92 to 1.49 nats in training. Entries are teacher-space weights, not physiological coupling; delta’s top weight is on the susceptibility-prone limbic network.

Figure 2. (a) Mean transport plan over training observations, folds and seeds, normalised per band (%, rounded); seed-to-seed total variation 0.11. Yeo-7 networks as in Figure 1: visual, somatomotor, dorsal and ventral attention, limbic, frontoparietal, default mode. (b) Change in ROC-AUC points (AUC × 100) when bands are withheld at test time, relative to each model’s full-band score.

Figure 1. Band2Net. Five band-wise EEG connectivity graphs and seven canonical fMRI network tokens are matched by a balanced transport plan Π∗\Pi^* computed per paired observation, whose transported targets tbt_b supervise the EEG student through a stop-gradient; the frozen fMRI branch is discarded at inference.

Fig. 2(b) withholds frequency bands at test time. Omitting alpha hurts every model most, consistent with alpha modulation under psilocybin [6]. Band2Net loses 2.9 AUC points against 4.8 for the EEG-only student, and 3.9 against 6.2 when two random bands are withheld, point estimates indicating an embedding less tied to one band. Omitting gamma costs every model under one AUC point, possibly reflecting residual muscle artefact at 30–45 Hz. Leaving one context out, Band2Net’s paired gain over the EEG-only student is 4.2 [0.6, 7.9] AUC points for rest, 3.6 [−0.1, 7.3] for meditation, 2.4 [−1.4, 6.2] for music and 0.8 [−3.2, 4.8] for movie; only the rest interval excludes zero.

The cohort is small, 56 participants from one data set. Session condition is partly confounded with acquisition order and day, although predicted probability does not drift with context position within pre-dose sessions (slope 0.004 per context, [−0.012, 0.020]); AUC is similar with and without prior meditation training (0.809 vs. 0.792). As the modalities were not recorded simultaneously, the plan aligns state-level representations, not neurovascular coupling.

Results and discussion

In Table 1, hand-crafted features are informative (band power 0.694, wPLI 0.728, Riemannian 0.739), while the EEG-only GNN (0.751) is the strongest EEG-only model. Band2Net reaches 0.801 AUC, 73.2% balanced accuracy and 72.8% macro-F1, with a primary gain of 5.0 AUC points [1.8, 8.2] over the EEG-only student. Pooled-target and global feature distillation perform similarly (0.772 vs. 0.768); nominal margins over these baselines suggest a possible benefit from band-specific routing. Only margins over band power, wPLI, the unrestricted shuffle and the random teacher survive Holm correction.

Learned routing among distinct teacher tokens (mean pairwise cosine 0.47) nominally adds beyond target diversity (2.7 AUC points [0.4, 5.1], not Holm-significant); the row-softmax margin is also nominal (2.5 AUC points [0.3, 4.7]), and KL relaxation gives 0.791. The shared-plan difference (1.4 AUC points [−0.7, 3.5]) remains unresolved, consistent with similar variation between and within participants (FF ratio 1.16, permutation p=0.23p=0.23). Label-free network descriptors reach 0.793 (gap 0.8 AUC points [−1.4, 3.1]), recovering most of the gain through network connectivity summaries without a learned teacher. Band-specific targets offer little apparent benefit without learned routing (fixed random plan 0.774 vs. pooled-target 0.772), whereas a single learned plan shared across observations performs close to per-observation routing (0.787 vs. 0.801).

The condition-matched shuffle falls 2.6 AUC points [0.4, 4.9] below Band2Net, a nominal advantage suggesting a benefit from pairing each participant’s EEG with their own fMRI. The unrestricted shuffle (0.748) falls below the EEG-only student because it also corrupts condition supervision, and a randomly initialised frozen teacher (0.757) shows that the alignment term and architecture alone explain little of the gain. Band2Net’s only extra trainable module, the projector (about 0.02M parameters), is discarded at inference. The teacher reaches 0.95 AUC on its training participants and 0.836 on unseen ones, yet cross-fitted tokens, free of each participant’s own label, give 0.794, so in-sample tokens do not explain the gain.

Conclusion

Band2Net transfers fMRI network structure from separately acquired recordings into EEG-only decoding through balanced band-to-network transport. It improves decoding by 5.0 AUC points over an EEG-only student; secondary comparisons nominally favour it over logit, global-feature, contrastive and pooled-target distillation. Participant-specific pairing shows a nominal benefit, while label-free network descriptors recover most of the gain. EEG-only deployment supports repeated monitoring outside the scanner. Whether gains extend to order-balanced cohorts and continuous outcomes such as subjective drug intensity remains open; a single learned plan may suffice at this sample size.

Compliance with ethical standards

This is a secondary analysis of de-identified, publicly released data (CC0 waiver); the original study was approved by an institutional human research ethics committee and all participants gave informed consent.

Acknowledgments

The authors thank the PsiConnect investigators for releasing the data under an open licence. This research was funded by Khalifa University of Science and Technology. The authors have no relevant financial or non-financial interests to disclose. ChatGPT (OpenAI) assisted with writing and Claude Opus 5.5 (Anthropic) with coding and revision; the authors reviewed all AI-assisted output and take full responsibility for the content of this paper.

References

  1. [1]B. T. Thomas Yeo, Fenna M. Krienen, Jorge Sepulcre, et al., “The organization of the human cerebral cortex estimated by intrinsic functional connectivity,” J. Neurophysiol., vol. 106, no. 3, pp. 1125–1165, 2011. [2] Fani Deligianni, Maria Centeno, David W. Carmichael, et al., “Relating resting-state fMRI and EEG whole-brain connectomes across frequency bands,” Front. Neurosci., vol. 8, pp. 258, 2014.
  2. [3]Jonathan Wirsich, Anne-Lise Giraud, and Sepideh Sadaghiani, “Concurrent EEG- and fMRI-derived functional connectomes exhibit linked dynamics,” NeuroImage, vol. 219, pp. 116998, 2020.DOI
  3. [4]Jonathan Wirsich, João Jorge, Giannina R. Iannotti, et al., “The relationship between EEG and fMRI connectomes is reproducible across simultaneous EEG–fMRI studies from 1.5T to 7T,” NeuroImage, vol. 231, pp. 117864, 2021.DOI
  4. [5]Richard E. Daws, Christopher Timmermann, Bruna Giribaldi, et al., “Increased global integration in the brain after psilocybin therapy for depression,” Nat. Med., vol. 28, no. 4, pp. 844–851, 2022.DOI
  5. [6]Suresh D. Muthukumaraswamy, Robin L. Carhart-Harris, Rosalyn J. Moran, et al., “Broadband cortical desynchronization underlies the human psychedelic state,” J. Neurosci., vol. 33, no. 38, pp. 15171–15183, 2013.DOI
  6. [7]Sin-Yee Yap, Fuad Noman, Junn Yong Loo, et al., “Brain–MGF: Multimodal graph fusion network for EEG–fMRI brain connectivity analysis under psilocybin,” arXiv:2511.18325, 2025.arxiv.org/abs/2511.18325
  7. [8]Weiheng Yao, Zhihan Lyu, Mufti Mahmud, et al., “CATD: Unified representation learning for EEG-to-fMRI cross-modal generation,” IEEE Trans. Med. Imaging, vol. 44, no. 7, pp. 2757–2767, 2025.DOI
  8. [9]Vladimir Vapnik and Akshay Vashist, “A new learning paradigm: Learning using privileged information,” Neural Netw., vol. 22, no. 5–6, pp. 544–557, 2009.DOI
  9. [10]David Lopez-Paz, Léon Bottou, Bernhard Schölkopf, et al., “Unifying distillation and privileged information,” in Proc. ICLR, 2016.
  10. [11]Geoffrey Hinton, Oriol Vinyals, and Jeff Dean, “Distilling the knowledge in a neural network,” arXiv:1503.02531, 2015.arxiv.org/abs/1503.02531
  11. [12]Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, et al., “FitNets: Hints for thin deep nets,” in Proc. ICLR, 2015.DOI
  12. [13]Xin Lin, Yuxiao Liu, Wenfei Yu, et al., “CMD-AD: Cross-modality distillation from fMRI-EEG for single-modality Alzheimer’s disease diagnosis,” in Proc. MICCAI, 2026, vol. 16886 of LNCS, pp. 174–184.DOI
  13. [14]Jintao Zhang, Yidan Ding, et al., “FRIST: fMRI representation informed shared-space training improves EEG-only individual-finger BCI decoding,” arXiv:2609.12298, 2026.arxiv.org/abs/2609.12298
  14. [15]Liqun Chen, Dong Wang, Zhe Gan, et al., “Wasserstein contrastive representation distillation,” in Proc. CVPR, 2021, pp. 16296–16305.arxiv.org/abs/2012.08674
  15. [16]Muhammad Haseeb Aslam, Muhammad Osama Zeeshan, Soufiane Belharbi, et al., “Distilling privileged multimodal information for expression recognition using optimal transport,” in Proc. IEEE FG, 2024.arxiv.org/abs/2401.15489
  16. [17]Muhammad Haseeb Aslam, Marco Pedersoli, Alessandro Lameiras Koerich, et al., “Multi teacher privileged knowledge distillation for multimodal expression recognition,” arXiv:2408.09035, 2024.arxiv.org/abs/2408.09035
  17. [18]Martin Vinck, Robert Oostenveld, Marijn van Wingerden, et al., “An improved index of phase-synchronization for electrophysiological data in the presence of volume-conduction, noise and sample-size bias,” NeuroImage, vol. 55, no. 4, pp. 1488–1565, 2011.
  18. [19]Alexander Schaefer, Ru Kong, Evan M. Gordon, et al., “Local-global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI,” Cereb. Cortex, vol. 28, no. 9, pp. 3095–3114, 2018.DOI
  19. [20]Marco Cuturi, “Sinkhorn distances: Lightspeed computation of optimal transport,” in Proc. NeurIPS, 2013, pp. 2292–2300.
  20. [21]Leonardo Novelli, Devon Stoliker, Tamrin Barta, et al., “Psi-Connect: Multimodal neuroimaging of context-dependent brain and behaviour dynamics under psilocybin,” Sci. Data, vol. 13, 2026, Art. no. 1407, doi:10.1038/s41597-026-07312-1.DOI
  21. [22]Devon Stoliker, Leonardo Novelli, Moein Khajehnejad, et al., “Psychedelics align brain activity with context,” Nature, vol. 656, no. 8129, pp. 936–947, 2026, doi:10.1038/s41586-026-10910-z.DOI
  22. [23]Adeel Razi, Devon Stoliker, and Leonardo Novelli, “Psi-Connect,” 2026, OpenNeuro dataset ds006110, version 1.2.1, doi:10.18112/openneuro.ds006110.v1.2.1.
  23. [24]Oscar Esteban, Christopher J. Markiewicz, Ross W. Blair, et al., “fMRIprep: A robust preprocessing pipeline for functional MRI,” Nat. Methods, vol. 16, no. 1, pp. 111–116, 2019.DOI
  24. [25]Neil W. Bailey, Mana Bhabani, Aron T. Hill, et al., “Introducing RELAX: An automated pre-processing pipeline for cleaning EEG data – part 1: Algorithm and application to oscillations,” Clin. Neurophysiol., vol. 149, pp. 178–201, 2023.DOI
  25. [26]Alexandre Barachant, Stéphane Bonnet, Marco Congedo, et al., “Multiclass brain–computer interface classification by Riemannian geometry,” IEEE Trans. Biomed. Eng., vol. 59, no. 4, pp. 920–928, 2012.DOI
  26. [27]Yonglong Tian, Dilip Krishnan, and Phillip Isola, “Contrastive representation distillation,” in Proc. ICLR, 2020.arxiv.org/abs/1910.10699
  27. [28]Lénaïc Chizat, Gabriel Peyré, Bernhard Schmitzer, et al., “Scaling algorithms for unbalanced optimal transport problems,” Math. Comp., vol. 87, no. 314, pp. 2563–2609, 2018.DOI

Paper details

vlnc:W_ab64qq2q
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

ChatGPT, Claude Opus 5.5 (bare)

View statement
Looking for related papers…
Contents