Pith. sign in

REVIEW 4 major objections 6 minor 77 references

Decoupling instance learning from subject labels lifts EEG diagnosis accuracy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 01:39 UTC pith:76J32YIQ

load-bearing objection Solid method-level paper with a real 14/15 empirical result; the weak spots are tuned hyperparameters and a mechanism that may learn subject identity rather than disease. the 4 major comments →

arxiv 2607.27274 v2 pith:76J32YIQ submitted 2026-07-29 cs.LG stat.ML

Rethinking EEG-Based Disease Diagnosis: Decoupling Instance Representation Learning from Subject-Level Supervision

Zhiyuan Ma , Zeyuan Li , Zhiyi Lu , Jiacheng Hao , Youlang Du , Zhen Jiang , Xinche Zhang , Yuhao Sun
show 2 more authors
Xinke Shen Sen Song
This is my paper
classification cs.LG stat.ML
keywords EEGdisease diagnosismultiple instance learningself-supervised representation learningsubject-level supervisionattention MILfeature retentionmedical time series
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that the standard practice of training EEG classifiers on short windows with the subject's disease label copied onto every window is a supervision mismatch, because some windows carry little diagnostic evidence. It introduces BridgeMIL, a two-stage framework that first learns instance representations from abundant EEG windows without any inherited labels — by aligning nearby windows and independently sampled sub-bags from the same subject — and then transfers that encoder to an attention-based multiple-instance-learning head that receives supervision only at the subject level. Across three EEG disease datasets and five backbone architectures, BridgeMIL reports the highest mean subject-level accuracy in 14 of 15 settings and an overall mean accuracy of 76.57%, 4.28 percentage points above the strongest baseline. If this holds, it would mean the field's inherited-label training is not just a harmless shortcut but an actual source of error, and that the correct supervision level is attainable without losing access to the abundance of unlabeled EEG instances.

Core claim

BridgeMIL's central claim is that subject-level EEG diagnosis should be decomposed into two problems solved at different supervision levels. Stage 1 pretrains the EEG encoder without disease labels using two positive-pair constructions: temporally nearby windows are aligned to capture local consistency, and independently sampled within-subject sub-bags are aligned to capture shared recording structure; variance and covariance regularization prevent the representation from collapsing or becoming redundant. Stage 2 then freezes that representation space as a reference, fine-tunes the encoder inside an attention-based MIL aggregator where a learned gate weights each instance's contribution to t

What carries the argument

The load-bearing mechanism is the two-scale alignment objective in Stage 1 together with the feature-retention bridge into Stage 2. Stage 1 pairs each EEG window with a nearby window and with the mean-pooled representation of an independently sampled sub-bag from the same subject; a variance-invariance-covariance objective (with target standard deviation and off-diagonal covariance penalties) keeps these pairs close without relying on negative pairs, so the encoder learns temporal consistency and within-subject shared structure without needing any instance labels. Stage 2 uses an attention-based MIL aggregator: each instance embedding gets an independent sigmoid gate, the gated embeddings ar

Load-bearing premise

The method assumes that what nearby EEG windows and separately sampled sub-bags from the same subject share is diagnostic structure; if the dominant shared structure is subject identity, recording-condition drift, or artifacts, the pretrained encoder anchors Stage 2 to non-diagnostic features and the reported gains would not generalize.

What would settle it

Train Stage 1 with disease labels randomly permuted across subjects while keeping all within-subject structure intact; if subject-level accuracy after Stage 2 remains close to the reported levels, the pretraining signal is not carrying diagnostic information. Conversely, pretrain only on healthy-control recordings and test whether the features transfer to patient classification; poor transfer would indicate the alignments are not disease-relevant.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correct, clinical EEG pipelines should stop copying the subject diagnosis onto every window and instead reserve labels for the subject-level aggregation step.
  • The approach should give its largest gains precisely where labeled subjects are scarce and instances are abundant, since the paper finds subject scarcity is the binding constraint.
  • Attention weights from the trained model can rank windows by diagnostic relevance, giving a practical way to identify which parts of a recording carry evidence.
  • The finding that inherited-label reliability varies across instances means instance-level accuracy is not a trustworthy proxy for subject-level diagnosis; evaluation should shift to subject-level metrics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the same decoupling — unsupervised instance encoding plus labels only at the bag level — should transfer to other per-patient time-series diagnostics, such as arrhythmia detection from ECG, where one label covers a long recording.
  • Editorial inference: if Stage 1 alignments encode disease-relevant structure, the pretrained encoder should transfer across EEG datasets and recording sites; this could be tested by pretraining on one dataset and fine-tuning the MIL head on another.
  • Editorial inference: the paper does not directly probe what Stage 1's shared structure represents; linear probes for disease versus subject identity on the pretrained features would settle whether the gain comes from diagnostic content or from subject or artifact alignment.
  • Editorial inference: the subject-scarcity result implies that collecting more labeled patients should matter more than recording more minutes per patient, a concrete data-collection priority for clinical deployment.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper addresses subject-level EEG disease diagnosis, arguing that inherited-label instance training and end-to-end MIL both have weaknesses. It proposes BridgeMIL: Stage 1 uses a VICReg-style objective to align nearby windows and within-subject sub-bags without any instance labels, producing a pretrained encoder; Stage 2 attaches an attention-based MIL aggregator, fine-tunes with subject-level cross-entropy, and adds an instance-level feature-retention loss that anchors the encoder to its Stage 1 representation. The method is evaluated on ADFTD, Mumtaz2017, and Rockhill2021 with five backbones, reporting the highest mean accuracy in 14 of 15 dataset-backbone settings and an overall mean accuracy of 76.57%. Additional analyses study inherited-label reliability, subject vs. instance scarcity, ablations, and representation structure.

Significance. If the empirical results hold, BridgeMIL is a practically useful two-stage recipe for small-cohort EEG diagnosis: it obtains a strong encoder from abundant unlabeled instances and preserves it during a lightly supervised MIL fine-tune. The experimental design has real strengths: subject-level stratified cross-validation is used consistently, the coverage of multiple backbones and datasets is commendable, the baselines are reasonably comprehensive, and the headline accuracy arithmetic is reproducible from Table 1. However, the central mechanistic claim—that the Stage 1 alignments capture disease-relevant structure rather than subject identity or recording-condition drift—is not directly established. Given the small subject counts (88, 63, 31), this limits the strength of the external-generalization conclusion. The paper would be substantially strengthened by a subject-identity control or a held-out-site experiment, a nested or fixed hyperparameter-evaluation statement, and a deconfounded scarcity analysis.

major comments (4)
  1. [Section 3.1, Eq. (9), Figure 5] The Stage 1 objective (Eqs. (2)-(4), (9)) aligns nearby windows and same-subject sub-bags, so a representation that encodes subject identity or recording-condition drift would satisfy the objective as easily as one that encodes disease-relevant structure. VICReg prevents collapse but does not select disease-relevant variance over subject-identity variance. The paper's t-SNE (Fig. 5) actually displays distinct subject-wise clusters, and the SC/DB indices are computed only with diagnostic classes as targets; no subject-identity probe or clustering metric is reported. This is load-bearing for the central claim and for external validity: with 88, 63, and 31 subjects, subject-specific nuisance can correlate with diagnosis and inflate in-dataset accuracy. Please add a concrete control, e.g., a linear probe for subject identity versus diagnosis, a leave-one-site-out experiment, or a baseline th
  2. [Appendix F, Table 3] The hyperparameters λ_ret and K are tuned on ADFTD with EEGNet and then used for all datasets and backbones in Table 1, without nested cross-validation or a clear statement that they were fixed before the main evaluation. Table 3 shows the adopted values are the best among those tested on ADFTD, which means the ADFTD column (and hence the overall 14/15 and 76.57% claims) may be optimistically biased. Please either report nested CV for hyperparameter selection, show the sensitivity results on the other two datasets, or explicitly document that the same hyperparameters were chosen a priori and held fixed; otherwise the main comparison is not a fully honest evaluation.
  3. [Section RQ3, Figure 4] The subject-scarcity experiment is confounded. Each condition draws a random subset of 8, 6, 4, or 3 folds from the 10 folds and then selects validation and test folds from that subset, so the test subject set changes across conditions. The accuracy differences in Fig. 4 therefore reflect both the number of training folds and the identity/composition of the test folds, violating the intended controlled comparison. To support the claim that subject scarcity is more limiting than instance scarcity, fix the test and validation folds across all conditions and vary only the number of training subjects; alternatively use random-subject subsampling within a fixed split. The instance-scarcity arm is not affected by this issue.
  4. [Section RQ2, Figure 3 right] The Top-K/Bottom-K analysis is partially circular. Instances are ranked by BridgeMIL attention weights, which are learned from subject-level labels; selecting the top-K instances and then training an inherited-label classifier on them is likely to select instances that are label-consistent by construction, so the observed accuracy ordering does not independently establish that inherited labels are unreliable. It mainly shows that the attention model's own notion of relevance transfers to a different classifier. Add a non-model-based reliability criterion (e.g., spectral/artifact features or agreement across independently trained instance classifiers), or at least a permutation control, and report paired significance tests for the Top-K vs Bottom-K comparisons.
minor comments (6)
  1. [Eq. (12)] The symbol b_a is used for both the input bias inside tanh and the scalar output bias; the dimensions of the two are different. Please use distinct symbols and clarify whether the output gate bias is a scalar or a per-instance quantity.
  2. [Figure 3] The right panel does not label its horizontal axis as K (20, 50, 100, 200), and the error bars/confidence regions are not defined. Please add axis labels and a caption statement describing what the shaded bands represent.
  3. [Table 1] All results are reported as mean±std over five seeds, but no significance tests are provided. Given the small number of subjects, paired bootstrap or permutation tests (or confidence intervals) would make the 'highest mean accuracy in 14 of 15 settings' claim more convincing.
  4. [Appendix D] The validation-based checkpoint selection uses different burn-in epochs for different datasets (after epoch 10 for ADFTD, after epoch 15 for Rockhill2021, none for Mumtaz2017). Please explain this choice; otherwise it looks like dataset-specific tuning.
  5. [Appendix E, Table 2] For the 'w/o Sub-bag' and 'w/o Nearby' ablations, please state explicitly whether the removed loss term is simply dropped or replaced by the remaining term, and whether the loss weights are renormalized. This affects the interpretation of the ablation.
  6. [General] The terms 'seed' and 'fold' are used in overlapping ways (random seed vs. cross-validation fold). Define both once and use them consistently, especially in the scarcity and ablation sections.

Circularity Check

2 steps flagged

Main result is a fair empirical comparison; minor circularity in auxiliary reliability analysis and ADFTD hyperparameter selection.

specific steps
  1. fitted input called prediction [Section RQ2 (Inherited-Label Reliability Analysis), Figure 3 right panel]
    "To test whether inherited labels are more reliable for some instances than others, we compare EEGNet trained on Top-K and Bottom-K instances ranked by BridgeMIL attention weights."

    The attention weights are learned by BridgeMIL from the same subject-level labels, and instances are then ranked by these weights. Higher accuracy of classifiers trained on Top-K instances is thus partly a self-confirmation of the model's own attention, not an independent measure of inherited-label reliability. The paper uses this to conclude that inherited labels are unequally reliable, but the ranking and the outcome share the same learned model.

  2. fitted input called prediction [Appendix F (Hyperparameter Sensitivity), Table 3]
    "This analysis is conducted on ADFTD with EEGNet. We vary one hyperparameter at a time while holding the other at the setting used in the main experiments... the configuration used in the main experiments achieves the highest mean accuracy among the tested values."

    The hyperparameters λ_ret and K are selected on ADFTD based on the reported accuracy, and the same ADFTD accuracy is then presented as one of the 14/15 winning settings. This makes the ADFTD result partly a fitted value rather than an independent prediction. Since the other two datasets share the same hyperparameters without tuning, the impact is limited but still a circular step for the ADFTD claim.

full rationale

BridgeMIL's primary claim—highest mean accuracy in 14/15 dataset-backbone settings—comes from a held-out subject-level cross-validation against external baselines (majority vote, MIL, SupCon, masked reconstruction). Stage 1 uses no labels and Stage 2 supervises only subject-level predictions, so the accuracy result is not derived from the losses by construction. The reliability analysis (Top-K vs Bottom-K) uses BridgeMIL's own attention weights to rank instances, making the conclusion that instances differ in label reliability partially self-confirmatory; the agreement-ratio analysis provides independent support. Additionally, hyperparameters for the main experiments were selected on ADFTD and the ADFTD accuracy is reported for that same dataset, injecting a mild selection bias. Neither issue affects the independent content of the main accuracy comparison across the other datasets.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The framework is built from standard SSL and MIL machinery plus a domain assumption about temporal/sub-bag consistency in EEG. The main free parameters are hyperparameters tuned on one dataset, and no new physical or conceptual entities are introduced.

free parameters (6)
  • Feature-retention coefficient λret = 1e-5
    Controls how strongly Stage 2 anchors encoder to Stage 1 representation; tuned on ADFTD (Table 3) and fixed for all datasets.
  • Sub-bag size K = 8 windows
    Size of independently sampled sub-bags in Stage 1; tuned on ADFTD (Appendix F) and fixed for all datasets.
  • VICReg loss weights λinv, λvar, λcov = 25, 25, 1
    Weights in Eq (19); chosen by hand following VICReg defaults.
  • Temporal radius r = 2 steps
    Nearby-window pairing radius in Eq (2); chosen by hand (Appendix D).
  • Projection head dimensions = d→256→128
    Architecture of qφ in Eq (18); chosen by hand.
  • Stage-1/Stage-2 epochs and learning rates = 50/100 epochs; 1e-3/1e-4
    Training schedule; chosen by hand (Appendix D).
axioms (5)
  • standard math VICReg variance and covariance regularization prevents representation collapse without negative pairs.
    Invoked in Eqs (6)-(7); accepted SSL result from Bardes et al. (2022).
  • domain assumption Each subject has a single disease label and all instances in a bag inherit it for bag-level supervision only.
    Problem formulation Eq (1); this is the paper's setup, and BridgeMIL avoids assigning instance-level labels.
  • domain assumption Temporally nearby EEG windows and two random sub-bags from the same subject share latent structure relevant to the subject's diagnosis.
    Stage 1 objective Eqs (2)-(4) and Eq (9); if false, Stage 1 pretraining would learn non-diagnostic structure.
  • domain assumption Attention MIL with independent sigmoid gates and mean pooling can represent subject-level evidence.
    Aggregator in Eqs (12)-(13); unproven but supported by empirical comparison to TimeMIL.
  • domain assumption Channel-wise z-score statistics computed on training folds generalize to validation/test folds.
    Preprocessing in Section 4.1/Appendix D; a standard but unstated assumption.

pith-pipeline@v1.3.0-alltime-deepseek · 18158 in / 15212 out tokens · 149932 ms · 2026-08-03T01:39:10.166508+00:00 · methodology

0 comments
read the original abstract

EEG-based disease diagnosis requires one prediction per subject, yet common pipelines segment recordings into short instances, inherit the subject label for every instance, and train instance-level classifiers. This assumes that all instances provide equally reliable diagnostic evidence. Multiple instance learning (MIL) avoids inherited labels by treating each subject as a bag. However, EEG datasets contain far fewer subjects than instances, which can limit the quality of the representations learned by end-to-end MIL. We propose BridgeMIL, a two-stage framework that decouples instance representation learning from subject-level supervision. Stage 1 pretrains the encoder without inherited instance labels by aligning temporally nearby windows and independently sampled within-subject sub-bags. Variance and covariance regularization prevent collapse and reduce redundancy without negative pairs. Stage 2 transfers the encoder to an attention-based MIL aggregator, applies supervision only to subject predictions, and limits representation drift through feature retention. Across three EEG disease datasets and five representative backbones, BridgeMIL attains the highest mean accuracy in 14 of 15 dataset-backbone settings and an overall mean accuracy of 76.57%, 4.28 percentage points higher than the strongest baseline. Further analyses reveal substantial variation in inherited-label reliability across instances, greater performance sensitivity to subject scarcity than to instance scarcity, and a more structured representation space with distinct subject-wise clusters and improved separation between diagnostic classes. Together, these findings underscore the importance of aligning supervision with the subject-level prediction objective while learning from abundant EEG instances without assigning disease labels to individual instances.

Figures

Figures reproduced from arXiv: 2607.27274 by Jiacheng Hao, Sen Song, Xinche Zhang, Xinke Shen, Youlang Du, Yuhao Sun, Zeyuan Li, Zhen Jiang, Zhiyi Lu, Zhiyuan Ma.

Figure 1
Figure 1. Figure 1: Motivation for BridgeMIL. EEG segments from the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of BridgeMIL. Stage 1 learns the EEG encoder without inherited instance labels by aligning temporally [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Inherited-label reliability analysis. Left: Per￾subject instance agreement ratios across three datasets, with horizontal bars indicating the mean ratio for each dataset. Right: Subject-level accuracy on ADFTD using Top-K, Bottom-K, or all training instances. Results are mean ± stan￾dard deviation over five seeds. in 14 of 15 dataset–backbone settings and outperforms ma￾jority voting and every end-to-end MI… view at source ↗
Figure 5
Figure 5. Figure 5: Ablation and representation analysis on Rockhill2021 with EEGNet. Stage 1 compares raw windows, pretraining [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

77 extracted references · 2 linked inside Pith

  1. [1]

    Chen, X.; Qiu, P.; Zhu, W.; Li, H.; Wang, H.; Sotiras, A.; Wang, Y.; and Razi, A. 2024. TimeMIL: advancing multivariate time series classification via a time-aware multiple instance learning . In Proceedings of the 41st International Conference on Machine Learning, 7190--7206

    2024

  2. [2]

    Early, J.; Cheung, G.; Cutajar, K.; Xie, H.; Kandola, J.; and Twomey, N. 2024. Inherently interpretable time series classification via multiple instance learning . In International Conference on Learning Representations, volume 2024, 56800--56828

    2024

  3. [3]

    Ilse, M.; Tomczak, J.; and Welling, M. 2018. Attention-based deep multiple instance learning . In International conference on machine learning, 2127--2136. PMLR

    2018

  4. [4]

    A.; Juyal, D.; Padigela, H.; Taylor-Weiner, A.; Yu, L.; and Prakash, A

    Javed, S. A.; Juyal, D.; Padigela, H.; Taylor-Weiner, A.; Yu, L.; and Prakash, A. 2022. Additive mil: Intrinsically interpretable multiple instance learning for pathology . Advances in Neural Information Processing Systems, 35: 20689--20702

    2022

  5. [5]

    B.; Al-Shourbaji, I.; Jabbari, A.; Kirner, R.; Alameen, A.; Migdady, H.; and Abualigah, L

    Kachare, P.; Puri, D.; Sangle, S. B.; Al-Shourbaji, I.; Jabbari, A.; Kirner, R.; Alameen, A.; Migdady, H.; and Abualigah, L. 2024. LCADNet: a novel light CNN architecture for EEG-based Alzheimer disease detection . Physical and Engineering Sciences in Medicine, 47(3): 1037--1050

    2024

  6. [6]

    Khosla, P.; Teterwak, P.; Wang, C.; Sarna, A.; Tian, Y.; Isola, P.; Maschinot, A.; Liu, C.; and Krishnan, D. 2020. Supervised contrastive learning . Advances in neural information processing systems, 33: 18661--18673

    2020

  7. [7]

    J.; Solon, A

    Lawhern, V. J.; Solon, A. J.; Waytowich, N. R.; Gordon, S. M.; Hung, C. P.; and Lance, B. J. 2018. EEGNet: a compact convolutional neural network for EEG-based brain--computer interfaces . Journal of neural engineering, 15(5): 056013

    2018

  8. [8]

    Ma, Z.; Li, Z.; Qiu, Z.; Li, J.; Meng, L.; Zhang, X.; Liu, Y.; Shen, X.; and Song, S. 2026. DSAINet: An Efficient Dual-Scale Attentive Interaction Network for General EEG Decoding . arXiv preprint arXiv:2604.18095

    Pith/arXiv arXiv 2026

  9. [9]

    D.; Afrantou, T.; Ioannidis, P.; Grigoriadis, N.; Tsalikakis, D

    Miltiadous, A.; Tzimourta, K. D.; Afrantou, T.; Ioannidis, P.; Grigoriadis, N.; Tsalikakis, D. G.; Angelidis, P.; Tsipouras, M. G.; Glavas, E.; Giannakeas, N.; et al. 2023. A dataset of scalp EEG recordings of Alzheimer’s disease, frontotemporal dementia and healthy subjects from routine EEG . Data, 8(6): 95

    2023

  10. [10]

    Mumtaz, W.; Xia, L.; Ali, S. S. A.; Yasin, M. A. M.; Hussain, M.; and Malik, A. S. 2017. Electroencephalogram (EEG)-based computer-aided technique to diagnose major depressive disorder (MDD) . Biomedical Signal Processing and Control, 31: 108--115

    2017

  11. [11]

    P.; Jackson, N.; George, J.; Aron, A.; and Swann, N

    Rockhill, A. P.; Jackson, N.; George, J.; Aron, A.; and Swann, N. C. 2021. UC San Diego Resting State EEG Data from Patients with Parkinson's Disease . OpenNeuro

    2021

  12. [12]

    Song, Y.; Zheng, Q.; Liu, B.; and Gao, X. 2022. EEG conformer: Convolutional transformer for EEG decoding and visualization . IEEE Transactions on Neural Systems and Rehabilitation Engineering, 31: 710--719

    2022

  13. [13]

    Wang, X.; Yan, Y.; Tang, P.; Bai, X.; and Liu, W. 2018. Revisiting multiple instance neural networks . Pattern recognition, 74: 15--24

    2018

  14. [14]

    Zerveas, G.; Jayaraman, S.; Patel, D.; Bhamidipaty, A.; and Eickhoff, C. 2021. A transformer-based framework for multivariate time series representation learning . In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, 2114--2124

    2021

  15. [15]

    Zini, S.; Barbera, T.; Bianco, S.; and Napoletano, P. 2026. Alzheimer’s disease classification from EEG using a multiscale temporal deep network . Biomedical Signal Processing and Control, 114: 109321

    2026

  16. [16]

    2023 , publisher=

    Parsa, Mohsen and Rad, Habib Yousefi and Vaezi, Hadi and Hossein-Zadeh, Gholam-Ali and Setarehdan, Seyed Kamaledin and Rostami, Reza and Rostami, Hana and Vahabie, Abdol-Hossein , journal=. 2023 , publisher=

    2023

  17. [17]

    Journal of Neural Engineering , volume=

    Sadatnejad, Khadijeh and Rahmati, Mohammad and Rostami, Reza and Kazemi, Reza and Ghidary, Saeed S and M. Journal of Neural Engineering , volume=. 2019 , publisher=

    2019

  18. [18]

    2025 , organization=

    Gijsen, Sam and Ritter, Kerstin , booktitle=. 2025 , organization=

    2025

  19. [19]

    Liu, Kangning and Zhu, Weicheng and Shen, Yiqiu and Liu, Sheng and Razavian, Narges and Geras, Krzysztof J and Fernandez-Granda, Carlos , booktitle=

  20. [20]

    Li, Bin and Li, Yin and Eliceiri, Kevin W , booktitle=

  21. [21]

    Fan, Wei and Fei, Jingru and Guo, Dingyu and Yi, Kun and Song, Xiaozhuang and Xiang, Haolong and Ye, Hangting and Li, Min , booktitle=

  22. [22]

    Wang, Yihe and Huang, Nan and Li, Taida and Yan, Yujun and Zhang, Xiang , journal=

  23. [23]

    Ye, Jiexia and Zhang, Weiqi and Li, Ziyue and Li, Jia and Tsung, Fugee , booktitle=

  24. [24]

    Yu, Guoqi and Wang, Juncheng and Yang, Chen and Qin, Jing and Aviles-Rivero, Angelica I and Wang, Shujun , journal=

  25. [25]

    2024 , publisher=

    Chen, Sheng and Zhang, Chutian and Yang, Hongjun and Peng, Liang and Xie, Haiqun and Lv, Zeping and Hou, Zeng-Guang , journal=. 2024 , publisher=

    2024

  26. [26]

    2025 , publisher=

    Xiao, Jun and Qi, Feifei and Wang, Lingli and He, Yanbin and Yu, Jingang and Wu, Wei and Yu, Zhuliang and Li, Yuanqing and Gu, Zhenghui and Yu, Tianyou , journal=. 2025 , publisher=

    2025

  27. [27]

    2026 , publisher=

    Jin, Changgyun and Kim, Hanul and Na, Yeongseo and Kim, Seong-Eun , journal=. 2026 , publisher=

    2026

  28. [28]

    2025 , publisher=

    Larbi, Arezki and Abed, Mansour and Cardoso, Jaime S and Ouahabi, Abdeljalil , journal=. 2025 , publisher=

    2025

  29. [29]

    Jang, Jaeseok and Kwon, Hyuk-Yoon , booktitle=

  30. [30]

    Journal of Neural Engineering , volume=

    Banville, Hubert and Chehab, Omar and Hyv. Journal of Neural Engineering , volume=. 2021 , publisher=

    2021

  31. [31]

    Cai, Donghong and Chen, Junru and Yang, Yang and Liu, Teng and Li, Yafeng , booktitle=

  32. [32]

    Wang, Yihe and Han, Yu and Wang, Haishuai and Zhang, Xiang , journal=

  33. [33]

    Bardes, Adrien and Ponce, Jean and Lecun, Yann , booktitle=

  34. [34]

    Weng, Weining and Gu, Yang and Ma, Yuan and Liu, Yuchen and Zhang, Yingwei and Chen, Yiqiang , booktitle=

  35. [35]

    Jiang, Wei-Bang and Zhao, Liming and Lu, Bao-Liang , booktitle=

  36. [36]

    Wang, Jiquan and Zhao, Sha and Luo, Zhiling and Zhou, Yangxuan and Jiang, Haiteng and Li, Shijian and Li, Tao and Pan, Gang , booktitle=

  37. [37]

    Advances in Neural Information Processing Systems , volume=

    El Ouahidi, Yassine and Lys, Jonathan and Th. Advances in Neural Information Processing Systems , volume=

  38. [38]

    2017 , publisher=

    Mumtaz, Wajid and Xia, Likun and Ali, Syed Saad Azhar and Yasin, Mohd Azhar Mohd and Hussain, Muhammad and Malik, Aamir Saeed , journal=. 2017 , publisher=

    2017

  39. [39]

    2023 , publisher=

    Miltiadous, Andreas and Tzimourta, Katerina D and Afrantou, Theodora and Ioannidis, Panagiotis and Grigoriadis, Nikolaos and Tsalikakis, Dimitrios G and Angelidis, Pantelis and Tsipouras, Markos G and Glavas, Euripidis and Giannakeas, Nikolaos and others , journal=. 2023 , publisher=

    2023

  40. [40]

    Rockhill and Nicko Jackson and Jobi George and Adam Aron and Nicole C

    Alexander P. Rockhill and Nicko Jackson and Jobi George and Adam Aron and Nicole C. Swann , title=. 2021 , doi=

    2021

  41. [41]

    Chen, Xiwen and Qiu, Peijie and Zhu, Wenhui and Li, Huayu and Wang, Hao and Sotiras, Aristeidis and Wang, Yalin and Razi, Abolfazl , booktitle=

  42. [42]

    Early, Joseph and Cheung, Gavin and Cutajar, Kurt and Xie, Hanting and Kandola, Jas and Twomey, Niall , booktitle=

  43. [43]

    Khosla, Prannay and Teterwak, Piotr and Wang, Chen and Sarna, Aaron and Tian, Yonglong and Isola, Phillip and Maschinot, Aaron and Liu, Ce and Krishnan, Dilip , journal=

  44. [44]

    Zerveas, George and Jayaraman, Srideepika and Patel, Dhaval and Bhamidipaty, Anuradha and Eickhoff, Carsten , booktitle=

  45. [45]

    2018 , publisher=

    Wang, Xinggang and Yan, Yongluan and Tang, Peng and Bai, Xiang and Liu, Wenyu , journal=. 2018 , publisher=

    2018

  46. [46]

    2018 , organization=

    Ilse, Maximilian and Tomczak, Jakub and Welling, Max , booktitle=. 2018 , organization=

    2018

  47. [47]

    Javed, Syed Ashar and Juyal, Dinkar and Padigela, Harshith and Taylor-Weiner, Amaro and Yu, Limin and Prakash, Aaditya , journal=

  48. [48]

    Ma, Zhiyuan and Li, Zeyuan and Qiu, Zihao and Li, Jinhao and Meng, Lingqin and Zhang, Xinche and Liu, Yixuan and Shen, Xinke and Song, Sen , journal=

  49. [49]

    2018 , publisher=

    Lawhern, Vernon J and Solon, Amelia J and Waytowich, Nicholas R and Gordon, Stephen M and Hung, Chou P and Lance, Brent J , journal=. 2018 , publisher=

    2018

  50. [50]

    2022 , publisher=

    Song, Yonghao and Zheng, Qingqing and Liu, Bingchuan and Gao, Xiaorong , journal=. 2022 , publisher=

    2022

  51. [51]

    2024 , publisher=

    Kachare, Pramod and Puri, Digambar and Sangle, Sandeep B and Al-Shourbaji, Ibrahim and Jabbari, Abdoh and Kirner, Raimund and Alameen, Abdalla and Migdady, Hazem and Abualigah, Laith , journal=. 2024 , publisher=

    2024

  52. [52]

    2026 , publisher=

    Zini, Simone and Barbera, Thomas and Bianco, Simone and Napoletano, Paolo , journal=. 2026 , publisher=

    2026

  53. [53]

    Bomatter, Philipp and Gouk, Henry , journal=

  54. [54]

    Banville, H.; Chehab, O.; Hyv \"a rinen, A.; Engemann, D.-A.; and Gramfort, A. 2021. Uncovering the structure of clinical EEG signals with self-supervised learning . Journal of Neural Engineering, 18(4): 046020

    2021

  55. [55]

    Bardes, A.; Ponce, J.; and Lecun, Y. 2022. VICReg: Variance-Invariance-Covariance Regularization For Self-Supervised Learning . In ICLR 2022-International Conference on Learning Representations

    2022

  56. [56]

    Bomatter, P.; and Gouk, H. 2026. Is limited participant diversity impeding eeg-based machine learning? Advances in Neural Information Processing Systems, 38: 117447--117474

    2026

  57. [57]

    Cai, D.; Chen, J.; Yang, Y.; Liu, T.; and Li, Y. 2023. Mbrain: A multi-channel self-supervised learning framework for brain signals . In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 130--141

    2023

  58. [58]

    Chen, S.; Zhang, C.; Yang, H.; Peng, L.; Xie, H.; Lv, Z.; and Hou, Z.-G. 2024 a . A multi-modal classification method for early diagnosis of mild cognitive impairment and Alzheimer’s disease using three paradigms with various task difficulties . IEEE Transactions on Neural Systems and Rehabilitation Engineering, 32: 1477--1486

    2024

  59. [59]

    Chen, X.; Qiu, P.; Zhu, W.; Li, H.; Wang, H.; Sotiras, A.; Wang, Y.; and Razi, A. 2024 b . TimeMIL: advancing multivariate time series classification via a time-aware multiple instance learning . In Proceedings of the 41st International Conference on Machine Learning, 7190--7206

    2024

  60. [60]

    El Ouahidi, Y.; Lys, J.; Th \"o lke, P.; Farrugia, N.; Pasdeloup, B.; Gripon, V.; Jerbi, K.; and Lioi, G. 2026. REVE: A foundation model for EEG-adapting to any setup with large-scale pretraining on 25,000 subjects . Advances in Neural Information Processing Systems, 38: 22541--22577

    2026

  61. [61]

    Fan, W.; Fei, J.; Guo, D.; Yi, K.; Song, X.; Xiang, H.; Ye, H.; and Li, M. 2025. Towards multi-resolution spatiotemporal graph learning for medical time series classification . In Proceedings of the ACM on Web Conference 2025, 5054--5064

    2025

  62. [62]

    Gijsen, S.; and Ritter, K. 2025. EEG-Language Pretraining for Highly Label-Efficient Clinical Phenotyping . In International Conference on Machine Learning, 19480--19504. PMLR

    2025

  63. [63]

    Jang, J.; and Kwon, H.-Y. 2025. TAIL-MIL: time-aware and instance-learnable multiple instance learning for multivariate time series anomaly detection . In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 17582--17589

    2025

  64. [64]

    Jiang, W.-B.; Zhao, L.; and Lu, B.-L. 2024. Large brain model for learning generic representations with tremendous EEG data in BCI . In International Conference on Learning Representations, volume 2024, 16405--16426

    2024

  65. [65]

    Jin, C.; Kim, H.; Na, Y.; and Kim, S.-E. 2026. MILFNet: Multiple Instance Learning-Based FastSlow Network With Multitask Autoencoder for Cross-Subject EEG Emotion Recognition . IEEE Transactions on Instrumentation and Measurement

    2026

  66. [66]

    S.; and Ouahabi, A

    Larbi, A.; Abed, M.; Cardoso, J. S.; and Ouahabi, A. 2025. Neonatal EEG classification using a compact support separable kernel time--frequency distribution and attention-based CNN . Biomedical Signal Processing and Control, 110: 108072

    2025

  67. [67]

    Li, B.; Li, Y.; and Eliceiri, K. W. 2021. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning . In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 14318--14328

    2021

  68. [68]

    J.; and Fernandez-Granda, C

    Liu, K.; Zhu, W.; Shen, Y.; Liu, S.; Razavian, N.; Geras, K. J.; and Fernandez-Granda, C. 2023. Multiple instance learning via iterative self-paced supervised contrastive learning . In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 3355--3365

    2023

  69. [69]

    Y.; Vaezi, H.; Hossein-Zadeh, G.-A.; Setarehdan, S

    Parsa, M.; Rad, H. Y.; Vaezi, H.; Hossein-Zadeh, G.-A.; Setarehdan, S. K.; Rostami, R.; Rostami, H.; and Vahabie, A.-H. 2023. EEG-based classification of individuals with neuropsychiatric disorders using deep neural networks: A systematic review of current status and future directions . Computer Methods and Programs in Biomedicine, 240: 107683

    2023

  70. [70]

    S.; M \"u ller, A.; and Alimardani, F

    Sadatnejad, K.; Rahmati, M.; Rostami, R.; Kazemi, R.; Ghidary, S. S.; M \"u ller, A.; and Alimardani, F. 2019. EEG representation using multi-instance framework on the manifold of symmetric positive definite matrices . Journal of Neural Engineering, 16(3): 036016

    2019

  71. [71]

    Wang, J.; Zhao, S.; Luo, Z.; Zhou, Y.; Jiang, H.; Li, S.; Li, T.; and Pan, G. 2025. Cbramod: A criss-cross brain foundation model for eeg decoding . In International conference on learning representations, volume 2025, 75310--75346

    2025

  72. [72]

    Wang, Y.; Han, Y.; Wang, H.; and Zhang, X. 2023. Contrast everything: A hierarchical contrastive framework for medical time-series . Advances in Neural Information Processing Systems, 36: 55694--55717

    2023

  73. [73]

    Wang, Y.; Huang, N.; Li, T.; Yan, Y.; and Zhang, X. 2024. Medformer: A multi-granularity patching transformer for medical time-series classification . Advances in Neural Information Processing Systems, 37: 36314--36341

    2024

  74. [74]

    Weng, W.; Gu, Y.; Ma, Y.; Liu, Y.; Zhang, Y.; and Chen, Y. 2026. State Mamba: Spatiotemporal EEG State-Space Model with Dynamic Brain Alignment for Cross-Subject Representation . In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 17850--17858

    2026

  75. [75]

    Xiao, J.; Qi, F.; Wang, L.; He, Y.; Yu, J.; Wu, W.; Yu, Z.; Li, Y.; Gu, Z.; and Yu, T. 2025. EmotionMIL: An End-to-End Multiple Instance Learning Framework for Emotion Recognition From EEG Signals . IEEE Transactions on Affective Computing

    2025

  76. [76]

    Ye, J.; Zhang, W.; Li, Z.; Li, J.; and Tsung, F. 2026. MedSpaformer: a Transferable Transformer with Multi-granularity Token Sparsification for Medical Time Series Classification . In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 27791--27799

    2026

  77. [77]

    I.; and Wang, S

    Yu, G.; Wang, J.; Yang, C.; Qin, J.; Aviles-Rivero, A. I.; and Wang, S. 2026. Decentralized attention fails centralized signals: Rethinking transformers for medical time series . arXiv preprint arXiv:2602.18473

    Pith/arXiv arXiv 2026