Training Paradigms for Affective EEG

A training paradigm specifies what supervision a model receives and what learning signal shapes its parameters. The same EEG dataset can support supervised classification, self-supervised representation learning, unsupervised discovery, or reinforcement learning, but these formulations require different data, validation procedures, and claims. Selecting a paradigm should follow the available labels and deployment objective rather than the popularity of a model family.

Supervised Training

Supervised learning trains a model from EEG inputs $x_i$ and target labels $y_i$. For classification, the target may be a discrete emotion category; for regression, it may be valence, arousal, stress intensity, or another continuous rating. The model minimizes a loss that compares its prediction $f_\theta(x_i)$ with the target, for example cross-entropy for classes or mean squared error for a continuous score.

Supervised training is the standard choice when labels are sufficiently reliable and match the intended prediction unit. It requires careful handling of label imbalance, inherited trial labels, noisy self-reports, and subject or session variation. All label preprocessing, including thresholding continuous ratings into classes, should be defined using training data and held fixed for validation and testing.

Example scenario: A researcher uses the DEAP dataset to predict high versus low arousal from full trials. The median threshold is determined from training subjects only, the classifier is trained on their EEG, and the final test evaluates unseen subjects. This is supervised subject-independent classification, not continuous affect tracking.

Unsupervised Training

Unsupervised learning uses EEG inputs without affect labels to discover structure in the data. The goal may be clustering participants or signal states, estimating a latent representation, detecting unusual recordings, or modeling the density of normal EEG segments. Its outputs are not automatically emotion categories; they require scientific interpretation and, when possible, validation against independent labels or experimental conditions.

Typical methods include clustering, dimensionality reduction, density estimation, and unsupervised representation learning. The number of clusters, preprocessing choices, and selection criteria should be determined without inspecting test labels. Cluster stability across subjects, sessions, preprocessing variants, and random seeds is often more informative than a single clustering score.

Example scenario: In a naturalistic study with no trial-level emotion labels, a researcher clusters spectral-spatial embeddings of EEG windows to identify recurring arousal-related states. The discovered clusters are then compared with independently collected behavioral reports that were not used to construct the clusters.

Self-Supervised Training

Self-supervised learning constructs a training signal from the data itself, then transfers the learned representation to a labeled downstream task. It is useful when unlabeled EEG is plentiful but affect annotations are scarce, expensive, delayed, or noisy. The pretraining corpus must be separated from downstream test data according to the intended generalization claim: using test-subject recordings during pretraining may be acceptable only when transductive or target-domain adaptation is part of the declared protocol.

Generative Self-Supervision

Generative self-supervision learns to reconstruct, predict, or model part of the EEG signal from another part. Examples include masked-sample reconstruction, masked-channel prediction, future-window prediction, denoising, and autoencoding. The objective encourages representations that preserve temporal, spectral, and spatial structure.

Example scenario: A masked autoencoder receives a 10-second multichannel EEG segment with several time spans and channels hidden. It learns to reconstruct the missing values from visible context, then its encoder is fine-tuned to classify valence using a small labeled training set.

Contrastive Self-Supervision

Contrastive self-supervision brings embeddings of related EEG views closer together and separates unrelated views. Related views can be two augmentations of the same window, neighboring windows from the same trial, or recordings from the same subject under a carefully defined invariant. The choice of positive and negative pairs encodes the scientific assumption of what should remain stable.

Augmentations must preserve the property needed by the downstream task. An augmentation that destroys phase, a critical frequency band, or a spatial relationship may produce an invariant representation that is unsuitable for the target. False negatives are also a concern: windows from different trials may share the same affective state even if the contrastive objective treats them as unrelated.

Example scenario: Two mild augmentations of the same artifact-screened EEG window, such as small amplitude scaling and limited channel dropout, form a positive pair. Windows from other subjects form negatives. The pretrained encoder is later evaluated for cross-subject emotion classification using only labeled source-subject data.

Pseudo-Label Self-Supervision

Pseudo-label methods construct provisional targets from the data or from a teacher model. Examples include predicting temporal order, identifying which channel was masked, clustering embeddings and predicting cluster assignments, or training a student model on high-confidence predictions from a teacher. Pseudo-labels can increase the effective training signal, but they can also amplify early model mistakes or encode artifacts rather than affective structure.

Use confidence thresholds, teacher-student consistency checks, or repeated clustering stability to control pseudo-label quality. Pseudo-label generation must be fitted on training data or on explicitly allowed unlabeled target data; it must never use held-out test labels.

Example scenario: A teacher encoder clusters unlabeled training EEG windows into stable prototype groups. Only windows assigned consistently across augmentations receive a pseudo-label. A student model is trained on those pseudo-labels and then fine-tuned using the available trial-level emotion labels.

Representation Learning for Affective EEG

Representation learning aims to construct an embedding in which EEG segments that are relevant to the scientific task are close, while nuisance variation is separated or made easier to adapt away. It is broader than a particular training paradigm: the representation may be learned through supervised labels, self-supervised pretraining, contrastive objectives, multi-task learning, or a combination of these. The downstream classifier or regressor is then only one possible consumer of the learned representation.

For affective EEG, the representation should reflect both the structure of the signal and the structure of emotion. A useful embedding may preserve temporal, spectral, and spatial information while organizing samples according to valence, arousal, or another theory-informed affect space. This can be more transferable than optimizing a classifier for the label vocabulary of one dataset, especially when datasets use different discrete categories, rating scales, channel layouts, or sampling conventions.

V-A-Guided Contrastive Learning

Valence-arousal (V-A) coordinates provide one way to connect heterogeneous emotion annotations. Discrete categories can be assigned or mapped to approximate locations in the V-A plane, while continuous ratings can be used directly. Instead of treating every non-identical label as an equally strong negative, a soft-weighted contrastive objective can make emotionally similar samples attract one another more strongly than emotionally distant samples. The result is an embedding with a graded notion of affective similarity rather than only dataset-specific class boundaries.

The choice of positive and negative relationships remains a scientific assumption. V-A proximity does not guarantee equivalence in appraisal, physiology, or temporal dynamics, and labels inherited from a whole trial may not apply to every local window. Report the mapping from labels to V-A coordinates, the temperature and weighting rules of the contrastive loss, the augmentations used, and whether subject or session identity is controlled in pair construction.

EMOD as a Cross-Dataset Example

EMOD is a recent example of affect-aware representation learning. Chen et al. (2026) propose a unified EEG emotion representation framework that projects discrete and continuous emotion labels into a common valence-arousal space and uses V-A-guided soft-weighted supervised contrastive learning. Its Triple-Domain Encoder and Spatial-Temporal Transformer integrate temporal, spectral, and spatial information while accommodating heterogeneous EEG formats. The authors pretrain on eight public EEG datasets and evaluate on three benchmark datasets, making the work relevant to cross-dataset transfer rather than only within-dataset classification.

EMOD should be understood as an example of a representation-learning strategy, not automatically as a foundation model. Its cross-dataset pretraining and reusable affective embedding illustrate how a foundation-model program could begin: first define representations that transfer across recording conditions and label schemes, then test whether a larger and more diverse pretrained backbone supports many downstream tasks. The model still needs evidence about subject-disjoint pretraining, task and dataset overlap, calibration, robustness, and transfer to affective settings not represented during training.

Validation of Learned Representations

Representation quality should be assessed separately from the accuracy of one attached head. Useful checks include:

• linear probing with the encoder frozen, followed by controlled fine-tuning and parameter-efficient adaptation;
• cross-subject, cross-session, cross-device, and cross-dataset evaluation with the held-out unit matching the claim;
• label-budget curves showing how much performance is available with few labeled trials;
• ablations of the affect-space mapping, contrastive weighting, augmentations, and temporal, spectral, or spatial encoder components;
• tests for shortcut learning, such as predicting subject, dataset, channel montage, or artifact status from the embedding; and
• embedding stability across random seeds, preprocessing choices, and reasonable label perturbations.

Pretraining data and downstream test data must be separated according to the declared protocol. If unlabeled target recordings are used during pretraining, that is transductive or domain-adaptive evaluation and should not be reported as pure unseen-target generalization. These controls connect representation learning directly to the leakage and evidence principles developed elsewhere in this chapter.

Reinforcement Learning

Reinforcement learning learns a policy that selects actions from observations in order to maximize cumulative reward. In affective EEG, the observation may include current EEG features and interaction context; an action may adapt a stimulus, select a training exercise, change feedback, or request a calibration measurement. The objective is not simply to classify emotion accurately at each step, but to choose interventions that improve a longer-term outcome.

A reinforcement-learning formulation should define the state, action space, reward, transition assumptions, safety constraints, and evaluation environment. Offline datasets typically contain logged behavior rather than outcomes for every possible action, so naive off-policy evaluation can be biased. Initial work should favor simulation, conservative offline reinforcement learning, or tightly controlled studies before adaptive policies affect real users.

Example scenario: In a neurofeedback application, a policy observes a participant's causal arousal estimate, recent task performance, and session time. It chooses between lowering task difficulty, maintaining it, or inserting a short break. Reward combines sustained engagement and task performance while penalizing excessive interruptions. Evaluation compares the learned policy with fixed difficulty schedules under the same safety constraints.

Choosing a Paradigm

These paradigms can be combined. For example, an encoder may be pretrained with contrastive learning, fine-tuned with supervised emotion labels, and used within a reinforcement-learning policy. Each stage must still have its own training data, validation criteria, and leakage controls.

References

• Banville, H., Chehab, O., Hyvarinen, 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.
• Chen, Y., Zhao, S., Li, S., and Pan, G. (2026). EMOD: A Unified EEG Emotion Representation Framework Leveraging V-A Guided Contrastive Learning. Proceedings of the AAAI Conference on Artificial Intelligence, 40(21), 17427-17435. https://doi.org/10.1609/aaai.v40i21.38796
• Lotte, F., Bougrain, L., Clerc, M., et al. (2018). A review of classification algorithms for EEG-based brain-computer interfaces: A 10 year update. Journal of Neural Engineering, 15(3), 031005.
• Sutton, R. S., and Barto, A. G. (2018). Reinforcement Learning: An Introduction. MIT Press.