Deep learning has revolutionized computer vision and natural language processing, fields where models are routinely trained on millions of images or billions of tokens. However, applying these data-hungry architectures to non-invasive neural signals like electroencephalography (EEG) introduces a fundamental bottleneck: data scarcity.
Unlike text or images, high-quality brainwaves cannot be scraped from the internet. Collecting EEG data requires specialized hardware, intensive clinical or laboratory setups, and significant participant time. Furthermore, human subjects experience rapid physical and mental fatigue, strictly limiting the duration of a single recording session (Xie et al.). This constraints the field to small, fragmented datasets with high dimensionality and low signal-to-noise ratios, making deep neural networks highly prone to overfitting.
To build practical, "plug-and-play" Brain-Computer Interfaces (BCIs) without forcing users through exhausting calibration sessions, researchers have developed a multi-faceted machine learning paradigm designed specifically to overcome neural data scarcity.
Cross-Subject and Cross-Domain Transfer Learning
If we cannot collect thousands of samples from a single user, the next logical step is to "borrow" data from other individuals. However, brain signals exhibit massive inter-subject variability; a network trained on one person's motor imagery often fails when deployed on another.
To bypass this calibration burden, researchers use alignment techniques to map disparate brain distributions into a shared mathematical space. For instance, Rodrigues et al. utilize Euclidean Alignment (EA), a method that aligns the covariance matrices of different subjects' EEG signals before feeding them into deep neural networks. By centering and scaling the data relative to a reference state, the network can focus on invariant physiological features rather than subject-specific noise. Similarly, Jin et al. introduced a hybrid framework called KMM-TrAdaBoost. This approach uses Kernel Mean Matching (KMM) to minimize the distribution distance between a source subject and a target subject, combined with an instance-based boosting algorithm to intelligently select and weight only the most relevant historical training samples.
[Subject A EEG] ───> [ Covariance Alignment ] ───┐
├───> [ Shared Feature Space ] ───> Robust Decoder
[Subject B EEG] ───> [ Covariance Alignment ] ───┘
Taking transfer learning a step further, some frameworks bridge entirely different domains or modalities. Because computer vision models benefit from massive pre-existing datasets, Tan et al. converted raw EEG signals into a video-like format termed "EEG optical flow." By framing brain activity as a dynamic visual sequence, they successfully transferred representations from computer vision architectures to decode brain states without requiring expert medical labeling. In clinical settings where specific patient data is exceptionally rare, Jin et al. demonstrated that cross-modality transfer models can even map knowledge from healthy individuals' EEG data to stroke patients' functional near-infrared spectroscopy (fNIRS) signals, drastically reducing the data required for robotic rehabilitation.
Physics-Informed and Generative Data Augmentation
When real data is exhausted, synthetic data can fill the gaps. However, standard data augmentation techniques used in computer vision (such as rotating or cropping an image) destroy the delicate temporal frequencies and spatial layouts inherent to neural recordings.
To generate realistic synthetic brainwaves, researchers employ generative models constrained by physiological boundaries. Bao et al. developed VAE-D2GAN, a hybrid architecture combining a Variational Autoencoder with a Dual Discriminator Generative Adversarial Network. The VAE stabilizes the training process by learning a smooth latent space of brain features, while the dual discriminators force the generated samples to match both the high-frequency details and the global statistical distributions of real EEG data. To address class imbalances in emotional datasets, Zhang et al. introduced ESC-GAN, which explicitly edits majority-class brain signals into under-represented emotional subspaces, pushing synthetic samples toward the decision boundaries where classifiers struggle most.
Beyond deep generative models, geometric and biological properties of the brain can guide augmentation. Kalaganis et al. proposed using Graph Empirical Mode Decomposition (GEMD). Because the physical layout of EEG electrodes forms a non-Euclidean space over the scalp, they modeled the sensor configuration as a graph. GEMD decomposes the multi-channel signals into intrinsic spatiotemporal modes based on the graph's topology, allowing the synthesis of new, physically plausible training trials that preserve the true underlying functional connectivity of the brain.
Architectural Regularization and Inductive Biases
Another way to combat data scarcity is to restrict the model's search space. By embedding domain-specific assumptions—inductive biases—directly into the model architecture, we prevent complex networks from memorizing noise when training samples are limited.
A classic example is sensor selection. Raw EEG setups often record from dozens of channels, many of which are irrelevant to a specific cognitive task. Rakotomamonjy et al. addressed this by integrating sensor weighting directly into the optimization objective of Support Vector Machines (sw-SVM). By learning hyperparameter weights that penalize uninformative channels, the model mathematically prunes noisy dimensions, allowing it to generalize effectively from sparse datasets.
In deep learning, this principle manifests as structural sparsity. Petrich et al. demonstrated that applying Sparse Evolutionary Training (SET) to BCI networks—dynamically dropping and regrowing weights during training based on their magnitude—outperforms dense networks when per-participant data is highly constrained. The sparse connectivity acts as a powerful regularizer, preventing the network from forming the complex, spurious co-adaptations that typically cause overfitting on small sample sizes.
Dense Network (Prone to Overfitting) Sparse Network via SET (Robust to Noise)
[O] [O] [O] [O] [O] [O]
/ | \ / | \ / | \ | / \
[O] [O] [O] [O] [O] [O]
\ | / \ | / \ | / \ /
[O] [O] [O] [O] [O] [O]
Foundation Models for Neural Signals
The most recent shift in overcoming data scarcity mirrors the trajectory of modern natural language processing: moving away from training small, task-specific models from scratch, and moving toward massive, self-supervised pre-training.
By aggregating thousands of hours of unlabelled, heterogeneous EEG data across diverse clinical and experimental paradigms, researchers can train Foundation Models that learn a generalized "grammar" of brain activity. Once this baseline representation is established, the model can be fine-tuned on a specific downstream task—such as seizure detection or motor imagery decoding—using only a fraction of the labeled data previously required.
- LaBraM (Large Brain Model): Developed by Jiang et al., this architecture scales to a 369-million-parameter Transformer pre-trained on 2,500 hours of EEG data. It utilizes a specialized neural tokenizer to convert continuous, multi-channel timeseries into discrete tokens, allowing the model to learn robust, generalized representations of neural dynamics.
- EpilepsyFM: Focusing on clinical utility, Li et al. introduced a foundation model that bridges the gap between non-invasive scalp EEG and highly scarce, invasive intracranial stereoelectroencephalography (SEEG) data. By employing a channel-set masking strategy during self-supervised pre-training, the model learns to infer deep, localized brain dynamics even when downstream clinical data is severely limited.
Through this combination of cross-subject alignment, physics-informed generation, structural regularization, and large-scale self-supervised pre-training, the field is steadily moving past the constraints of data scarcity, unlocking practical and accessible neural decoders.
Key Foundational Work
sw-SVM: sensor weighting support vector machines for EEG-based brain–computer interfaces (2011) — contributed 1 concept elements to this direction.
Compressed Sensing, Sparsity, and Dimensionality in Neuronal Information Processing and Data Analysis (2012) — contributed 1 concept elements to this direction.