Every brain-computer interface (a system that turns recorded brain activity into commands) has to be recalibrated for each new user. Skull thickness, cortical folding and electrode placement change how brain activity reaches the scalp electrodes, and scalp EEG (electroencephalography, voltage recorded from the scalp) is so noisy that a decoder trained on other people often fails on a new person. Wireless engineers face the same problem with radio channels. They send a few known "pilot" symbols so the receiver can learn the channel. A recent line of work skips the explicit channel estimate altogether: a single network reads the raw pilots and outputs the final decision. With far fewer pilots, it matches or beats the classical estimate-then-decode chain.
We propose to build that kind of receiver for EEG. A handful of labelled calibration trials from a new user plays the role of the pilot symbols. One network takes these trials and the trial to be decoded as joint input and outputs the decision directly. It never runs per-user gradient updates and never explicitly estimates the user's spatial filters. The question is empirical. When calibration trials are few and noisy, does conditioning on them inside one network beat the two accepted routes? The first route estimates per-user covariance or spatial filters and then classifies. The second fine-tunes a compact convolutional decoder on the new user's trials.
Core Claim
The claim is that per-user EEG calibration is a channel-estimation problem in disguise. In that setting, the lesson from monolithic receivers should carry over: the intermediate estimate is where the error enters. Learning to Beamform for Intelligent Reflecting Surface with Implicit Channel Estimate maps raw pilots straight to beamforming outputs and trains on the final sum-rate. It gets within 6.7% of the perfect-channel-knowledge bound with 54 pilots, where the classical least-squares-plus-optimisation pipeline needs about 180. Deep Learning for Channel Sensing and Hybrid Precoding explains why. A channel estimate tuned to minimise squared error is not the estimate that maximises the downstream objective.
In EEG, the analogous intermediate is the per-user spatial covariance or common-spatial-pattern filter. Common spatial patterns (CSP) is the classical method that finds electrode weightings separating two classes by their signal power. The standard decoding pipeline, the filter-bank version FBCSP from Filter Bank Common Spatial Pattern Algorithm on BCI Competition IV Datasets 2a and 2b, computes this filter from the user's own trials. Estimating a covariance matrix from a handful of low-SNR trials is exactly the regime where plug-in estimates fail.
Formally, the modular route computes $\hat y = g(x_q; \hat W)$. Here $x_q$ is the query trial, $\hat W$ is a user-specific filter estimated from the $K$ calibration trials $\{(x_i, y_i)\}_{i=1}^{K}$, and $g$ is a classifier. Our receiver instead computes
$$\hat y = f_\theta\big(x_q;\ \{(x_i, y_i)\}_{i=1}^{K}\big),$$
where $f_\theta$ is a single network trained across many simulated "new user" episodes. Its only loss is decoding accuracy on the query trial.
The prediction is specific. The in-context receiver should win most clearly at small $K$, roughly 2 to 10 trials per class. The advantage should shrink as $K$ grows and covariance estimates become reliable.
This differs from the closest wireless work in two ways. Learning How to Demodulate from Few Pilots via Meta-Learning uses MAML (meta-learning an initialisation that adapts with one gradient step) on synthetic amplifier distortions, not on real biological signals. The implicit-beamforming network handles only synthetic fading, with a plain multilayer perceptron. It also differs from the closest neural work. EEGNet and LaBraM are end-to-end from raw EEG to label, but they take no calibration trials as input. New users are handled by retraining or fine-tuning. So "monolithic EEG decoding" by itself is not new. What is new is treating the calibration block as a network input.
Research Program
The receiver has three parts.
The first is a shared front end that extracts features from each trial automatically. It follows EEGNet's design: temporal convolutions act as learned band-pass filters, and depthwise convolutions act as learned spatial filters. It is optionally initialised from the public LaBraM encoder (LaBraM code). We keep the compact EEGNet-style front end as the default because the data are small. Residual and Plain CNNs for 3D Brain MRI found that extra depth bought nothing on a small cohort (AUC 0.88 plain versus 0.87 residual), and EEGNet reaches competitive accuracy with about 1,100 parameters.
The second is a cross-attention block. The query trial's features attend over the features of the calibration trials, each paired with its label embedding. This produces a user-conditioned representation.
The third is a small head that outputs the class, or, for visual decoding, a vector in CLIP embedding space.
Training is episodic. Each episode holds out one subject (or one recording session), draws $K$ labelled calibration trials from it, and scores the network on further trials from the same subject. Low SNR enters through $K$: we vary it from 1 to the full session, so that noise in the calibration trials is the controlled variable.
We also add an auxiliary objective borrowed from EEG2Rep, which predicts masked contiguous blocks in latent space rather than raw samples. It is included only if it helps. EEG2Rep showed that reconstructing raw samples wastes capacity on artifacts. Our pure task loss avoids reconstruction, but a latent-prediction loss may regularise the front end when few subjects are available.
A mechanistic analysis follows the main experiments. We probe whether the cross-attention output linearly encodes the held-out user's covariance structure. We also check whether saliency concentrates on contralateral motor cortex, as EEGNet's DeepLIFT analysis found for its decoder. This addresses the opacity concern raised in Deep Learning for Wireless Communications.
Baselines
The do-nothing reference is a pooled decoder with no calibration. It is an EEGNet trained on all training subjects and applied zero-shot to the held-out subject. This measures the subject gap that calibration must close. Visual Decoding and Reconstruction via EEG Embeddings with Guided Diffusion reports that the subject gap is severe but gives no cross-subject number in the file. So there is no reported result found on Lacuna for cross-subject zero-shot accuracy on these benchmarks, and we will measure it ourselves.
The classical modular competitor is FBCSP with a naive Bayes Parzen-window classifier, estimated only from the $K$ calibration trials. We add a stronger modular variant that shrinks toward the pooled covariance, so that the comparison is not against a strawman plug-in estimator. With full per-subject training data, FBCSP reports a mean kappa of 0.569 on BCI Competition IV 2a and 0.600 on 2b (Ang et al.). We use these as the full-calibration ceiling for the classical route.
The strongest deep competitor is an EEGNet pretrained on the training subjects and then fine-tuned on the $K$ calibration trials. We tune the learning rate, step count and choice of frozen layers per $K$ on validation subjects, with the same search budget as our receiver's hyperparameters. We also include a MAML-initialised EEGNet in the style of Park et al., which is the direct wireless analogue. For the unlabelled-adaptation competitor we use TSMNet, which adapts to a new user through domain-specific batch normalisation on the covariance matrices of that user's unlabelled trials. We give it the same $K$ trials without labels, plus any unlabelled trials available.
Two ablations isolate the core ingredient. In the first, the calibration trials are replaced by an explicitly computed user covariance matrix. This is fed to the same head at the same parameter count and trained end-to-end. It tests "implicit versus explicit channel" while holding capacity fixed. In the second, the calibration labels are shuffled, which tests whether the receiver uses label information or merely normalises statistics.
All methods see identical subjects, splits, trial counts and preprocessing. Beating the field means a higher mean kappa (for BCI IV) or top-1 retrieval (for THINGS-EEG2) than both fine-tuned EEGNet and TSMNet at $K \le 10$ per class, with the difference holding across held-out subjects under a paired test.
Why not just…?
Why not just fine-tune EEGNet on the calibration trials? This is the most serious threat. Fine-tuning ~1,100 parameters on 10 trials may already be near-optimal, and nothing in the file shows it failing. The only indirect evidence is Park et al.'s finding that jointly trained initialisations adapt poorly from few pilots, and that comes from synthetic radio data. Our $K$-sweep settles the question directly. If fine-tuned EEGNet matches the receiver at every $K$, the claim is false, and we will report it as such. The residual advantage of the receiver would then only be that it adapts without any gradient steps on the device.
Why not just use FBCSP estimated from the calibration trials? At full calibration it is a strong, interpretable baseline. But its covariance estimates come from the very trials whose scarcity defines our regime, and EEGNet already matches or beats it across four paradigms in Lawhern et al.. We include a shrinkage version so that any receiver advantage cannot be blamed on a poor estimator. The explicit-covariance ablation then separates "better estimator" from "no estimator."
Why not just align unlabelled statistics, as TSMNet does, and skip labels altogether? If most of the subject gap is a shift in second-order statistics, unlabelled re-centering may capture it for free. That would make labelled pilots unnecessary. The label-shuffling ablation and the TSMNet baseline together decide this. If shuffled-label pilots perform as well as true ones, the receiver is only doing normalisation, and TSMNet is the more economical answer.
Proposed Experiments
The primary testbed is BCI Competition IV 2a (22 channels, 4 classes, 9 subjects) and 2b (3 bipolar channels, 2 classes), evaluated leave-one-subject-out. Episodes are drawn across both subjects and sessions, so that the two recording sessions supply extra shift.
The main figure is kappa against $K$, the number of calibration trials per class, for all baselines. The decisive check is whether the receiver's gap over fine-tuned EEGNet and TSMNet is largest at small $K$ and closes as $K$ grows, as the estimation-error argument predicts. A flat or reversed gap would refute the mechanism even if average accuracy were higher.
A second test is THINGS-EEG2 visual decoding, where calibration pilots are a few labelled image-EEG pairs from the held-out subject. The target is CLIP-space retrieval among 200 candidates. The subject-dependent reference is ATM's reported 28.6% top-1 (Li et al.). For cross-subject retrieval with few pilots there is no reported result found on Lacuna for THINGS-EEG2, so the fraction of the subject-dependent score recovered at each $K$ is itself the result.
Because nine or ten training subjects is thin for episodic training, we repeat the BCI experiments with the front end initialised from LaBraM. We report whether subject diversity from pretraining, rather than the in-context mechanism, explains any gain. A final robustness check corrupts calibration trials with added artifacts. It asks whether the receiver learns to down-weight bad pilots, the way the impulsive-noise receiver in Bin Wang et al. learned to ignore amplitude spikes.
References
- Learning to Reflect and to Beamform for Intelligent Reflecting Surface With Implicit Channel Estimation
- Deep Learning for Channel Sensing and Hybrid Precoding in TDD Massive MIMO OFDM Systems
- Filter Bank Common Spatial Pattern Algorithm on BCI Competition IV Datasets 2a and 2b
- Learning How to Demodulate from Few Pilots via Meta-Learning
- EEGNet: A Compact Convolutional Network for EEG-based Brain-Computer Interfaces
- Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI
- Residual and plain convolutional neural networks for 3D brain MRI classification
- EEG2Rep: Enhancing Self-supervised EEG Representation Through Informative Masked Inputs
- Deep Learning for Wireless Communications: An Emerging Interdisciplinary Paradigm
- Visual Decoding and Reconstruction via EEG Embeddings with Guided Diffusion
- An intelligent signal processing method against impulsive noise interference in AIoT