A brain-computer interface (BCI) must be calibrated for each new user. Someone sits for a few minutes imagining hand or foot movements while their electroencephalogram (EEG, voltage recorded from scalp electrodes) is logged. A decoder then learns to read those signals one trial at a time, with no averaging over repeats. The field has two kinds of decoder, and each fails at a different point in that session. Small, fixed models such as spatial filter banks or compact convolutional networks work from the first dozen trials but stop improving. Large pretrained models with per-user adapters reach higher accuracy, but they overfit when the user has given only a handful of trials. Today practitioners choose one model size before the session starts, although the right size clearly depends on how many trials have been collected so far and how clean this user's signals are.
We propose to make model capacity a quantity that is re-chosen during the calibration session. It would grow from zero user-specific parameters toward a full-rank per-user adapter as trials accumulate. The choice would come from a population-level map of where simple models stop winning, not from cross-validation on the user's own tiny dataset. The mechanism is a pretrained shared EEG decoder with a single "nested" per-user low-rank adapter. The adapter is trained so that any prefix of its rank is a usable model, and it is merged into the weights after training, so on-device inference cost does not depend on the chosen rank. The scientific deliverable is the first measured crossover curve for single-trial EEG decoding: accuracy as a function of calibration trials and adapter capacity. The engineering deliverable is a calibration procedure that tracks the best size on that curve in real time.
Core Claim
In the low-data regime, inner cross-validation over a handful of trials is too noisy to pick model size reliably. A selector that uses the population crossover curve, conditioned on the trial count and one cheap per-user separability statistic, will pick better. We further claim that a nested adapter trained once can stand in for a family of separately trained adapters, so tracking the best capacity costs one training run per update rather than a grid search. With the adapter merged, inference latency and microcontroller memory match the frozen backbone. That removes the usual argument that more capacity means a slower device.
Relation to Prior Work
The motivating evidence comes from three EEG studies that each see one side of the tradeoff. Combining Euclidean Alignment and Data Augmentation for BCI decoding compares ShallowNet, Deep4Net, EEGNet and EEG-Conformer. It finds that the high-capacity models overfit on individual data and gain most from alignment plus augmentation (13.60% for EEG-Conformer), but it uses fixed architectures and never varies trial count systematically. DBConformer shows that accuracy peaks at two Transformer layers and degrades beyond that, and it reaches roughly 76.67% leave-one-subject-out accuracy with about 92k parameters. That is evidence that EEG saturates capacity early, but DBConformer picks its depth once for all users.
The closest method is SuLoRA. It freezes a shared backbone and trains a per-subject low-rank update $W_s = W_{general} + A_s B_s$, where $A_s$ and $B_s$ are thin matrices of rank $r$. With $r=32$ fixed for everyone, it reaches 64.72% on BCI Competition IV 2a, against 55.00% subject-agnostic and 59.36% subject-specific. We reuse its adapter form. We differ in three ways: rank is chosen per user and re-chosen as trials arrive, the adapter is nested so all ranks come from one run, and we evaluate in a simulated streaming calibration session rather than on a fixed split.
The idea that a crossover point exists comes from outside neuroscience. Neural network language models for low resource languages found feed-forward language models beating recurrent ones at about 100k tokens, with the recurrent models winning only past about 500k. State Encoders in Reinforcement Learning for Recommendation found attention losing to MLPs on a 290-user dataset and winning on a 15,400-user one. Both show a crossover exists in their domains. Neither turns it into a selection rule, and neither measures it in the tens-to-hundreds-of-trials regime typical of BCI calibration.
The deployment constraint comes from the EEGNet edge-computing study. It shows that feature-map RAM, not model size, blocks microcontroller deployment, and it reaches 101 ms and 4.28 mJ per inference on a Cortex-M4F by shrinking the input. We deliberately add no runtime structure: merged adapters leave that memory profile unchanged. This is also why we do not pursue early-exit or dynamic-depth inference. The cost we care about is paid during calibration, not at every inference.
Generalizable, real-time neural decoding with hybrid state-space models supplies a useful contrast. Its recurrent streaming backbone works on invasive spikes, where cross-species pretraining makes few-shot transfer strong (97.73% handwriting accuracy against 81.36% for a PCA-KNN baseline). It does not study how capacity should scale when pretraining data is weaker, which is the typical scalp-EEG situation.
Research Program
The first component is the crossover map. For each target user we hold the user out and pretrain a shared backbone on all other users, using Euclidean alignment (re-centering each user's channel covariance to the identity). We use an EEGNet-sized convolutional backbone and a DBConformer-sized dual-branch backbone.
We then define a capacity ladder. Rung zero is the frozen backbone with no user parameters. The next rung refits only the final classifier. The rungs above that add low-rank adapters of rank 1, 2, 4, 8, 16 and 32 on every layer. The top rung is full fine-tuning. Alongside the ladder we run FBCSP, the filter-bank spatial-filter pipeline of the BCI Competition IV FBCSP paper, which needs no pretraining.
For each user we sweep calibration size $N$ from 5 to the full session in trials per class, with many random draws per size, and record held-out accuracy $a(N, r)$ for every rung $r$. We fit, per rung, a saturating curve
$$a(N, r) = \alpha_r - \beta_r N^{-\gamma_r},$$
where $\alpha_r$ is the rung's asymptotic accuracy, $\beta_r$ is its small-data penalty, and $\gamma_r$ is how fast it recovers. We use this power-law form because it is the standard shape of learning curves; the fit residuals will tell us if it is wrong. The crossover between two rungs is the $N$ at which their curves meet.
The second component is the nested adapter. During training we sample a truncation rank $k$ and use only the first $k$ columns of $A_s$ and the first $k$ rows of $B_s$. This forces the leading directions to carry the most useful user-specific correction. One training run then yields every rung from 1 to 32.
The third component is the selector. It predicts the best rung from $N$ and one cheap statistic: the user's FBCSP cross-validated accuracy on the current calibration trials, which serves as a proxy for signal-to-noise. The prediction is read off the fitted population curves, with each curve's parameters shifted linearly in the statistic. The selector is fitted on training users only and evaluated on held-out users.
The fourth component is a streaming calibration simulation. Trials are replayed in recorded order. Every 10 trials the system retrains the nested adapter within a fixed wall-clock budget of a few seconds on a laptop CPU, picks a rank, merges it, and decodes the next block. We report accuracy on each block as the session unfolds.
Benchmarks are BCI Competition IV 2a (9 users, 4-class motor imagery), the 105-user PhysioNet motor imagery set, the MOABB suite (BNCI2014001, Cho2017, Shin2017A) used by Rodrigues et al., and the 64-user Tsinghua RSVP target-detection set. The RSVP set tests whether the crossover moves for evoked responses with rare targets (1–4% of images), where AUC replaces accuracy. The per-user variation in PhysioNet makes it the main set for fitting the selector.
Baselines
The naive reference points are the frozen shared backbone with Euclidean alignment and no user parameters, and FBCSP trained on the calibration trials alone. On 2a, the SuLoRA paper reports 55.00% for its subject-agnostic model, and the FBCSP paper reports a mean kappa of 0.569 on the competition protocol. Any adaptive scheme that falls below the better of these two at small $N$ has failed at its main job.
The strongest existing method is SuLoRA with fixed $r=32$. It reports 64.72% accuracy on BCI Competition IV 2a (SuLoRA). We re-implement it on the same backbone, alignment, calibration draws and tuning budget as our method, and we evaluate it at every $N$, not only the full session. Beating it means two things. First, higher mean accuracy at $N \le 20$ trials per class. Second, no statistically significant loss at the full session, tested with a paired test across users and draws. The adaptive rule must not pay for its small-data gains at large $N$.
On PhysioNet the relevant reference is the optimized EEGNet, reported at 65.07% globally and above 70% with subject-specific transfer (EEGNet edge BCI). On Tsinghua RSVP it is PSAEEGNet at AUC 0.9341 (PSAEEGNet). We run both under our calibration-size sweep.
The oracle rung, chosen per user and per $N$ by held-out test accuracy, is an upper bound. The selector should close most of the gap between fixed $r=32$ and the oracle.
Two ablations isolate the ingredients. For the adapter, the nested version is compared with separately trained adapters at each rank. It should match them within the draw-to-draw noise at a fraction of the training time. For the selector, the population-curve rule is compared with the same ladder chosen by inner cross-validation on the calibration trials. This comparison is the core test of our claim.
Why not just…?
Why not just pick the rank by cross-validation on the user's calibration data? This is the real threat. If it works at small $N$, our selector is unnecessary. With 10 trials per class and four classes, a validation fold holds only a handful of trials, so the accuracy estimate varies by several points between folds. Our hypothesis is that this noise makes inner cross-validation pick overly large rungs by chance. The selector-versus-inner-CV ablation settles the question directly. If inner cross-validation matches the selector at every $N$, we report that result and the contribution reduces to the crossover map and the cheap nested adapter.
Why not just always use the biggest adapter with strong regularization? Rodrigues et al. show that high-capacity models can be rescued by alignment plus augmentation. SuLoRA's single $r=32$ already beats both simpler references at full data. We include fixed $r=32$ with the Rodrigues segmentation-and-reconstruction augmentation as an additional baseline. If it dominates across all $N$, there is no crossover to exploit and we say so.
Why not just use FBCSP or EEGNet and stop worrying about capacity? They are cheap and robust. On 2a, however, the file's evidence shows personalized adapters exceeding FBCSP-level performance at full data. The crossover map will show exactly how many trials the fixed small models stay ahead for. That number is useful to practitioners regardless of which side wins.
Why not just make inference adaptive (early exit, dynamic depth) instead? Merging the adapter makes inference cost independent of capacity. Dynamic inference would add microcontroller RAM pressure, which the EEGNet edge study identifies as the binding constraint, and the file gives no evidence that inference-time capacity is what limits accuracy.
Decisive Checks
Four results decide the project. First, the map must show a crossover. If the rank-32 adapter beats FBCSP and the frozen backbone even at 5 trials per class on all four benchmarks, the premise is wrong. Second, at $N \le 20$ trials per class, the selector must beat inner cross-validation and fixed $r=32$ on held-out PhysioNet users, and the gain must transfer to 2a and MOABB without refitting. Third, nested training must match separately trained adapters, and one retraining step must fit inside the between-block budget on CPU. Fourth, in the streaming simulation, the adaptive decoder's cumulative accuracy over the session must be at least as good as the better of FBCSP and fixed SuLoRA at every block.
A secondary result is whether the crossover $N$ differs between motor imagery and RSVP. A difference would indicate that the capacity schedule depends on the task, not just on data volume.
References
- Combining Euclidean Alignment and Data Augmentation for BCI decoding
- DBConformer: Dual-Branch Convolutional Transformer for EEG Decoding
- Mitigating Subject Dependency in EEG Decoding with Subject-Specific Low-Rank Adapters
- Neural network language models for low resource languages
- State Encoders in Reinforcement Learning for Recommendation: A Reproducibility Study
- An Accurate EEGNet-based Motor-Imagery Brain-Computer Interface for Low-Power Edge Computing
- Generalizable, real-time neural decoding with hybrid state-space models
- Filter Bank Common Spatial Pattern Algorithm on BCI Competition IV Datasets 2a and 2b
- PSAEEGNet: pyramid squeeze attention mechanism-based CNN for single-trial EEG classification in RSVP task