EEG2ERP: Uncertainty-Aware ERP Estimation

• EEG2ERP is an uncertainty-aware deep learning framework that maps a few noisy, stimulus-locked EEG trials to subject-specific ERP waveforms.
• It utilizes a split-half autoencoder architecture with trial count embedding and explicit uncertainty modeling to address low-SNR, latency jitter, and inter-subject variability.
• Empirical results on ERP CORE and other datasets demonstrate that EEG2ERP outperforms classical averaging and alignment-based approaches in the few-trial regime.

Searching arXiv for papers on EEG2ERP and closely related ERP estimation/detection methods. arXiv search query: "EEG2ERP event-related potential EEG ERP estimation detection" EEG2ERP denotes, in its narrowest and most explicit usage, an uncertainty-aware autoencoder approach that maps an arbitrary number of electroencephalography (EEG) trials to their associated event-related potential (ERP), with the stated aim of reducing the number of trials necessary for ERP research (Nørskov et al., 28 Nov 2025). In a broader methodological sense, the term also refers to a family of mappings from stimulus-locked EEG to ERP-related objects, including enhanced average waveforms, component detections, latency estimates, and classification-oriented ERP representations. The topic is therefore situated between classical ensemble averaging, alignment-based enhancement, wavelet-based component isolation, and subject-generalizing deep representation learning. Its central technical problem is the same across these variants: ERP structure is time-locked but low-SNR, temporally variable, and strongly modulated by inter-subject differences.

1. Scope, definitions, and problem setting

In conventional ERP analysis, repeated trials are segmented around an event and averaged samplewise to suppress unrelated EEG. EEG2ERP begins from the observation that this procedure becomes unstable in the few-trial regime, is sensitive to outliers, and blurs components when latency jitter is present. The specific EEG2ERP model introduced in 2025 addresses exactly this setting: given only a small number of noisy stimulus-locked EEG trials from a subject and task, estimate the subject-specific ERP that would ordinarily be approximated by averaging many trials (Nørskov et al., 28 Nov 2025).

A stricter definition of EEG2ERP reserves the term for models that output an ERP waveform estimate. Under that definition, some recent ERP-from-EEG systems are adjacent rather than identical. For example, the Deep-Match framework learns from multichannel EEG to detect whether an ERP is present and where in time it occurs, but its output is a temporal detection signal rather than a reconstructed ERP waveform (Zylinski et al., 11 Mar 2026). A plausible implication is that the literature now contains at least two distinct meanings of EEG2ERP: waveform estimation in the narrow sense, and ERP-aware detection or representation learning in the broader sense.

The few-trial regime is the decisive constraint. Standard averaging assumes that trials are equally reliable and temporally aligned, whereas EEG2ERP explicitly targets settings in which only a handful of artifact-free trials are available, such as clinical, pediatric, aging, or brain-computer interface contexts (Nørskov et al., 28 Nov 2025).

2. Classical antecedents: averaging, wavelets, and temporal realignment

Before deep EEG2ERP models, the dominant strategies either improved the classical average or attempted component-specific detection from single trials. One line of work used wavelet-domain similarity measures. In a consumer-grade P300 setting, a 5-channel Emotiv Insight headset was used with 1 s post-stimulus epochs, averaging levels $N \in \{5,10,15,20\}$, and a continuous wavelet transform restricted to scales 1–64 and 230–700 ms; the scalar score was

$S = \sum_{b=1}^{W} \max(C_{a,b}),$

and the reported average classification accuracy with 20 target stimuli was $54.4 \pm 7.7\%$, with a minimum of $33.33\%$ and a maximum of $81.48\%$ (Agapov et al., 2016). The authors’ main conclusion was that subject variability dominated wavelet-family choice.

A more component-specific wavelet approach targeted oscillatory ERPs in single trials. For N170 detection, an asymmetry-based continuous-wavelet method exploited both coefficient magnitude and a characteristic time separation between a negative and a positive coefficient extremum. On 1200 trials, it achieved $96.7\%$ positive-class accuracy, $86.0\%$ negative-class accuracy, and $91.33\%$ overall accuracy, compared with $42.66\%$ for the matching wavelet method and $83.3\%$ for t-CWT (A et al., 2014). This established that wavelet-domain temporal structure could carry component-specific evidence beyond simple template matching.

A second lineage addressed temporal misalignment directly. A modified dynamic time warping procedure first computes a conventional average, then aligns each trial to that template, reconstructs an averaging-compatible warped signal, low-pass filters it, and averages the aligned trials. On public ERP datasets, the filtered DTW-based average increased median and mean P200 peak by $S = \sum_{b=1}^{W} \max(C_{a,b}),$0 and $S = \sum_{b=1}^{W} \max(C_{a,b}),$1, and median and mean P200 amplitude by $S = \sum_{b=1}^{W} \max(C_{a,b}),$2 and $S = \sum_{b=1}^{W} \max(C_{a,b}),$3, relative to conventional averaging (Molina et al., 2024). This suggests that an important prehistory of EEG2ERP is not only denoising, but explicit correction of latency and jitter before averaging.

3. EEG2ERP as an uncertainty-aware deep waveform estimator

The model explicitly named EEG2ERP is built from CSLP-AE, a contrastive split-latent permutation autoencoder, but is repurposed from single-trial self-reconstruction to ERP denoising (Nørskov et al., 28 Nov 2025). Instead of reconstructing each input trial, it learns to map an average of a small number of trials to a higher-quality ERP target. Its encoder and decoder are defined as

$S = \sum_{b=1}^{W} \max(C_{a,b}),$4

with a split bottleneck containing a subject-specific latent and a task-specific latent. The encoder acts as an implicit conditioner, so explicit subject or task labels are not required at inference.

Support for an arbitrary number of input trials is implemented by averaging however many trials are available and conditioning the network on the count $S = \sum_{b=1}^{W} \max(C_{a,b}),$5. During training,

$S = \sum_{b=1}^{W} \max(C_{a,b}),$6

so small trial counts are sampled more often (Nørskov et al., 28 Nov 2025). This inverse weighting biases learning toward the few-trial regime. The count is embedded with a sinusoidal positional embedding and projected into the network before the bottleneck transformer.

A defining design choice is split-half training. For each subject-task pair, trials are divided into disjoint input and target halves. Inputs are formed only from the input half, whereas targets are bootstrapped from the target half. The paper argues that using the full-set average as target would leak shared single-trial noise into both predictor and target, artificially inflating correlation and understating the effect of latency jitter (Nørskov et al., 28 Nov 2025).

Uncertainty modeling is explicit. The network predicts both a mean ERP and a pointwise standard deviation through a separate variance decoder with softplus positivity. Training uses bootstrapped target ERPs under a factorized Gaussian likelihood, and the total loss combines reconstruction, contrastive, and latent-permutation terms. The appendix also gives an inverse-variance aggregation rule for a single-trial variant: $S = \sum_{b=1}^{W} \max(C_{a,b}),$7 with corresponding variance

$S = \sum_{b=1}^{W} \max(C_{a,b}),$8

Architecturally, EEG2ERP replaces ReLUs with gated linear units and replaces stride-based residual resizing with linear interpolation in the residual branch, while moving instance normalization into the GLU gate to preserve amplitude information important for ERP morphology (Nørskov et al., 28 Nov 2025).

4. Datasets, preprocessing, and evaluation protocol

EEG2ERP was evaluated in a zero-shot cross-subject setting on three public data sources: ERP CORE, the P300 Speller dataset, and the Wakeman–Henson face-perception dataset containing both EEG and magnetoencephalography (MEG) (Nørskov et al., 28 Nov 2025). ERP CORE includes 40 subjects, 30 EEG channels, and over 50,000 trials across six paradigms and seven named ERP components; the split was 28 train, 4 validation, and 8 test subjects. The Wakeman–Henson dataset includes 70 EEG channels and 102 MEG channels sampled at 200 Hz after preprocessing, with subject split 11 train, 2 validation, and 3 test. The P300 Speller dataset includes 55 subjects and 64 EEG channels, split 38 train, 5 validation, and 12 test.

Preprocessing for the M/EEG datasets was summarized as bad-channel detection and interpolation, 1 Hz high-pass, 50 Hz notch, epoching, artifact rejection for jump artifacts and EOG-based artifacts, baseline correction over $S = \sum_{b=1}^{W} \max(C_{a,b}),$9 to 0 ms, average re-reference for EEG, resampling to 200 Hz, and cropping to $54.4 \pm 7.7\%$0 to 800 ms (Nørskov et al., 28 Nov 2025). Formal evaluation compares the predicted ERP to the ERP from the held-out target half, using RMSE and $54.4 \pm 7.7\%$1 computed over $54.4 \pm 7.7\%$2 bootstrapped input samples.

The baseline suite was broad: simple averaging, the tanh weighting estimator of Leonowicz et al., DTW averaging of Molina et al., Woody’s latency alignment, RIDE, a global grand-average template, nearest-neighbor template selection, and xDAWN on the P300 Speller task (Nørskov et al., 28 Nov 2025). This is important because EEG2ERP was not evaluated only against naive averaging, but against both robust averaging and alignment-based baselines.

5. Empirical findings, ablations, and operating regime

EEG2ERP’s strongest results occur in the few-trial regime. On ERP CORE, $54.4 \pm 7.7\%$3 was $54.4 \pm 7.7\%$4 at $54.4 \pm 7.7\%$5, $54.4 \pm 7.7\%$6 at $54.4 \pm 7.7\%$7, $54.4 \pm 7.7\%$8 at $54.4 \pm 7.7\%$9, and $33.33\%$0 at $33.33\%$1 (Nørskov et al., 28 Nov 2025). By contrast, simple averaging gave $33.33\%$2, $33.33\%$3, $33.33\%$4, and $33.33\%$5, while DTW gave $33.33\%$6, $33.33\%$7, and $33.33\%$8 at $33.33\%$9, $81.48\%$0, and $81.48\%$1. The paper notes a particularly striking result: EEG2ERP at $81.48\%$2 can exceed simple averaging even when the latter uses $81.48\%$3 of the input-half trials.

On Wakeman–Henson EEG, EEG2ERP again led at small $81.48\%$4, with $81.48\%$5 at $81.48\%$6, $81.48\%$7 at $81.48\%$8, $81.48\%$9 at $96.7\%$0, and $96.7\%$1 at $96.7\%$2 (Nørskov et al., 28 Nov 2025). On MEG, the same qualitative pattern held at low trial counts, but at $96.7\%$3 simple and weighted averaging were much higher than EEG2ERP. On the P300 Speller target-character ERP task, EEG2ERP reached $96.7\%$4 at $96.7\%$5, $96.7\%$6 at $96.7\%$7, and $96.7\%$8 at $96.7\%$9, slightly exceeding xDAWN at $86.0\%$0 for target ERPs. For non-target ERPs, all methods struggled, and xDAWN was best at $86.0\%$1 with $86.0\%$2 (Nørskov et al., 28 Nov 2025).

These results define EEG2ERP’s operating regime precisely. It is not primarily a replacement for classical averages when many trials are available; rather, it is strongest when only one or a few trials can be used. The paper states that performance varies across paradigms, with MMN and N2pc remaining difficult, and that latency recovery is less improved than amplitude recovery because the input itself is already trial-averaged, so some latency variability has been smeared before the network sees it (Nørskov et al., 28 Nov 2025).

Ablations support the importance of the model’s defining components. On ERP CORE, removing uncertainty modeling reduced $86.0\%$3 to $86.0\%$4, $86.0\%$5, and $86.0\%$6 at $86.0\%$7, $86.0\%$8, and $86.0\%$9, while removing both uncertainty and trial-count embedding yielded $91.33\%$0, $91.33\%$1, and $91.33\%$2 (Nørskov et al., 28 Nov 2025). The predicted standard deviation also tracked actual RMSE qualitatively, which the authors interpret as reasonable calibration.

6. Adjacent paradigms, misconceptions, and unresolved questions

A common misconception is that all recent EEG2ERP-adjacent systems are waveform estimators. Several are not. “ERP-LSTM” uses a classical ERP extraction stage—72 trials per image partitioned into 6 sets of 12, averaged into ERP sequences—and then trains an LSTM for visual classification; it is therefore an ERP-based EEG decoding pipeline rather than a direct EEG-to-ERP model (Zheng et al., 2020). Deep-Match is an ERP-from-EEG detection/localization method, and ERP-XTTN maps single-trial EEG into an interpretable ERP-prototype routing map for cross-subject classification, but neither reconstructs a continuous ERP waveform (Zylinski et al., 11 Mar 2026, Wyman et al., 1 Jun 2026). A broader misunderstanding is to equate ERP benchmarking with ERP estimation: the 2026 ERP benchmark compares manual features, deep models, and foundation models for single-trial ERP classification and disease detection, and concludes that supervised deep learning usually outperforms handcrafted features, while current EEG foundation models do not demonstrate clear performance advantages over supervised models trained from scratch (Wang et al., 2 Jan 2026).

Another unresolved issue concerns the relation between estimation and inference. EEG2ERP estimates waveforms from few trials, whereas SLAM is a semiparametric latent ANOVA model that unifies ERP component identification and covariate association at the waveform-analysis stage, with explicit latent stationary points and group-level inference on latency (Yu et al., 2023). This suggests that future EEG2ERP pipelines may need tighter integration between waveform recovery and downstream statistical modeling.

The open problems are explicit. EEG2ERP remains strongest in controlled laboratory datasets and can be surpassed by DTW or by simple or weighted averaging when many trials are available (Nørskov et al., 28 Nov 2025). The paper itself points toward combining alignment-based preprocessing with EEG2ERP, modeling explicit latency structure, exploring alternative $91.33\%$3 sampling schedules, and building on newer self-supervised EEG foundation models such as BENDR, Neuro-GPT, and LaBraM (Nørskov et al., 28 Nov 2025). In that sense, EEG2ERP marks a transition rather than an endpoint: from ERP estimation as fixed averaging to ERP estimation as a subject-generalizing, uncertainty-aware, data-efficient inference problem.