Content-Routed Experts for Few-Trial ERP Estimation

Brain responses to an event, such as reading an unexpected word or seeing a face, unfold as a sequence of voltage bumps on the scalp. These bumps are called event-related potential (ERP) components. Each has a typical timing (a "P1" near 100 ms, an "N400" near 400 ms, a "P300" between 300 and 800 ms), but the real timing drifts from person to person and from trial to trial. Recovering these waveforms normally requires averaging dozens to hundreds of repeated trials. The best learned shortcut, EEG2ERP, recovers the averaged waveform from five trials, yet its authors report that it still places components at the wrong times. We propose a sparse mixture-of-experts model for this problem. A mixture of experts is a network in which a small learned "router" sends each input piece to one of several specialist sub-networks. Here, each short time slice of the recording is sent to an expert based on what the slice looks like, not where it sits in the epoch. If one expert learns "N400-shaped activity," it should catch the N400 whether it arrives at 380 ms or 450 ms. When a component is absent, as when a deceived reader shows no N400, the slice can go to an expert that does nothing.

Getting such a model to train requires solving the main failure mode of mixtures of experts: expert collapse. In collapse, a few experts win most of the traffic early, and the others stop learning. We argue that ERP data makes collapse worse in a specific, diagnosable way. Most of an epoch is background noise. The early sensory responses and eye artifacts are large and produce strong gradients, while the late, subtle components that carry the cognitive signal are small. An unconstrained router therefore collapses onto "large-amplitude" experts. A router forced into uniform balance makes every expert fit noise, so the experts become interchangeable. The proposal is to treat noise as a legitimate routing destination and to balance only the signal experts, measured over whole datasets rather than per batch. We then test whether this produces component-specialized experts that reduce latency errors.

Core Claim

Content-based routing of time slices, combined with a zero-output "null" expert and population-level balancing restricted to signal experts, will do three things on ERP CORE in the 1-to-5-trial regime. It will yield experts whose assignments track canonical ERP components across shifts in latency. It will reduce peak-latency error relative to the monolithic EEG2ERP model. It will avoid both dead experts and interchangeable experts. Standard mixture-of-experts training recipes, which have no null expert and use batch-level or capacity-forced balance, will fail in this setting in a measurable way.

What is new relative to the closest work is that the experts are organized by processing stage, not by electrode location or data source, and that collapse is studied as the central obstacle. BrainStack assigns one expert per anatomical scalp region for silent-speech decoding. It blends experts with a learnable gate that has no load balancing, and it never measures collapse. EEG-MoCE assigns hyperbolic experts to separate modalities such as EEG, audio and video. BrainMoE blends fMRI experts that were each pre-trained on a different cognitive task, and it does so post hoc. fMRI is far too slow in time to separate ERP components. None of these three models routes within a single time-locked epoch.

Research Program

The model starts from the EEG2ERP architecture. Estimating the Event-Related Potential from Few EEG Trials introduced an autoencoder that splits its internal code into a subject part and a task part. Its decoder is conditioned on the number of trials K, and a separate head outputs pointwise uncertainty. We keep its subject/task split, its K-conditioning and its variance head, which together supply the method. We replace the feed-forward blocks of the task encoder with sparse expert layers. The input epoch is cut into short slices of roughly 25–40 ms across all channels. Each slice becomes a token, and each token goes to one expert.

The router follows On the Representation Collapse of Sparse Mixture of Experts. That paper shows that high-dimensional dot-product routers pull token representations toward expert centroids and make assignments flip under small input shifts. It addresses this with a low-dimensional, normalized router with a learned temperature. Flipping under small shifts is exactly what would wreck latency tolerance, so we adopt that router design.

The pool contains E signal experts plus one null expert, which outputs zeros and costs nothing. Balancing follows φ-Balancing for Mixture-of-Experts Training. That paper argues that per-batch balance destroys specialization because mini-batches are unrepresentative of the full data. It instead balances an exponential moving average of historical load through a convex potential. We apply that potential only to the signal experts, so the null expert's share is free. The training objective is

$$\mathcal{L} = \mathcal{L}_{\text{EEG2ERP}} + \lambda\, D_\phi\big(\bar{p}_{1:E} \,\|\, u_E\big),$$

where $\mathcal{L}_{\text{EEG2ERP}}$ is the original reconstruction-plus-contrastive loss, and $\bar{p}_{1:E}$ is the moving-average routing share of the E signal experts, renormalized to exclude the null expert. $u_E$ is the uniform distribution over those experts, $D_\phi$ is the Bregman divergence of negative entropy (a KL divergence), and $\lambda$ is a weight. We also test whether the orthogonality loss on expert outputs from Advancing Expert Specialization for Better MoE adds anything. That paper shows that strict load balancing makes experts redundant and that penalizing collinear expert outputs restores diversity.

The analysis half matters as much as the model. We record routing over training and track three quantities: the fraction of dead experts, the similarity between experts' outputs, and the mutual information between expert assignment and canonical component windows. The component windows are re-centred per subject on that subject's grand-average peak, so that genuine latency shifts count as agreement. This measures directly whether collapse in this setting goes toward high-amplitude early components, as we predict, and whether the null expert absorbs pre-stimulus and between-component slices.

Baselines

The naive reference point is plain averaging of the K available trials. EEG2ERP reports that on ERP CORE at K=5 this gives negative R² values, meaning a flat line beats the sample average. Any learned model must clear this easily. It anchors the scale.

The strongest existing method is EEG2ERP itself. It reports R² = 31.7% at K=5 on ERP CORE (Estimating the Event-Related Potential from Few EEG Trials). We will retrain it from the authors' description under our exact subject splits, preprocessing, epoch length and hyperparameter search budget. We report both our reproduction and the published 31.7%. Beating it means two things: a statistically reliable gain in R² over subject-level resampling at K=1 through 5, and a reduction in absolute peak-latency error for the latency-sensitive components (N2, N400, P3) of at least 10 ms on average. Latency is where the published model admits weakness. An R² gain without a latency gain would not support the claim.

The partition-axis competitor is a spatial-expert model in the style of BrainStack. It uses one expert per scalp region and a learnable fusion gate, and we transplant it into the same EEG2ERP backbone and parameter count. This tests whether experts should be organized by where the signal is or by what stage it represents. BrainStack reports 41.87% accuracy on its silent-speech benchmark against 29.50% for TCNet. No reported result was found on Lacuna for spatial-expert models on ERP CORE.

The routing-recipe competitors keep our architecture and swap the collapse control. Each gets the same tuning budget for its balancing weight. The first is no balancing at all. The second is the Switch-style per-batch auxiliary loss. The third is expert-choice routing, studied for vision in Routers in Vision Mixture of Experts: An Empirical Study, in which experts pick tokens up to a fixed capacity. The fourth is exact balanced assignment in the style of BASE Layers. No reported result was found on Lacuna for any of these on EEG data.

Two ablations isolate the key ingredients. Removing the null expert while keeping population balancing tests our noise-sink hypothesis. Replacing content routing with a fixed position-to-expert map tests whether routing matters at all. For the fixed map, each expert owns a canonical time window: 0–200, 200–400 and 400–800 ms.

Why not just…?

Why not just use fixed time-window experts, since the component windows are known? This is the most direct threat. If components stayed in their textbook windows, a fixed partition would capture all of the benefit with no risk of collapse. Our bet is that between-subject latency spread and conditional absence break fixed windows. Under a fixed map, a late N400 lands in the "P3 expert," and a missing N400 still consumes the N400 expert's capacity. The fixed-map ablation settles this. If it matches content routing on latency error, both on natural data and under the injected-shift test below, the routing claim is false. The remaining contribution would then be the collapse characterization alone, and we would report that.

Why not just add time-shift augmentation, or an explicit latency-alignment step, to the dense EEG2ERP model? This is also a real threat. Shift augmentation teaches invariance cheaply, and alignment methods such as RIDE and dynamic time warping were already among EEG2ERP's comparison baselines. The difference is that augmentation shifts the whole epoch together. Different components within one epoch, however, drift by different amounts: early sensory timing is stable while late evaluative timing varies. Per-slice routing can follow each component separately, whereas global augmentation cannot. We will train EEG2ERP with random global shifts matched to the injected-shift range and treat it as a full baseline, not a strawman. The per-component latency error will show whether separate following is real.

Why not just apply the standard load-balancing loss and move on? The file's evidence suggests it is the wrong default here. Advancing Expert Specialization for Better MoE finds that rigid balancing makes experts redundant. φ-Balancing finds that per-batch statistics destroy domain preferences. In ERP epochs, where most slices are noise, uniform per-batch balance forces every expert to take its share of noise. That evidence comes from language models, though, so this is a prediction rather than a fact for EEG. The routing-recipe comparison, using expert-output similarity and the component mutual-information measure, tests it directly.

Proposed Experiments

The main benchmark is ERP CORE: 40 participants, six paradigms covering N170, mismatch negativity, P3, N400, N2 and error-related negativity, and more than 50,000 trials. We follow EEG2ERP's cross-subject protocol and report R² at K=1 through 5 together with per-component peak latency and amplitude error.

The decisive check is a controlled-jitter experiment. Starting from held-out subjects, we shift the late portion of each trial (after 250 ms) by a known per-subject offset of up to ±60 ms, leaving early components untouched. We then measure how well each model recovers the shifted peak latencies. The proposal succeeds only if content routing tracks the injected offsets better than the fixed-window map and better than shift-augmented EEG2ERP. The test fails if expert assignment does not move with the shifted component.

A second check is conditional engagement. In paradigms with a condition contrast, such as N400 for related versus unrelated words and P3 for rare versus frequent stimuli, the null expert's share in the component window should rise in the condition that lacks the component. This mirrors the all-or-nothing N400/P600 pattern that How do Humans Process AI-generated Hallucination Contents observed in deceived readers.

Collapse diagnostics run across five seeds per routing recipe. They report dead-expert fraction over training, pairwise expert-output cosine similarity, and which component windows the surviving experts cover. These diagnostics test the prediction that unbalanced routing collapses onto early high-amplitude slices.

For transfer, the ERP CORE-trained model is evaluated on Wakeman-Henson face perception (N170) and on the P300 speller. No numerical results are reported on Lacuna for EEG2ERP on either dataset, so we report our reproduced EEG2ERP numbers alongside ours. If the AMBER hallucination EEG recordings are obtainable, a secondary test asks whether late-component expert usage predicts sentence-level hallucination detection. The reference point there is the AUC of 0.931 reported by Zhu et al. for their attention-based EEG model. This test is exploratory and is not part of the core claim.