Latent Prediction Raises EEG Transfer Accuracy Across Benchmarks

EEG foundation models are commonly pretrained by masking voltage samples and reconstructing the missing waveform. That objective is convenient, but it can reward recovery of predictable background activity, recording conditions, and artifacts rather than the neural features needed for downstream transfer. EEG-JEPA changes the prediction target: instead of estimating masked signals directly, it asks a model to infer latent representations of masked electrode–time regions from their surrounding context.

The paper evaluates that shift under matched backbone, corpus, and training-duration conditions. Its headline comparison is substantial for frozen transfer: EEG-JEPA reaches 50.42% macro balanced accuracy across 14 EEG-FM-Bench tasks, compared with 40.49% for CBraMod-style masked waveform reconstruction. Continuing pretraining on multiple sources raises the reported average to 52.94%, the best average among the foundation models included in that evaluation.

Core Contribution

The central contribution is a structured latent-prediction objective for EEG, called EEG-JEPA. The authors organize its supervision along three axes: target content, the representation being predicted; target support, the electrode–time locations selected for prediction; and target depth, the encoder layers at which targets are imposed. This is more specific than simply replacing a reconstruction loss. It treats EEG as a spatial-temporal signal whose useful context can span electrode relations, local temporal structure, and increasingly abstract features across encoder depth.

The distinction from waveform masking matters because the target encoder sees the complete crop, while the context encoder receives a masked version. The predictor must therefore use available context to estimate what the full-input encoder represents at selected locations. The authors’ premise is that this makes the pretraining task less dependent on reproducing noisy voltages and more dependent on modeling latent structure that survives transfer.

Technical Approach

Neurotopology-Aware Multi-scale Electrode-Temporal Masking, or N-MET, selects structured targets on the electrode–time lattice rather than relying on unstructured masking. The supplied masking diagram positions this as a mechanism for choosing target regions at multiple scales while respecting the topology of EEG channels and time.

Figure 2 shows the structured target-selection component: target locations are organized across electrodes and time rather than treated as independent missing samples.

The full framework uses a context encoder on the masked crop and an exponential-moving-average target encoder on the unmasked crop. The target branch produces stop-gradient contextual representations. A shared predictor then combines hierarchical context features with learned, position-conditioned target queries to estimate targets from layers L3, L6, L9, and the final layer. Training combines latent prediction with a context-consistency term and variance/covariance regularization. The multi-depth setup supplies target representations from both intermediate and final encoder layers during pretraining.

Figure 3 lays out this asymmetry between masked context processing and complete-crop target generation, as well as the four target depths used during pretraining.

Results and Analysis

The strongest evidence is the controlled frozen-transfer comparison. With the same backbone, pretraining corpus, and training duration, EEG-JEPA improves the 14-task macro balanced accuracy by 9.93 percentage points over waveform reconstruction, from 40.49% to 50.42%. The multi-source continuation result adds 2.52 points, reaching 52.94%. That scale of gain supports the framework's advantage within the reported matched setup.

The full fine-tuning result is smaller: the nine-task average balanced accuracy rises from 68.98% to 70.65%, a 1.67-point improvement. This gap between frozen and fine-tuned gains is informative. EEG-JEPA shows a larger reported gain for frozen representations, while retaining an advantage after full fine-tuning. The result figure separates the 14-task frozen suite from the nine-task fine-tuning suite, so the two averages should not be read as a single directly comparable benchmark.

Figure 4 also includes a parameter-efficiency comparison, but the supplied material does not provide sufficiently legible numerical values to support a quantitative editorial claim about it. The reported transfer outcomes nevertheless support the paper’s main conclusion: latent prediction is a stronger pretraining signal than masked waveform reconstruction under the evaluated protocols, particularly for frozen representations.

Limits in Practice

The evidence is broad across downstream tasks, but it is still bounded by the selected EEG-FM-Bench suites and the paper’s protocol-matched comparisons. The reported frozen and full-fine-tuning results use different task counts, 14 and 9 respectively. The supplied material also does not establish performance on every acquisition setting, clinical deployment condition, or artifact profile. Finally, the paper demonstrates an accuracy advantage, not a direct measurement that the learned features isolate neural structure from non-neural confounds.