A linear probe is a single linear classifier trained on top of a frozen pre-trained network. It is the standard way to ask what a foundation model has learned. For brain recordings such as scalp EEG (electroencephalography, voltages measured through the skull), probe accuracy is weak evidence. The signal is tiny compared with muscle, eye and electrical interference. Some of that interference is correlated with the label: people move their eyes during emotional film clips, and they tense muscles during imagined movements. A probe can therefore score well by reading artifacts or quirks of one subject rather than brain activity. We propose to separate these cases using a fact from causal identifiability research. When many subjects receive the same synchronized stimulus, the covariance computed between two different subjects' embeddings contains none of either subject's private noise, however large that noise is. This holds whether the noise is Gaussian or not. The cross-subject covariance therefore identifies a stimulus-driven subspace of the embedding. Any linear probe can then be split into a part that reads this shared subspace and a part that reads subject-private directions.
The deliverable is a measurement procedure plus one testable prediction. The procedure scores how much of a probe's decoding comes from the identified shared subspace, and checks the answer by deleting each subspace in turn. The prediction is that probes restricted to the shared subspace transfer to unseen subjects better than ordinary probes, even when they score lower within a subject. If that holds, the procedure explains part of the "linear probing fails on EEG" result in the benchmarks. It also gives a principled way to compare pre-training objectives, including the unsettled question of whether reconstructing masked raw signal or predicting masked latents produces more usable representations.
Core Claim
Two benchmarks report that linear probes on frozen EEG foundation models perform poorly: Are EEG Foundation Models Worth It? and EEG-FM-Bench. They measure this only through accuracy, and they cannot say whether the label-relevant information that is linearly present is neural, stimulus-locked, or subject-specific noise. We claim the decomposition is identifiable from frozen embeddings alone, given synchronized multi-subject data.
Model the frozen embedding of subject $s$ at stimulus time $t$ as
$$z_s(t) = A_s\, c(t) + n_s(t),$$
where $c(t)$ is a low-dimensional stimulus-driven component shared by all subjects, $A_s$ is a subject-specific linear map, and $n_s(t)$ is everything private to the subject: artifacts, background activity, and subject idiosyncrasies. If private terms are independent across subjects, then for $s \neq t$
$$\mathrm{Cov}(z_s, z_t) = A_s\, \Sigma_c\, A_t^{\top},$$
where $\Sigma_c$ is the covariance of $c$. The column space of this matrix is the shared subspace. It is untouched by the noise-to-signal ratio, because the noise terms do not appear. Only the estimation variance grows with noise, roughly as noise variance over the number of synchronized samples, since the cross term is an average of products in which the noise has zero mean.
This is the second-order mechanism that Multi-View Causal Discovery without Non-Gaussianity uses to identify causal ordering across subjects in sensor space. The measurement-error literature makes the contrast sharp. CAMME shows that within-view correlations and partial correlations are biased by the error-variance ratio. TIN shows that suitably transformed observations can still separate latent causes at noise four times the signal. We move this reasoning out of sensor space and into the latent space of pre-trained encoders, where nobody has tested it. We then use it to audit probes rather than to discover graphs.
Research Program
The work has four parts.
Shared-subspace estimator. On a stimulus-synchronized dataset, we extract frozen embeddings for every subject and form all pairwise cross-subject covariance matrices. We recover the shared subspace per subject with a generalized multi-set canonical correlation step, keeping $k$ dimensions. We choose $k$ by permutation: shuffling the time alignment between subjects destroys true sharing, so it gives a null distribution for each candidate dimension. The estimator itself is a standard multi-set canonical correlation method, and we do not claim it as new. What we test is whether it is valid on foundation-model embeddings. That would fail if pre-training mixed private noise nonlinearly into the shared directions. In that case the independence argument would hold in sensor space but break after the encoder.
Probe audit. Each standard linear probe's weight vector is projected onto the shared and private subspaces. We report the fraction of decoding performance carried by each. We then run deletion tests in the spirit of Measuring the Reliability of Causal Probing Methods, which scores interventions on text-model representations by two quantities. Completeness asks whether removing a subspace removes the property. Selectivity asks whether it leaves everything else intact. That paper found linear deletions unreliable in intermediate layers of language models, which is a warning for us. We therefore report both quantities rather than one, and we test whether the trade-off it found also appears in noisy neural embeddings.
Shared-subspace probe. A linear probe is trained only on the identified $k$-dimensional coordinates. It remains strictly linear: the subspace is fixed before any label is seen, so the readout is still a hyperplane on frozen features. This avoids the hidden nonlinear pooling heads that Are EEG Foundation Models Worth It? found inflating earlier "linear probing" claims.
Pre-training comparison. The audit is applied across encoders trained with different objectives. ST-EEGFormer reconstructs masked raw signal. EEG2Rep predicts masked regions in latent space, motivated by the argument that raw reconstruction wastes capacity on noise. LaBraM and EEGPT are evaluated as they appear in the benchmarks. EEG-FM-Bench found reconstruction gradients nearly orthogonal to label gradients. Our audit asks the representational version of that question: which objective places more label-relevant variance in the cross-subject shared subspace?
Baselines
Naive reference. The standard frozen-backbone linear probe, trained per subject and evaluated within subject and leave-one-subject-out. This is the protocol both EEG benchmarks used. Its within-subject accuracy is the number our method must explain, and its leave-one-subject-out accuracy is the number we expect to beat. We give it the same regularization grid, the same epochs and the same embeddings as the shared-subspace probe. We will not use any reported benchmark figure for this probe, because no per-dataset linear-probe number for SEED or DREAMER was found on Lacuna. We rerun it ourselves.
Strongest existing method. Full or low-rank fine-tuning of the same backbones, which EEG-FM-Bench shows consistently beats frozen probing. It is the strongest competitor on raw accuracy. However, it gives up the diagnostic purpose, and the cross-subject benchmark reports it erodes population robustness ("LOO Drop"). We run it with matched tuning budgets and report both accuracy and drop. No reported numeric result was found on Lacuna for fine-tuned models on SEED or DREAMER. We do not expect to beat fine-tuning within subject. Beating it means matching or exceeding its leave-one-subject-out accuracy with no gradient updates to the encoder.
Classical anchor. FBCSP, a filter-bank common spatial pattern pipeline, reports a mean kappa of 0.569 on BCI Competition IV 2a and 0.600 on 2b. We include it on 2a as a sanity anchor showing where non-neural decoders stand. That dataset lacks stimulus-synchronized content, so it serves only as an anchor and not as a test of the shared-subspace method.
Ablations. Two ablations isolate the key ingredient, which is the use of cross-subject covariance. The first replaces the shared subspace with a same-rank subspace from pooled within-subject principal components. This keeps the dimensionality reduction but loses the noise-cancelling cross term. The second uses a same-rank random subspace. If either matches the shared-subspace probe on leave-one-subject-out accuracy and on the artifact-detection test below, the identifiability argument is doing no work.
Why not just…?
Why not just run multi-set CCA or hyperalignment and call it done? This is the real threat to novelty, because our estimator is a multi-set canonical correlation method. We do not claim the estimator. The contribution lies elsewhere. First, we test whether its noise-immunity guarantee survives a nonlinear pre-trained encoder, using the semi-synthetic ground truth below. Second, we turn it into an audit of existing probes, with completeness and selectivity deletion tests. Third, we use it to rank pre-training objectives. If reviewers judge that already standard, the experiments still settle an open empirical question that neither EEG benchmark addresses.
Why not just remove artifacts first, with eye and muscle reference channels or ICA cleaning, and probe the cleaned data? Cleaning removes artifacts that the cleaner can recognize. It does not remove subject-specific neural idiosyncrasies that fail to transfer. It also leaves stimulus-locked artifacts partly intact. We include an ICA-cleaned-input condition for every encoder. If cleaning alone closes the leave-one-subject-out gap, our method is unnecessary for that dataset, and we will say so.
Why not just report leave-one-subject-out accuracy of the ordinary probe? That number shows that a probe fails to transfer but not which directions caused the failure. It also cannot flag a probe that transfers only because an artifact is shared across subjects. The semi-synthetic test below measures exactly what leave-one-subject-out accuracy misses.
Why not just use a small nonlinear probe or fine-tune? EEG-FM-Bench already shows fine-tuning wins on accuracy. A nonlinear head can build the missing structure itself, which defeats the question of what the frozen representation contains. We report both as upper references but do not treat them as alternatives to an audit.
Why not just apply TIN or LiMVAM directly to the raw sensor data? These methods recover causal structure among a few variables. TIN was evaluated on tabular data such as the Sachs protein network, where it reached a graph distance of 0.33 against 0.8 for ICA-LiNGAM. Neither method says anything about what a foundation model's latent space contains. We do run the sensor-space version of the shared-subspace analysis, as a check on whether the encoder preserves or destroys sharing that exists in the raw data.
Decisive Experiments
Semi-synthetic benchmark. This comes first because it has ground truth. We take real background EEG from the Temple University Hospital corpus and inject three kinds of signal. The first is a small set of non-Gaussian sources with a known linear causal graph, synchronized across simulated subjects and mixed through subject-specific forward matrices. The second is a label-correlated "artifact" source injected privately, with independent timing per subject. The third, as a hard case, is a label-correlated artifact injected synchronously in all subjects. We sweep the noise-to-signal ratio from 0.25 to 4.0, matching the range in which TIN was validated, and pass everything through each frozen encoder.
Three predictions follow. The shared-subspace fraction of an ordinary probe should fall when the private artifact is made predictive. Leave-one-subject-out accuracy of the shared-subspace probe should stay flat across the noise sweep. The synchronous artifact should be classified as shared. That last outcome is a known limit: the method separates "stimulus-driven" from "private", not "neural" from "artifact", and we will state this as the method's boundary. The test fails if the recovered subspace's principal angles to the true injected sources grow with noise-to-signal ratio after the encoder but not in sensor space. That would show the encoder breaks the independence the argument needs.
Real data. We use SEED and DREAMER, in which all subjects watch the same film clips, so clip time gives the synchronization. Encoders are ST-EEGFormer, EEG2Rep, LaBraM and EEGPT. The measures are balanced accuracy and kappa, within subject and leave-one-subject-out, plus the shared fraction and the deletion completeness and selectivity for each probe. For DREAMER, the only reported figure found is EEG2Rep's linear-probe margin of "over 4%" against its nearest baseline. No absolute accuracy is reported in the file, so we recompute all baselines.
To bound the shared-artifact limit, we estimate the shared subspace from simultaneously recorded frontal (eye-dominated) channels alone and measure its principal angles to the full-montage shared subspace. A large overlap means the "shared" signal on these datasets is substantially ocular, and we will report it as such.
What would decide the matter. The proposal succeeds if three things hold. The shared-subspace probe beats the ordinary probe on leave-one-subject-out accuracy on both datasets. It beats both ablations. Deleting the shared subspace drops leave-one-subject-out decoding to near chance while deleting the private subspace does not. The proposal fails if the within-subject principal-component ablation does as well, or if the semi-synthetic test shows the encoders destroy cross-subject identifiability. Either outcome is informative. The comparison across encoders then gives the first representational evidence on whether masked raw reconstruction or latent prediction preserves more of the stimulus-driven signal.
References
- Are EEG Foundation Models Worth It? Comparative Evaluation with Traditional Decoders in Diverse BCI Tasks
- EEG-FM-Bench: A Comprehensive Benchmark for the Systematic Evaluation and Diagnostic Analyses of EEG Foundation Models
- Multi-View Causal Discovery without Non-Gaussianity: Identifiability and Algorithms
- Causal Discovery with Linear Non-Gaussian Models under Measurement Error: Structural Identifiability Results
- Independence Testing-Based Approach to Causal Discovery under Measurement Error and Linear Non-Gaussian Models
- Measuring the Reliability of Causal Probing Methods: Tradeoffs, Limitations, and the Plight of Nullifying Interventions
- EEG2Rep: Enhancing Self-supervised EEG Representation Through Informative Masked Inputs
- Filter Bank Common Spatial Pattern Algorithm on BCI Competition IV Datasets 2a and 2b