Brain Decoding Alignment Erases User Identity Without Changing Thought Geometry

Brain-computer interface systems that decode imagined hand or foot movements from electrical brain activity struggle with a fundamental limitation known as inter-subject variability, where neural patterns differ wildly from one person to another. To make brain-computer decoders work across different users without requiring lengthy custom training for each new patient, researchers commonly rely on subject-wise alignment techniques. These mathematical transformations shift each person’s neural data onto a shared reference point before the system attempts to decode their intentions. Although this preprocessing step reliably boosts classification accuracy across users, a new study reveals that it achieves this feat by stripping away individual user identity rather than by genuinely aligning the internal geometric arrangement of task thoughts.

The research team behind the investigation, consisting of Anna Notaro, Letizia Rossato, and Sadasivan Puthusserypady, examined how alignment alters the internal geometry of neural representations using electroencephalogram recordings. Their findings were published on the arXiv preprint server in November 2026. The authors demonstrate that decoding accuracy conflates two entirely different properties, namely subject identity and relational structure. Subject identity measures whether a single burst of brain activity reveals who produced it, whereas relational structure refers to how different mental tasks relate to one another within an individual mind, capturing which imagined movements are processed as similar or distinct.

Why Accuracy Conflates De-Identification and Real Alignment

When scientists evaluate brain-computer interfaces, they typically look at decoding accuracy as proof that cross-subject transfer is working. However, accuracy only shows whether a classifier can successfully categorize tasks across users, leaving the underlying representational geometry unexamined. To separate these effects, the authors analyzed three distinct motor-imagery electroencephalogram cohorts containing a combined total of 38 participants. These datasets included 19 participants from an in-house dry-electrode study, 9 participants from a public 4-class benchmark, and 10 participants from a public 6-class dataset.

Each brain recording trial was translated into an eight-channel spatial covariance matrix, which captures the rhythmic power and electrical channel interactions modulated during imagined movements. Covariance matrices form a curved mathematical space known as a symmetric positive-definite manifold, where standard flat-space distance calculations fail. The team applied per-subject recentering, a standard alignment technique that calculates each participant’s average covariance point on the manifold and maps it directly to the identity matrix. They measured subject identity using a dedicated probing classifier and evaluated relational organization using representational similarity analysis.

Probing the Collapse of User Identity and Geometric Invariance

The results exposed a striking dissociation between de-identification and relational geometry. Following per-subject recentering, the subject-identification probe accuracy plummeted from 0.959 down to 0.054 on the in-house cohort, effectively reaching the baseline chance level of 0.053. Similar collapses occurred across the public datasets, where identity probes fell from 0.978 to 0.076 and from 0.981 to 0.066. Single trials completely ceased to betray their source user, confirming that the alignment successfully removed individual distributional offsets.

At the same time, the relational geometry measured by the affine-invariant Riemannian distance remained entirely unchanged. The mathematical proof established that recentering acts as a congruence transformation which preserves all pairwise distances between condition means within a subject’s representational space. Consequently, the representational dissimilarity matrices computed before and after alignment exhibited maximum entrywise changes on the order of 10 to the power of negative fourteen, which is merely a floating-point numerical artifact. Cross-subject agreement in task organization remained significantly above zero but below the estimator noise ceiling, proving that removing a person’s digital fingerprint did not make their internal thought organization match anyone else’s.

Testing Differentiable Surrogates and Training Pressures

To test whether artificial neural networks could be forced to achieve true relational alignment during training, the researchers experimented with a differentiable similarity regularizer based on Centered Kernel Alignment. By adding this loss penalty with varying weights up to thirty times the baseline, the network optimized directly for cross-subject similarity. While batch-level similarity scores increased during training, held-out cross-subject agreement failed to improve.

The analysis revealed that apparent reductions in between-subject differences were actually driven by a collapsing reliability ceiling rather than genuine geometric convergence. The authors noted that spatial dimension reduction in certain deep architectures, such as reducing eight-channel matrices down to smaller dimensions, inadvertently stripped away valid class-relevant neural variations. These experiments highlight that current alignment pipelines succeed by shifting coordinate systems and erasing user-identifying biases rather than by restructuring the geometric relationships between cognitive tasks.

The study was posted to the arXiv preprint server on November 11, 2026.