BCIwiki (bciwiki.com) — According to arXiv, the preprint "Neural Decoding as Cognitive Inference" was uploaded on October 8, 2026, and has not been peer reviewed. The study examines how stable cognitive states can be extracted from continually changing neural activity, recasting neural decoding as cognitive inference constrained by brain-intrinsic priors.
The approach draws on Bayesian brain theory, using prior information to interpret observations. The authors argue that existing decoding methods map neural observations to predefined external labels under a stimulus-response principle and may capture recording-specific spurious correlations. Their method instead applies brain-intrinsic prior constraints to produce what they call "meta-neural semantic representations": high-level semantic representations of neural observations.
Decoding experiments covered five neural recording modalities and three cognitive domains: motor, perception and internal mentation. The study reports that the method reorganized the geometry of neural observation representations. The resulting meta-neural semantic representations exhibited consistent geometric relationships across cognitive tasks and enabled recovery of stable cognitive states from variable neural observations.
The authors interpret these results as suggesting that the brain may sustain relatively stable cognition by inferring cognitive states from changing neural activity, without requiring fixed activity patterns. This is a proposed mechanistic explanation, not an established account of how the brain maintains cognitive stability.