Data-driven identification of functional networks in artificial and biological neural networks

Understanding how the brain represents information is a central challenge in neuroscience and a practical bottleneck for brain-computer interfaces. Existing analytical tools cannot identify neural representations directly from neural activity data. We introduce MultiPEC, a data-driven method that discovers neural representations by quantifying how sets of signals jointly reduce context prediction error. Applied to artificial neural networks, MultiPEC uncovered class-specific subnetworks whose targeted ablation disproportionately impaired performance, demonstrating their causal role in feature recognition. Applied to EEG recordings from 24 participants, MultiPEC revealed functional signatures of auditory and visual processing. Classification analyses showed that MultiPEC captured fine-grained stimulus submodalities (levels of stimulus meaningfulness) more effectively than broad stimulus domains, highlighting the context-sensitive nature of neural representations. Together, these results establish MultiPEC as a scalable approach for identifying data driven (natural) representations in both biological and artificial systems, with potential applications in adaptive neurotechnology, clinical diagnostics and cognitive neuroscience.