Computational demands shape seizure susceptibility in recurrent neural networks

Why do some brain areas slip into seizures more readily than others? Anatomy and physiology are part of the story, but in this talk we argue that the computation a network performs is itself a determinant of its vulnerability. Working in the language of recurrent neural networks and attractor dynamics, we contrast two computational regimes: networks that represent continuous variables (e.g. tracking a continuous position in space) establish barrier-free state manifolds, whereas networks that store discrete memories have to separate different states by establishing deep wells. A simple energy-landscape picture suggests that continuous networks can amplify perturbations into runaway activity more easily while discrete networks contain them. We test this idea across three systems: handcrafted spiking attractor networks, recurrent networks trained on continuous or discrete computational tasks, and in vivo Neuropixels recordings comparing medial entorhinal cortex (continuous, grid-cell dynamics) with hippocampal CA3 (discrete memory). Under a shared disinhibitory "seizure perturbation," the continuous systems destabilize sooner and drive stronger epileptiform activity, and causal silencing shows this depends on intact entorhinal output. Together, these results establish a direct link between the computation a network is built to perform and its susceptibility to seizures, showing that the very features that enable a network to process information also shape its vulnerability to pathological transitions.