Issue 28 · Pick 08 Neuroscience ✓ read
Computational demands shape seizure susceptibility in recurrent neural networks
The computation a brain region performs may set how easily it tips into a seizure. Not its excitability, not its inhibition, not its wiring density in the usual sense — but the shape of the state space its recurrent network is built to represent. Networks that encode continuous variables (like a grid cell's position) sit on a flat valley of near-equivalent states, and the authors argue that this very flatness gives runaway activity an easy road to travel. Networks that store discrete memories sit in separated wells with walls between them, and those walls contain the fire.
That is the whole thesis, and what makes it worth your time is that they test it three ways that fail in different directions — handcrafted attractors, trained spiking RNNs, and live mouse recordings with a causal knockout — and all three point the same way.
Why regional seizure risk is a real puzzle
Focal seizures start in one place. Sometimes that's because that place is already damaged. But seizures also start focally when the whole brain is perturbed uniformly — a fever, a drug, an electrolyte swing. If the insult is global and the brain is otherwise healthy, then something intrinsic makes one region light up first. The usual suspects are local: intrinsic excitability, weak inhibition, excitatory circuit motifs. All plausible, all hard to compare across regions, and none tied to what the region is actually for.
The move here is to ask whether the job description itself carries the risk.
The energy-landscape intuition
Picture a recurrent network's dynamics as a ball rolling on an energy landscape over the space of population activity patterns. Low points are the states the network likes to settle into.
A continuous attractor — the standard model for grid cells maintaining a spatial estimate — has a continuous manifold of low-energy states. You can slide the activity bump smoothly from one location to a neighboring one at essentially no energetic cost. That flatness is exactly the feature that lets it track a continuous variable.
A discrete attractor — the standard model for CA3 storing distinct episodic memories — has isolated wells separated by high barriers. That separation is exactly the feature that keeps memories from smearing into each other (pattern separation).
Now push both networks toward more excitation. In the continuous case, extra activity has a frictionless direction to spread: it flows along the manifold, bumps grow, drift, merge, and eventually you get traveling wavefronts of activity sweeping the sheet. In the discrete case, the barriers block spread; activity that can't escape a well just makes the neurons in that well fire and go refractory together — a synchronized burst, not a sustained storm.
The authors connect this to the fluctuation–dissipation theorem: a physical system that responds strongly to inputs also fluctuates strongly on its own. A continuous representation is built to respond — that responsiveness is the same thing as fragility. That framing, function and dysfunction as two faces of one property, is the paper's most quotable idea.
Building it by hand, then letting training discover it
Handcrafted attractors. They build 2D spiking sheets with excitatory and inhibitory neurons at each site. Crucially, they match the two network types on everything they can: same neuron model, same I→E broad inhibition, and — importantly — the same number of incoming and outgoing synapses per neuron. The only difference is topology: continuous networks wire excitation locally (short-range), discrete networks wire it into spatially interleaved assemblies via a degree-preserving edge swap. So any difference in seizure behavior can't be blamed on more synapses or higher connectivity.
They then apply a disinhibitory current (lowering inhibitory neurons' resting potential — mimicking a GABA-A block). The continuous network degrades earlier and climbs to higher population activity; at high disinhibition it produces traveling wavefronts. The discrete network holds its states longer and, when it breaks, bursts rather than sustains. This held across a parameter sweep: 58 of 240 parameter combinations produced a valid seizure-like transition, and within those the continuous-vs-discrete asymmetry was robust.
Trained RNNs. The handcrafted result could be an artifact of the specific wiring rule, so they redo it with optimization. A recurrent layer of spiking LIF neurons is trained by backprop-through-time (surrogate gradients) on two tasks that differ only in the input–output map: the continuous task must reproduce an input bump's exact position; the discrete task must report which half it fell in. Same input, same architecture, same output format.
The clever part is the regularizer stack. They penalize deviations from a shared target population activity level, a shared distribution of recurrent weights, and a shared distribution of biases. This forces the two trained networks to be matched on incidental statistics, so when they perturb (uniformly shift all recurrent biases up) and the continuous-trained RNN blows up into sustained high activity while the discrete-trained one merely hops between its two target states, you can attribute the difference to task identity, not to one network happening to be more excitable. The continuous-trained networks fail even within the 100-step training horizon; the discrete ones survive it. Analysis of the learned weight matrices confirms they discovered attractor-like structure on their own.
The animals
Now the real test. Simultaneous Neuropixels recordings from mouse MEC (grid cells, continuous attractor) and CA3 (episodic memory, discrete attractor), under urethane anesthesia to strip out behavior and external drive. The uniform perturbation from the models becomes a systemic i.p. injection of PTZ, a GABA-A antagonist.
Three cohorts of 5 mice each for the high-dose seizure experiments. In wildtype mice, high-dose PTZ produces epileptiform discharges in both regions, and a Granger directionality index shows MEC LFP leads CA3 — the negative direction dominates. Then the causal handle: virally express tetanus toxin light chain (which blocks synaptic release without stopping spikes) selectively in superficial MEC excitatory neurons or in CA3 pyramidal cells.
- Silence MEC output → discharges drop in both MEC and CA3, and the MEC-leads-CA3 directionality weakens. MEC's outputs are necessary for the discharge in both regions.
- Silence CA3 output → CA3 discharges drop but MEC's are unchanged, and directionality shifts toward MEC. CA3 is a follower.
For the mechanism prediction, they use a low PTZ dose that perturbs without seizing, spike-sort single units (6 mice per group), and compute temporal continuity — cosine similarity of the detrended population state across successive 200 ms bins. Their models predict continuous attractors drift smoothly (high continuity) while discrete ones hop (continuity near zero). In wildtype mice, MEC shows high temporal continuity and CA3 low, exactly the predicted split — despite no significant difference in overall firing rate between the regions. And silencing MEC output pushes MEC's temporal continuity toward zero: the smooth trajectories require intact MEC recurrence. Silencing CA3 leaves MEC's continuity untouched.
That last result is the one I find most persuasive, because it's not just a correlation between region and dynamics — it shows the smooth "continuous-attractor" signature depends on the recurrent connectivity the theory says matters.
How strong is this, really
The convergence is the strength. Handcrafted networks control for synapse count, trained RNNs control for activity and weight statistics via explicit regularizers, and the animal experiments add a causal knockout rather than a mere comparison. Each approach's confounds are different, so their agreement is hard to explain away.
Where to stay skeptical:
- Only two regions, one perturbation class. MEC vs CA3 is a single contrast, and the models' clean continuous/discrete dichotomy is an idealization — both regions are heterogeneous. The authors are candid that this is "a contrast between networks skewed towards different computational topologies," not proof either region is a pure attractor.
- In vivo the confounds can't be regularized away. MEC and CA3 differ in cell types, inhibitory architecture, and much else. The temporal-continuity match and the tetanus-toxin causality are the load-bearing evidence that the computational difference is what matters, not just any difference.
- Small n and anesthesia. Five to six mice per group; urethane anesthesia (chosen deliberately to isolate intrinsic dynamics, but a departure from awake seizures). Many key numbers live only in the figures, not the text.
- PTZ is global and acute. This tests intrinsic susceptibility to a uniform insult, which is the point — but it says less about chronic epileptogenesis.
What to read
Go to the Results section on the trained RNNs and Figure 3: the task design and the matched-statistics regularizers are the cleverest part of the paper and the clearest demonstration that the effect follows computation rather than incidental network properties. Then read the Figure 5 / low-dose PTZ single-unit analysis, where temporal continuity plus the tetMEC dependence turns a modeling prediction into a testable, falsified-if-wrong claim about real cortex — and it holds.
If it generalizes, the reframing is genuinely useful: to predict which region will seize under a systemic insult, ask what kind of representation it maintains. Continuous, low-dimensional, manifold-like codes should carry more intrinsic risk than discrete, well-separated ones — a hypothesis you could go test in thalamus, prefrontal working-memory circuits, or head-direction systems next.