First-hand information for everyone
Researchers show that the pattern of connections in a neural circuit can be used to program chaotic activity in a way that suits different computing tasks. Rather than only changing properties of individual neurons, they change the network topology—the way neurons are wired—to move the circuit between ordered and chaotic behavior. This gives a new, reconfigurable knob for computation that affects both how random the activity is and how fast signals spread.
The team used a continuous-time neural-circuit model made of excitatory and inhibitory neurons. Each neuron has a simple nonlinear response (a tanh activation) and each outgoing connection from a neuron is either +1 or –1 to respect Dale’s law (cells are strictly excitatory or inhibitory). The connection pattern was varied with a single disorder parameter β that interpolates between a regular lattice, a small-world network, and a random (Erdős–Rényi) graph. In examples shown, they used circuits of N = 300 nodes, connection sparsity p0 = 0.07, and an excitatory fraction E = 0.2, and studied dynamics for β values near 10−4 (regular), 0.1 (small-world), and 0.8 (random).
To quantify dynamics they introduce two time measures. The memory timescale τc comes from how fast the system’s autocorrelation decays: small τc means the system quickly loses memory and behaves more randomly. The propagation latency τp measures how long it takes a local perturbation to spread across the network: small τp means fast signal spreading. They also tracked static network measures like the clustering coefficient C (local feedback loops) and the average path length L (how many hops separate nodes). As β increases, clustering falls and path length shortens, and the circuit moves from slow, stable dynamics (regular networks), through a small-world regime with low latency and near-critical dynamics, to rapidly spreading, highly chaotic dynamics (random networks).
Why this matters: the study shows that network wiring is a practical control parameter for task-specific computation. By jointly tuning neuron-level properties and topology the authors map a “chaos–latency” phase diagram that identifies network regimes suited to different tasks. In particular, small-world networks near the edge of chaos appear useful when a task needs both low latency and some memory or retention. The authors also report that rapid on–off switching of chaos can be achieved by rewiring edges in small-world networks, and they support their claims with reservoir-computing benchmarks run across different topologies.
Important caveats: the results come from simulations of a specific mathematical model with chosen activation, sign constraints, and sparsity, not from physical devices or biological tissue. The excerpt notes some robustness checks (for example, on τp) but does not give full details of the benchmarks or how general the findings are across other models or real hardware. The paper therefore demonstrates a promising design principle—using topology as a reconfigurable control over chaos—but further work is needed to test how broadly it applies and how it performs outside the simulated setting.