CogSpike verifies spiking neural nets with 17x state reduction
New tool cuts state space exponentially, enabling verification of stochastic SNNs.
Spiking neural networks (SNNs) model biological dynamics more faithfully than standard artificial networks, but their stochastic, event-driven behavior requires probabilistic models that suffer from exponential state space growth in verification. Researchers from the team of Nikan Zandian Jazi, Elisabetta De Maria, and Christopher Leturc present CogSpike, a unified workbench that integrates SNN design, simulation, and PRISM-based formal verification. Their key innovation is a weight-discretized quotient model abstraction that maps continuous synaptic weights to a compact integer range while preserving each synapse's relative contribution. This yields exponential state space reduction—approximately 17× per neuron for discretization parameter W=3—verified across seven canonical topologies.
The tool comes with formal correctness guarantees: a two-sided fidelity theorem confines firing disagreements to a bounded gray zone around threshold, and an Asymptotic Silence theorem ensures that unforced neurons fall permanently silent. These mathematical assurances make CogSpike a rigorous platform for verifying properties of probabilistic SNNs that were previously intractable. Paper accepted at ICANN 26, the work opens new avenues for safe deployment of neuromorphic computing in mission-critical applications where deterministic abstractions are inadequate.
- CogSpike integrates SNN design, simulation, and PRISM-based formal verification into one isomorphic tool chain.
- Achieves ~17x state space reduction per neuron using weight-discretized quotient model with W=3 discretization parameter.
- Provides formal guarantees: two-sided fidelity theorem for firing disagreements and asymptotic silence for unforced neurons.
Why It Matters
Makes verification of biologically realistic, stochastic SNNs feasible for safety-critical neuromorphic applications.