Reliable computation with stochastic, asynchronous neuron ensembles

Establish formal methods and conditions under which ensembles of asynchronous, event-driven, and intrinsically stochastic neurons can be constructed to perform reliable and arbitrary computations on neuromorphic systems despite non-deterministic neural behavior.

Background

The paper emphasizes that neuromorphic computation is event-driven and highly parallel, making it inherently asynchronous and often non-deterministic. Biological neural systems exhibit variability and stochasticity, which may be a feature rather than a limitation, but ensuring reliable computation in such settings poses theoretical challenges.

The authors note that while computation abstracted above the neuron level can be reliable (e.g., sampling algorithms approximating fixed distributions), understanding how to construct neuron-level ensembles that yield reliable and arbitrary computations remains unresolved, aligning with current questions in theoretical neuroscience.