Bounded substrate comparison

Artificial and spiking neural networks

When is a spike-based model meaningfully different from a conventional neural network?

software or accelerator

Conventional ANN

Valid claim: Performs the declared tensor computation under its numeric and timing model.

event-based dynamical model

Spiking neural network

Valid claim: Performs the declared state-and-event computation under its neuron and coding model.

Comparable axes

  • Same task and data split
  • Matched quality constraint
  • Full inference latency and resources

Non-equivalences

  • A spike is not automatically a biological action potential.
  • Operation counts are architecture-specific.

Comparison procedure

  1. 1.Freeze task and quality.
  2. 2.Document conversion or training.
  3. 3.Measure complete pipelines.

Prohibited inference

Do not infer biological realism, intelligence, or efficiency from the presence of spikes without matched task evidence and complete resource accounting.

Connected concepts and sources

Spiking neural networksBenchmarking, energy, and task equivalence
  1. NeuroBench: Advancing Neuromorphic Computing Through Collaborative, Fair and Representative Benchmarking · National Institute of Standards and Technology