Turns out spiking neural networks don't need multiple spikes per neuron to be just as capable.
A new theoretical paper proves that spiking neural networks built from neurons that fire only once are mathematically equivalent, in approximation power, to networks where neurons fire many times. The researchers show that any approximation bound proven for multi-spike networks also holds for single-spike networks, as long as the number of neurons is scaled up or down linearly. The reverse holds too: results proven for single-spike networks carry over to the multi-spike case. The equivalence covers a broad class of neuron models, including the widely used leaky integrate-and-fire model with subtractive reset.
Single-spike neurons are commonly assumed to be a limitation, since firing only once is a much simpler and more constrained way for a neuron to communicate than firing repeatedly. Almost all existing approximation-bound research for spiking networks assumes single-spike neurons for exactly that reason. This paper's equivalence means that research was not working around an artificial restriction: it applies with equal force to multi-spike networks.
The paper is theoretical, not a benchmark win, but it removes a specific technical excuse for dismissing single-spike designs as underpowered.