About SpikeTime training of NMNIST

[ahmetakman]

Hello,
I am curious about something. The tutorial on the repo uses SpikeRate for the NMNIST training. And it was said " Since the target spike train is not known for this problem, we use SpikeRate loss and target high spiking rate for true class and low spiking rate for false class. "
So, if we assign some arbitrary distinct 10 spike trains as target spike trains, would not it be better to use the SpikeTime method which may provide better efficiency and latency. Am I wrong with my assumption?

Sincerely,

[bamsumit]

If you use SpikeTime, you are asking the network to match the 10 spike times exactly whereas if you are using SpikeRate corresponding to 10 spikes per sample interval, then 10 spikes anywhere is a valid solution. The SpikeTime solution is a special case of SpikeRate solution. This means that using rate, the final solution is flexible and, therefore, easier to train.

[ahmetakman]

Thank you very much for the explanation.