Real DVS Gesture recordings, classified in your browser by a spiking CNN and two ordinary CNNs

Events, step t

32 × 32 event counts in one 60 ms bin (blue ON, orange OFF): what the spiking network receives at this step.

All 30 steps summed

What a network without time sees: one "event image" of the whole 1.8 s window.

Spiking CNN: output spikes so far

CNN on every frame: average so far

CNN on the summed image

3D view: what the camera produces vs. what each network receives

Test accuracy during training: which surrogate gradient?

Where the three networks go wrong (264 test recordings, 6 people they never saw)

Spiking CNN over time

CNN on every frame