A spiking CNN, one time step at a time
What the network receives at each step:
the image (direct)
random spikes (rate)
events (saccades, like N-MNIST)
Draw a digit
Clear
Random test digit
Input at step t
time steps T
show step t
▶ Play
Layer 1: 12 feature maps, spikes at step t
Layer 2 neurons over time (row = neuron, column = step)
Output: 10 neurons over time
Output spike counts (SNN)
ReLU CNN on the image, softmax
Accuracy on the 10,000 MNIST test images vs. number of time steps
spiking CNN, image as input (trained at T = 25)
spiking CNN, saccade events (trained at T = 30)
ReLU CNN