Figure 3.

We used EventProp and a time-to-first-spike loss function to train a two-layer leaky integrate-and-fire network on the Yin-Yang dataset. (A) Illustration of the two-dimensional training dataset. The three different classes are shown in red, green and blue. This dataset was encoded using spike time latencies (see D). (B, C) Training results in terms of test error and loss averaged over 10 different random seeds (individual traces shown as grey lines). (D) Data points (x, y) were transformed into $(x,1-x,y,1-y)$ and encoded using spike time latencies. We added a fixed spike at time $t_{\text{bias}}$. (E) Spike time latencies $\mathrm{\Delta }t$ of the three output neurons (encoding the blue, red or green class) after training, for all samples in the test set and a specific random seed. Latencies are relative to the first spike among the three neurons and given in units of $t_{\max }$. A latency of zero (bright yellow dots) implies that the corresponding neuron fired the first spike, determining the class assignment. Missing spikes are denoted using green crosses