Summary: Supplementary Methods
This section gives additional information about the choice of parameters and the initial conditions
used in the numerical simulations of Figs. 4-8. We have checked that none of the reported results
required fine-tuning of the model parameters.
Number of synapses. All shown simulations except Fig. 4a and Fig. 7c-d were made with N =
500. This choice was dictated by computation time and memory constraints. In Fig. 4a, we also
show the tempotron's capacity limit for N = 1000 and N = 1500. In Fig. 7c-d, where each afferent
spiked three times, the total number of input spikes was approximately preserved by reducing N
to 168. By simulating up to N = 10, 000 for selected parameters and reduced sample size, we
checked that the shown behavior approximates the large N limit.
Time scales. In most of the results, we used the value = 15 ms for the PSP decay time constant.
This is close to the optimal value (Fig. 4b) when assuming a reasonable pattern duration of T =
500 ms. For the decay time constant of synaptic currents, we have chosen s = /4, which for
= 15 ms is compatible with experimental results for fast synapses.
Tempotron learning. In the tempotron learning, we chose the maximal value of synaptic changes
to be = 10-4
/V0 (except in Fig. 4). This yields near optimal convergence time with N = 500,
T = 500 ms, and = 15 ms, the parameter valuesq used throughout the work. When is far
from its optimal, learning times become sensitive to the choice of . Empirically we found that
= 3 × 10-3