Referee 3's review:

Using computer simulations the paper shows that a neuron may learn to fire just in response to a single pattern which is presented repeatedly, if synaptic initial conditions are appropriately prepared and the STDP function is biased towards depression. Similar findings have been reported previously, but the paper considers somewhat more general input statistics. Up to the 6 points mentioned below, the paper seems technically sound.

Minor Comments

1) Page 7, Second but last paragraph.

The authors should properly specify the stochastic process generating the instantaneous firing rate.

2) Page 7, Last paragraph.

In generating input spike patterns a special additional rule is introduced assuring that there is a least one spike every 50 ms. The authors claim that this is important, but I do not understand why. Further it is hard to imagine that such a rule is implemented in the brain

3) Page 8, Definition of the EPSP.

Two normalizing constants, K and 1/(1-\tau_s/\tau_m) are used. One would be enough.

4) Page 9, Definition of the reset kernel \nu

Four normalizing constants are used, two would do.

5) Page 9, Fourth paragraph

The authors say that the membrane potential is only computed at input spike times. How are the output spike times detected?

6) Page 9, Fifth paragraph

It is claimed that the event driven integration is much faster than using a fixed step size. Given that the are 2000 afferents spiking at some 50 HZ, an event occurs on average every 0.01 ms. Since the membrane time constant is 10 ms, a time constant of 0.01 ms in a fixed step size routine would seem extremely conservative. It is not clear how event driven integration can speed up for such dense events.

----------
N.B. These are the general comments made by the reviewer when reviewing this paper in light of which the manuscript was revised. Specific points addressed during revision of the paper are not shown.