Open akorgor opened 4 months ago
@akorgor That's a very impressive plot. Let me test this idea with the N-MNIST task. I will post here the results.
@akorgor Here the plot
Clearly the change, although using the same parameters, does not work very well for this case.
Perhaps a parameter resetting might help. I will keep looking into this.
@akorgor Here the plot
Clearly the change, although using the same parameters, does not work very well for this case.
Perhaps a parameter resetting might help. I will keep looking into this.
Could you please try again with the parameter changes introduced in https://github.com/JesusEV/nest-simulator/pull/20/commits/0dfaad73bf1a6a0d6b0c06102c866cb0ba0b1136?
For a brief experiment of 30 iterations, the primary reason for the failure to converge in the n-mnist task appears to be the omission of the factor (1-kappa) in the filtering process of the eligibility trace.
Pull request automatically marked stale!
This PR removes the
regular_spike_arrival
flag and the corresponding mechanism, which was initially in place to reproduce the behavior of the original model. SinceP_z_in=1
that variable is no longer needed and is removed. To match the regular spike arrival at the end of the time step (the commit message erroneously says at the beginning of the time step), the spike threshold crossing reset is moved to the current time step $t$, resulting in the following timing in the membrane potential update equation:instead of the previous
To use the new propagator in the sine waves task requires scaling the spike threshold with the old default propagator $v\mathrm{th, new} = v\mathrm{th, old} / (1-\mathrm{exp}(\mathrm{d}t / \tau\mathrm{m}))$. To make the sine waves task with the new propagator consistent with the evidence accumulation task and since the convergence benefits from it, the membrane time constants in the system are decreased from 30 ms to 20 ms, leading to a new threshold potential equal to the one in the evidence accumulation task $v\mathrm{th, new} = 0.03 / (1-\mathrm{exp}(\mathrm{d}t / 20))=0.6$
The convergence on the sine waves task improves with this change: