Incorrect summation of synaptic currents for iaf_psc_exp_ps

[apdavison]

When two inputs, with equal but opposite weights, converge on an iaf_psc_exp_ps neuron with different time constants for the ex and in synapses, the voltage trace is flat. This problem does not occur with iaf_psc_exp or if the inputs arrive at different times.

To reproduce, run the following once with precise=True, again with precise=False:

import matplotlib.pyplot as plt
import nest

precise = True
relative_delay = 0.0  # the problem does not occur if this is positive

nest.SetKernelStatus({'off_grid_spiking': precise})

if precise:
    neuron_model = 'iaf_psc_exp_ps'
    filename = "with_ps"
else:
    neuron_model = 'iaf_psc_exp'
    filename = "without_ps"

if relative_delay:
    filename += "_delay"

neuron = nest.Create(neuron_model, 1, {'tau_syn_ex': 12.0, 'tau_syn_in': 4.0, 't_ref': 0.1})

spike_sourceD = nest.Create('spike_generator', 1)
spike_sourceR = nest.Create('spike_generator', 1)
nest.SetStatus(spike_sourceD, 'spike_times', [[40.0]])
nest.SetStatus(spike_sourceR, 'spike_times', [[40.0]])

nest.CopyModel('static_synapse', 'decay_syn', {'weight': 200.0, 'delay': 0.5})
nest.CopyModel('static_synapse', 'rise_syn', {'weight': -200.0, 'delay': 0.5 + relative_delay})

nest.Connect(spike_sourceD, neuron, syn_spec='decay_syn')
nest.Connect(spike_sourceR, neuron, syn_spec='rise_syn')

recorder = nest.Create('multimeter')
nest.SetStatus(recorder, {'withtime': True, 'record_from': ['V_m']})
nest.Connect(recorder, neuron, 'all_to_all')

nest.Simulate(200.0)

data = nest.GetStatus(recorder)[0]['events']
senders = data['senders']
times = data['times']
vm = data['V_m']

for n in neuron:
    plt.plot(times[senders==n], vm[senders==n])

plt.savefig(filename + ".png")

precise = False

precise = True

NEST: 2.10.0, Python: 3.4.2

Thanks to Hajime Yamauchi for the original bug report to the NeuralEnsemble Google group.

[heplesser]

One ignores Donald K at ones own peril: premature optimization is the root of all evil .... The problem is that SliceRingBuffer::get_next_spike() sums all spikes arriving at exactly the same time and then returns the summed weight. In the test case, this sum is zero, thus the flat membrane potential.

This was not a problem really in the original iaf_psc_alpha_canon model, because that had only a single synaptic time constant for both excitatory and inhibitory input. iaf_psc_exp_ps allows for different time constants, but still puts all spikes into one buffer and only distinguishes between E and I spikes when the buffer is read out. At this point, it gets the summed weights of simultaneous spikes and we are in trouble.

The proper solution for this is to place excitatory and inhibitory spikes in separate buffers in the handle() method, as is done in iaf_psc_exp.

One should probably also remove the spike accumulation from SliceRingBuffer::get_next_spike(), since the probability of simultaneous spikes in a precise simulation is vanishing (spike offsets would have to agree to the last bit).

[terhorstd]

Note also the discussion on buffers in multi-synapse models in the 2016-09-19-Open-NEST-Developer-Video-Conference...