This version of the documentation is NOT an official release. You are looking at ‘latest’, which is in active and ongoing development. You can change versions on the bottom left of the screen.

correlation_detector – Device for evaluating cross correlation between two spike sources


The correlation_detector device is a recording device. It is used to record spikes from two pools of spike inputs and calculates the count_histogram of inter-spike intervals (raw cross correlation) binned to bins of duration \(\delta_\tau\). The result can be obtained via GetStatus under the key /count_histogram. In parallel it records a weighted histogram, where the connection weights are used to weight every count. In order to minimize numerical errors, the Kahan summation algorithm is used when calculating the weighted histogram. Both are arrays of \(2*\tau_{max}/\delta_{\tau}+1\) values containing the histogram counts in the following way:

Let \(t_{1,i}\) be the spike times of source 1, \(t_{2,j}\) the spike times of source 2. histogram[n] then contains the sum of products of the weight \(w_{1,i}*w_{2,j}\), count_histogram[n] contains 1 summed over all events with \(t_{2,j}-t_{1,i}\) in

\[n*\delta_\tau - \tau_{max} - \delta_\tau/2 n*\delta_\tau - \tau_{max} + \delta_\tau/2\]

The bins are centered around the time difference they represent, but are left-closed and right-open. This means that events with time difference -tau_max-delta_tau/2 are counted in the leftmost bin, but event with difference tau_max+delta_tau/2 are not counted at all.

The correlation detector has two inputs, which are selected via the receptor_port of the incoming connection: All incoming connections with receptor_port = 0 will be pooled as the spike source 1, the ones with receptor_port = 1 will be used as spike source 2.




Time when to start counting events. This time


be set to at least start + tau_max in order to


edge effects of the correlation counts.



Time when to stop counting events. This time


be set to at most Tsim - tau_max, where Tsim is


duration of simulation, in order to avoid edge effects of the correlation counts.



Bin width. This has to be an odd multiple of the resolution, to allow the symmetry between positive and negative time-lags.



One-sided width. In the lower triagnular part events with differences in [0,


are counted. On the diagonal and in the upper triangular part events with differences in (0, tau_max+delta_tau/2].



The number of pools. This defines the range of receptor_type. Default is 1. Setting N_channels clears count_covariance, covariance and n_events.



read-only - raw, weighted, cross-correlation


synaptic weights

Unit depends on model


list of integers

read-only - Correction factors for Kahan summation algoritm


list of integers

Number of events from source 0 and 1. By setting n_events to [0,0], the histogram is cleared.


This recorder does not record to file, screen or memory in the usual sense.

Correlation detectors IGNORE any connection delays.

Correlation detector breaks with the persistence scheme as follows: the internal buffers for storing spikes are part of State_, but are initialized by init_buffers_().

@todo The correlation detector could be made more efficient as follows (HEP 2008-07-01): - incoming_ is vector of two deques - let handle() push_back() entries in incoming_ and do nothing else - keep index to last “old spike” in each incoming_; cannot

be iterator since that may change

  • update() deletes all entries before now-tau_max, sorts the new entries, then registers new entries in histogram


See Auto- and crosscorrelation functions for spike trains[cross_check_mip_corrdet.py] in pynest/examples.



See also

Device, Detector