Benchmark Tasks for Evaluating the Computational Power of NMCs

The universality of a computational model for neural microcircuits can not be tested by evaluating their performance for a single computational task.

Instead, each microcircuit model should be tested on a large variety of computational benchmark tasks. Hence it is desirable that many users test the circuit models on their favorite computational problem. Furthermore it seems advantageous to have a set of benchmark tests on which different models can be compared on a qualitative basis.

In the following we describe some benchmark tasks which have already been used, and for which code is provided in the Learning-Tool package

Classification of jittered Spike Trains

Two arrays of d (default d=1) Poisson spike trains (freqency freq=20Hz, length Tmax=0.5sec) are generated, and fixed as templates 1 and 2. That is a template is a spatio temporal spike pattern. For i=1,2 one generates jittered versions of template i by varying each spike in template i by a random drawn amount (Gaussion distribuion with zero mean and a given STD; this STD is called jitter (default jitter=4ms)).

The task is to output the number of the template from which the spike train was generated. 0.65*nRuns jittered versions of the templates are used for training. The accuracy of the output is estimated testing on 0.35*nRuns new jittered spike trains (nRuns=100).

Parameters that one might want to vary for further exploration

    Parameter Default value  Description
    d1 the number of spike trains in each array
    jitter4 ms the STD of the Gaussion added to each spike
    freq 20 Hz frequency of the Poisson spike trains used as templates
    Tmax 0.5 sec length of templates

See the demo lsm/learning/demos/spike_train_classification for details.

Classification of Segments of jitterd Spike trains

All spike trains are of length Tmax=1.0 sec and consist of n=4 segments of Tmax / n=250 ms each. For each segment m=2 templates are generated randomly (Poisson spike train with a frequency of freq=20 Hz.

The actual input spike trains of length Tmax=1.0 sec are generated by choosing for each segment one of the m=2 associated templates, and then generating a jittered version of it.

The task is to output for each of the n=4 segments the number of the template from wich the corresponding segment of the input was generated.

Parameters that one might want to vary for further exploration

    Parameter Default value  Description
    d 1  the number of input spike trains
    jitter 4ms  the STD of the Gaussion added to each spike
    freq 20Hz  frequency of the Poisson spike trains used as templates
    Tmax 1sec  length of templates
    n 4  number of segments
    m 2  number of templates per segment

See the demo lsm/learning/demos/segment_classification for details.

Retrieval of Delayed Sum of Rates

See section 6.1 of (Maass et. al. 2002a) for a descriptio of this task.

The demo is not yet ready.

Multi-tasking in real-time

See Fig. 2 of (Maass et. al. 2002) for an example.

The demo is not yet ready.

 
(C) 2003, Thomas Natschläger last modified 07/10/2006