Competitive learning in ANNarchy

[Nestor Rueda]

Dear all, thank you in advance, for all the information which ANNarchy team have given me. 

I try to develop a simple model of competitive learning but I have two questions:

1. I want set the inputs to population1(spiking neurons) periodically in a fixed duration trials. I am using SpikeSourceArray to set the inputs,  Can it be done as follows?

spikeTrain = getSpikeTrain(nNeurons)

sTrain = SpikeSourceArray(spikeTrain)

learP = Population(nNeurons, neuron=ann.IF_cond_alpha)

projlpNet = Projection(pre=sTrain, post=learP, target='exc', synapse = STDP)

projlpNet.connect_all_to_all(weights=1.0)

compile()

for trial in range(10):

    sTrain.spike_times = getSpikeTrain(nNeurons)

    simulate(5.0)

2. I want to develop a competitive learning model as (Masquelier et al. 2009) but I do not know if the next model can be implemented in ANNarchy:

(see attachment picture)

As the picture show, a set of spike trains (may be generated through SpikeSourceArray) are the inputs to a population of neurons, then, if a neuron fires, it sends a inhibitory postsynaptic potential to its neighbors. The next code doesn't work correctly, and I believe that is wrong. Can I implementing this recurrent inhibitory behaviour in ANNarchy?

spikeTrain = getSpikeTrain(nNeurons)

sTrain = SpikeSourceArray(spikeTrain)

learP = Population(nNeurons, neuron=IF_cond_alpha)

projlpNet = Projection(pre=sTrain, post=learP, target='exc',  synapse = STDP)

projlpNet.connect_all_to_all(weights=1.0)

projlpNet = Projection(pre=learP, post=learP, target='inh', synapse = STDP)

projlpNet.connect_all_to_all(weights=1.0)

Could you help me with the above questions? 

Thanks again to all,

All the best, 

[Julien Vitay]

Hi Nestor,

1. Unfortunately, in a SpikeSourceArray, you need to provide the absolute time of the spikes, they are not relative to each call to simulate(). I will think about a way to allow relative spike times, but meanwhile I think it is faster if you modify your getSpikeTrain method so that it takes a reference time (what would be t=0 for your periodic input, but may be 0, 5, 10, 15ms... inside your loop) and adds it to all the spike times. You can access the current time with get_time() in ANNarchy.

2. I see nothing wrong about the code, it reflects the picture. What is not working exactly? I'll try to have a look at the Masquelier paper, it could be a nice example for ANNarchy.

Best

Julien

[Nestor Rueda]

Hi Julien, thank you for your fast reply.

It would be great allow relative spike times, but meanwhile, I will set the inputs of population periodically in a loop. 

May be the competitive learning example didn't work for a misuse of SpikeSourceArray. I will fix it.

I have attached a copy of Masquelier's paper, may be helpful. Would be a good example for ANNarchy.

All the best,

Nestor.

[Julien Vitay]

Hi,

sorry it took so long.... SpikeSourceArray now can be reset, so the spike times are relative:

for trial in range(10):

    sTrain.reset()

    simulate(5.0)

There is now a documentation in the manual on how to set inputs:

http://annarchy.readthedocs.io/en/latest/manual/Inputs.html

It is still in the develop branch, will be integrated to the 4.6.0 release.

Best

Julien