Behaviour of Random Number Generation

[andreARibe]

Hello!

To experiment with stochastic models in GeNN I developed the following code ("Beta" is unused bc this is a simplification):

NUM_NEURONS = 256
BETA_INIT   = 1
TIMESTEPS   = 10

sig_model = create_neuron_model(
    "Sigmoidal",
    params=["Beta"],
    sim_code=
    """
    if(id == 0){
        printf(\" 1:%f \", Isyn);
    }
    if(id == 2){
        printf(\" 2:%f \", Isyn);
    }
    """,
    threshold_condition_code=
    """
    Isyn < 0
    """
    )


if __name__ == "__main__":
    model = GeNNModel("float", "Stochastic_Hopfield_Neurons")
    i_start = {"Beta": BETA_INIT}
    pn    = model.add_neuron_population("pn", NUM_NEURONS, sig_model, i_start, {})
    pn.spike_recording_enabled = True
    cs = model.add_current_source("noise", "GaussianNoise", pn,
                                {"mean": 0.0, "sd": 1.0})

    model.build()
    model.load(num_recording_timesteps = TIMESTEPS)
   
    for t in range(TIMESTEPS):
        model.step_time()
   
    model.pull_recording_buffers_from_device()

    pn_spike_times, pn_spike_ids = pn.spike_recording_data[0]

    print(f"\nGENN: {len(pn_spike_ids)/NUM_NEURONS}")
    print(f"NUMPY: {np.sum(np.random.normal(size=TIMESTEPS*NUM_NEURONS) < 0)/NUM_NEURONS}")

The output is the following:

1:0.004593 2:0.004593 1:0.008632 2:0.008632 1:0.012542 2:0.012542 1:0.016356 2:0.016356 1:0.020096 2:0.020096 1:0.023776 2:0.023776 1:0.027405 2:0.027405 1:0.030989 2:0.030989 1:0.034534 2:0.034534 1:0.038044 2:0.038044
GENN: 0.0
NUMPY: 5.0078125

I was expecting each neuron to spike about 1/2 of the time, as it happens when I try to mimic the behavior in numpy.

I don´t understand:

• Why are the values sampled by the Gaussian distribution strictly increasing;
• And why are the random values in the two neurons printed equal?

What am I doing wrong?

Thank you in advance for the support.

[neworderofjamie]

Apologies about this - definitely a bug. I have fixed in the rng_fix branch which I will merge later this morning.

[andreARibe]

Thank you!