Voltage increase affected by tau_ref for LIF neuron

General Discussion
1

Hi all,

Comparing the voltages of single neurons, I am looking at how each parameter influences the spiking behaviour. From the documentation (both the parameter describtion as well as the step_math function) I expected that tau_ref would only affect the duration of the refractory period (i.e. the time that the voltage is kept at 0 after a spike.). However, for two neurons that are identical except for their tau_ref value, the neuron with a larger tau_ref reaches the spiking threshold faster (see figure). I do not see why this happens, as I expexted both neurons to reach the spiking threshold at the same time. Can someone explain to me why this is the case, that the potential increases quicker under the same input?

tau_ref
tau_ref809Ă—480 37.8 KB

Code:

import numpy as np
import matplotlib.pyplot as plt
import nengo

model = nengo.Network(label="NEF summary")

def input_func(t):
    if t < .1:
        return 0
    else:
        return 1


with model:

    input = nengo.Node(lambda t:input_func(t))

    ens1 = nengo.Ensemble(
                    1,
                    dimensions=1,
                    intercepts=[0.99],
                    max_rates=[20],
                    encoders=[[1]],
                    label="ens1",
                    neuron_type=nengo.neurons.LIF(tau_ref=0.002),
                    seed=1
    )
    
    ens2 = nengo.Ensemble(
                    1,
                    dimensions=1,
                    intercepts=[0.99],
                    max_rates=[20],
                    encoders=[[1]],
                    neuron_type=nengo.neurons.LIF(tau_ref=0.01),
                    label="ens2",
                    seed=1
    )

    nengo.Connection(input, ens1, synapse=0)
    nengo.Connection(input, ens2, synapse=0)


    input_probe = nengo.Probe(input)
    voltage_ens1 = nengo.Probe(ens1.neurons, "voltage")
    voltage_ens2 = nengo.Probe(ens2.neurons, "voltage")

sim= nengo.Simulator(model)

sim.run(0.3)

t = sim.trange()

data_input = sim.data[input_probe]
data_ens1 = sim.data[voltage_ens1]
data_ens2 = sim.data[voltage_ens2]

# print("data_ens1:")
# print(data_ens1)
# print("data_ens2:")
# print(data_ens2)

plt.plot(t, data_ens1, label="ens1 (tau_ref=0.002)")
plt.plot(t, data_ens2, label="ens2 (tau_ref=0.01)")
plt.plot(t, data_input, label="input")
plt.xlim(right=t[-1])
plt.ylim(-0.1,1.1)
plt.legend()
plt.show()
2

It’s because you’re setting max_rates and intercepts. These determine the gain/bias transform based on the neuron’s tuning curve, such that the neuron has the desired intercept, and achieves the desired max rate for an input value of 1. Since your neurons have different tuning curves, you’ll get different gain/bias values for each ensemble, resulting in different behaviour beyond the change in refractory period. As you can see, both of your neurons are firing at the same rate of 20 Hz, which makes sense as this is what you set as the desired max rate for each neuron (and giving an input of 1 will drive the neurons at their max rate).

If you set gain and bias directly instead, then this will ensure that both neurons have the same input transform.

To get good values, you can use the NeuronType.gain_bias function:

neuron_type = nengo.LIF()
gain, bias = neuron_type.gain_bias(max_rates=20, intercepts=0.5)

Just make sure to only do this once and use the same values for both ensembles. Also note that I’m using an intercept of 0.5, since 0.99 is quite a high intercept and results in quite large gain and bias values.

1 Like
3

As a quick demonstration, here’s your code modified to set the gains and biases directly (instead of setting the max rates and intercepts):

import matplotlib.pyplot as plt
import nengo
import numpy as np

model = nengo.Network(label="NEF summary")


def input_func(t):
    if t < 0.1:
        return 0
    else:
        return 1


with model:

    input = nengo.Node(lambda t: input_func(t))

    ens1 = nengo.Ensemble(
        1,
        dimensions=1,
        gain=[15.6],
        bias=[-14.5],
        encoders=[[1]],
        label="ens1",
        neuron_type=nengo.neurons.LIF(tau_ref=0.002),
        seed=1,
    )

    ens2 = nengo.Ensemble(
        1,
        dimensions=1,
        gain=[15.6],
        bias=[-14.5],
        encoders=[[1]],
        neuron_type=nengo.neurons.LIF(tau_ref=0.01),
        label="ens2",
        seed=1,
    )

    nengo.Connection(input, ens1, synapse=0)
    nengo.Connection(input, ens2, synapse=0)

    input_probe = nengo.Probe(input)
    voltage_ens1 = nengo.Probe(ens1.neurons, "voltage")
    voltage_ens2 = nengo.Probe(ens2.neurons, "voltage")

with nengo.Simulator(model) as sim:
    sim.run(0.3)

t = sim.trange()

data_input = sim.data[input_probe]
data_ens1 = sim.data[voltage_ens1]
data_ens2 = sim.data[voltage_ens2]

# print("data_ens1:")
# print(data_ens1)
# print("data_ens2:")
# print(data_ens2)

plt.plot(t, data_ens1, label="ens1 (tau_ref=0.002)")
plt.plot(t, data_ens2, label="ens2 (tau_ref=0.01)")
plt.plot(t, data_input, label="input")
plt.xlim(right=t[-1])
plt.ylim(-0.1, 1.1)
plt.legend()
plt.show()

And if you plot the voltages, you’ll see that setting the gains and biases directly lead to both neurons having the same responses, with the only difference being the tau_ref.

image
image728Ă—532 71.6 KB

As an aside, here is where Nengo determines what the gain and bias of the neurons should be, and you’ll notice that it takes the tau_ref value into account when doing this computation.

1 Like