Ablating neurons in a SPA model

The model used here is a SPA model that goes through a set sequence of semantic pointers, as shown in this NengoSPA example.

This is primarily a demonstration of how to apply the ablate_ensemble function to all the ensembles in a network or SPA module. If you haven’t gone through the ablate_ensemble example in this directory, go through that example first.

[1]:
%matplotlib inline
import numpy as np
import matplotlib.pyplot as plt
import nengo
from nengo import spa
[2]:
dimensions = 16


def start(t):
    if t < 0.05:
        return "A"
    else:
        return "0"


with spa.SPA() as model:
    model.cortex = spa.Buffer(dimensions=dimensions)
    actions = spa.Actions(
        "dot(cortex, A) --> cortex = B",
        "dot(cortex, B) --> cortex = C",
        "dot(cortex, C) --> cortex = D",
        "dot(cortex, D) --> cortex = E",
        "dot(cortex, E) --> cortex = A",
    )
    model.bg = spa.BasalGanglia(actions=actions)
    model.thal = spa.Thalamus(model.bg)
    model.input = spa.Input(cortex=start)

    cortex = nengo.Probe(model.cortex.state.output, synapse=0.01)
    actions = nengo.Probe(model.thal.actions.output, synapse=0.01)
    utility = nengo.Probe(model.bg.input, synapse=0.01)
[3]:
with nengo.Simulator(model) as sim:
    sim.run(0.5)
0%
 
/home/tbekolay/Code/nengo/nengo/builder/builder.py:234: UserWarning: Object <BasalGanglia "bg"> has already been built.
  warnings.warn(f"Object {obj} has already been built.")
/home/tbekolay/Code/nengo/nengo/builder/builder.py:234: UserWarning: Object <Thalamus "thal"> has already been built.
  warnings.warn(f"Object {obj} has already been built.")
0%
 
[4]:
def plot():
    fig = plt.figure(figsize=(12, 8))
    p1 = fig.add_subplot(3, 1, 1)

    p1.plot(sim.trange(), model.similarity(sim.data, cortex))
    p1.legend(model.get_output_vocab("cortex").keys, fontsize="x-small")
    p1.set_ylabel("State")

    p2 = fig.add_subplot(3, 1, 2)
    p2.plot(sim.trange(), sim.data[actions])
    p2_legend_txt = [a.effect for a in model.bg.actions.actions]
    p2.legend(p2_legend_txt, fontsize="x-small")
    p2.set_ylabel("Action")

    p3 = fig.add_subplot(3, 1, 3)
    p3.plot(sim.trange(), sim.data[utility])
    p3_legend_txt = [a.condition for a in model.bg.actions.actions]
    p3.legend(p3_legend_txt, fontsize="x-small")
    p3.set_ylabel("Utility")

    fig.subplots_adjust(hspace=0.2)


plot()
../_images/ablate_ablate-spa_4_0.png

We will use the same ablate_ensemble function as before.

[5]:
def ablate_ensemble(ens, proportion, sim, bias=True):
    """Ablate a proportion of the neurons in an ensemble.

    The ablation is done by setting the encoder and gain associated
    with a neuron to zero. Since no input current being injected,
    the neuron will generally be silent. However, if there is direct
    current injected with a neuron-to-neuron connection, then the
    cell may still fire. To counter that in most cases, we set the
    bias associated with the neuron to a large negative value.
    """

    n_neurons = min(int(ens.n_neurons * proportion), ens.n_neurons)
    idx = np.random.choice(np.arange(ens.n_neurons), replace=False, size=n_neurons)

    encoder_sig = sim.signals[sim.model.sig[ens]["encoders"]]
    encoder_sig.setflags(write=True)
    encoder_sig[idx] = 0.0
    encoder_sig.setflags(write=False)

    if bias:
        bias_sig = sim.signals[sim.model.sig[ens.neurons]["bias"]]
        bias_sig.setflags(write=True)
        bias_sig[idx] = -1000

This time, however, instead of applying it to a single ensemble, we will apply it to all of the ensembles in a SPA module.

We can therefore see the effects of ablation in different parts of the model. Let’s see what happens when ablating 25% of the neurons in the cortex.

[6]:
with nengo.Simulator(model) as sim:
    for ens in model.cortex.all_ensembles:
        ablate_ensemble(ens, 0.25, sim)
    sim.run(0.5)
plot()
0%
 
/home/tbekolay/Code/nengo/nengo/builder/builder.py:234: UserWarning: Object <BasalGanglia "bg"> has already been built.
  warnings.warn(f"Object {obj} has already been built.")
/home/tbekolay/Code/nengo/nengo/builder/builder.py:234: UserWarning: Object <Thalamus "thal"> has already been built.
  warnings.warn(f"Object {obj} has already been built.")
0%
 
../_images/ablate_ablate-spa_8_5.png

In some cases, the sequence stops after some time, which is interesting!

It saves us a tiny bit of typing to make a helper function to ablate whole networks (or SPA modules). So let’s do that, then look at the effects of ablating the other parts of the model.

[7]:
def ablate_network(net, proportion, sim, bias=True):
    """Ablate a proportion of the neurons in all ensembles in a network."""
    for ens in net.all_ensembles:
        ablate_ensemble(ens, proportion, sim, bias=bias)
[8]:
with nengo.Simulator(model) as sim:
    ablate_network(model.bg, 0.25, sim)
    sim.run(0.5)
plot()
0%
 
/home/tbekolay/Code/nengo/nengo/builder/builder.py:234: UserWarning: Object <BasalGanglia "bg"> has already been built.
  warnings.warn(f"Object {obj} has already been built.")
/home/tbekolay/Code/nengo/nengo/builder/builder.py:234: UserWarning: Object <Thalamus "thal"> has already been built.
  warnings.warn(f"Object {obj} has already been built.")
0%
 
../_images/ablate_ablate-spa_11_5.png
[9]:
with nengo.Simulator(model) as sim:
    ablate_network(model.thal, 0.25, sim)
    sim.run(0.5)
plot()
0%
 
/home/tbekolay/Code/nengo/nengo/builder/builder.py:234: UserWarning: Object <BasalGanglia "bg"> has already been built.
  warnings.warn(f"Object {obj} has already been built.")
/home/tbekolay/Code/nengo/nengo/builder/builder.py:234: UserWarning: Object <Thalamus "thal"> has already been built.
  warnings.warn(f"Object {obj} has already been built.")
0%
 
../_images/ablate_ablate-spa_12_5.png

Finally, let’s look at what happens when we ablate 25% of the neurons in the entire model.

[10]:
with nengo.Simulator(model) as sim:
    ablate_network(model, 0.25, sim)
    sim.run(0.5)
plot()
0%
 
/home/tbekolay/Code/nengo/nengo/builder/builder.py:234: UserWarning: Object <BasalGanglia "bg"> has already been built.
  warnings.warn(f"Object {obj} has already been built.")
/home/tbekolay/Code/nengo/nengo/builder/builder.py:234: UserWarning: Object <Thalamus "thal"> has already been built.
  warnings.warn(f"Object {obj} has already been built.")
0%
 
../_images/ablate_ablate-spa_14_5.png

Not silencing the neurons seems to hurt performance, possibly due to the added noise of the random spikes.

[11]:
with nengo.Simulator(model) as sim:
    ablate_network(model, 0.25, sim, bias=False)
    sim.run(0.5)
plot()
0%
 
/home/tbekolay/Code/nengo/nengo/builder/builder.py:234: UserWarning: Object <BasalGanglia "bg"> has already been built.
  warnings.warn(f"Object {obj} has already been built.")
/home/tbekolay/Code/nengo/nengo/builder/builder.py:234: UserWarning: Object <Thalamus "thal"> has already been built.
  warnings.warn(f"Object {obj} has already been built.")
0%
 
../_images/ablate_ablate-spa_16_5.png

To verify that this is indeed working as expected, we can ablate all of the neurons and confirm that there is no activity in the model.

[12]:
with nengo.Simulator(model) as sim:
    ablate_network(model, 1, sim)
    sim.run(0.5)
plot()
0%
 
/home/tbekolay/Code/nengo/nengo/builder/builder.py:234: UserWarning: Object <BasalGanglia "bg"> has already been built.
  warnings.warn(f"Object {obj} has already been built.")
/home/tbekolay/Code/nengo/nengo/builder/builder.py:234: UserWarning: Object <Thalamus "thal"> has already been built.
  warnings.warn(f"Object {obj} has already been built.")
0%
 
../_images/ablate_ablate-spa_18_5.png