Computing squared length from spa.State with feedback

I have a network that is behaving in a weird way and I’m not sure what I am missing (or if it is a bug). I am trying to compute the squared length of a spa.State (acting as memory with a feedback) by adding an additional output to the internal ensemble array to compute the squares. These are then projected into a single ensemble. For some reason the result is considerably larger than expected. If I probe the output of that spa.State and do the calculations with NumPy, I get the expected result, so the represented value is correct.

If I feed the output of the State into another State and calculate the squared length in the same way (adding squared output, projecting to 1d ensemble), it works.

Here is some code:

import matplotlib.pyplot as plt
import nengo
from nengo import spa

d = 128
with spa.SPA() as model:
    model.diff = spa.State(d)
    model.state = spa.State(d, subdimensions=1, feedback=1.)
    model.copy = spa.State(d, subdimensions=1)
    nengo.Connection(model.diff.output, model.state.input, transform=10.)
    nengo.Connection(model.state.output, model.diff.input, transform=-1.)
    nengo.Connection(model.state.output, model.copy.input)
    state_sq = model.state.state_ensembles.add_output('sq', np.square)
    copy_sq = model.copy.state_ensembles.add_output('sq', np.square)
    state_sq_length = nengo.Ensemble(150, 1)
    copy_sq_length = nengo.Ensemble(150, 1)
    nengo.Connection(state_sq, state_sq_length, transform=np.ones((1, d)))
    nengo.Connection(copy_sq, copy_sq_length, transform=np.ones((1, d)))
    model.stim = spa.Input(diff='A')
    p_state = nengo.Probe(model.state.output, synapse=0.01)
    p_copy = nengo.Probe(model.copy.output, synapse=0.01)
    p_state_sq_length = nengo.Probe(state_sq_length, synapse=0.01)
    p_copy_sq_length = nengo.Probe(copy_sq_length, synapse=0.01)

with nengo.Simulator(model) as sim:

plt.plot(sim.trange(),[p_state_sq_length], label='state_sq_length')
plt.plot(sim.trange(), np.sum(np.square([p_state]), axis=1), label='state')
plt.plot(sim.trange(),[p_copy_sq_length], label='copy_sq_length')
plt.plot(sim.trange(), np.sum(np.square([p_copy]), axis=1), label='copy')
plt.xlabel('Time (s)')
plt.ylabel('Squared length')

Output plot:

The blue line is the initial attempt that is way larger than the analytical calculation from the probed output (green line). By feeding the output into another State neural and analytical result (red and purple line) agree.

Can anyone explain to me why this is happening?

Not a bug. It’s because of the recurrent connection you have between diff and state. The recurrent connection within state is 0.1, but the connection between diff and state is 0.005 (default). This results in a very high frequency oscillation in the output of the state buffer. When you compute the squared length directly on that output, you get an amplification of the noise (so you get a bigger result).

The reason why you don’t get the amplified results when you a) calculate it with numpy or b) use the copy ensemble is because of the extra filtering that takes place on the output of state. To fix, simply:

nengo.Connection(model.diff.output, model.state.input, transform=10., synapse=0.1)

(i.e. match the synapse on any inputs to state with the synapse on the recurrent connection in state)


On an unrelated note, this ‘bug’ (I couldn’t up the gain – i.e. the transform=10 – high enough without causing the high frequency oscillations) was what caused me to abandon using this approach (i.e. a difference based memory) for the first version of Spaun. But I’ve fixed it now (see:, so it’s back in use for Spaun 2.0! :smiley:

1 Like