Reproducing Sinusoid oscillator of Supervised learning in spiking neural networks with FORCE training

[schmitts]

No description provided.

[mstimberg]

Many thanks (again 😊)! I wonder whether we cannot replace the two network operations with a "creative use" of synapses – this might be a nice showcase for doing this and would make everything more efficient (and enable us to use standalone mode!). It will make things quite a bit less readable, though, not sure...

[schmitts]

Yes, I thought the same thing. I was not really happy with the explicit matrix operations. On the other hand it makes the math more accessible/direct. Do you want to draft a version with Synapse operations? It's easier to discuss/compare with an example at hand.

[mstimberg]

I have hacked together a solution that only uses Synapses, summed variables, and run_regularly operations, but… it isn't pretty. In particular, we don't have a nice syntax to set the timing of several summed variable operations, but here they need to be executed in the right order to faithfully reproduce the algorithm in your network_operation. Actually, the result doesn't seem to be exactly the same, so there might be a bug as well – but either way, I don't think this should go into an example like this. You can compare the two implementations here: schmitts/brian2@Nicola_Clopath_2017...brian-team:brian2:Nicola_Clopath_2017_no_network_op

The best solution is probably to go with your implementation for now, but also think about how we can improve syntax to make these things more straightforward in the future. For example, I had one idea that would work quite nicely as an extension of our current syntax, I think: we could extend (summed) to NeuronGroups, to make it possible to sum expressions into a shared variable. This would allow users to replace this

# linear readout
@network_operation(dt=defaultclock.dt)
def readout(t):
    neurons.z = np.dot(neurons.BPhi, neurons.r)

by:

"""...
z = BPhi * r : 1 (summed, shared)
"""

in the equations.

[schmitts]

Hi!

Thanks for your effort! Did you do some benchmarking between the two implementations?

Allowing for

"""...
z = BPhi * r : 1 (summed, shared)
"""

would be cool!

Cheers,

Sebastian

[mstimberg]

Did you do some benchmarking between the two implementations?

It was faster, but not dramatically so. Instead of 1m20s, it ran in something like 42s. With C++ standalone mode, it went further down to 22s. But I think the network_operation can be slightly sped up as well, I'll try out a few things.

[mstimberg]

I tried a few things and nothing makes it significantly faster. In general, I'd recommend to use e.g. neurons.v[:] instead of neurons.v when passing Brian variable to functions, since neurons.v[index] adds quite a bit of overhead to support string-based indexing, etc. In this case, the numpy functions deal with this in a decent way, and it only shaved off a second or two here and there. For simplicity, I'll therefore leave it as it is and go ahead with the merge.

We can use this example as a use case for the syntax I mentioned earlier (more general discussion: #180).

[mstimberg]

Thanks again 🙏

[mstimberg]

And here it is: https://brian2.readthedocs.io/en/latest/examples/frompapers.Nicola_Clopath_2017.html