pure event-driven simulation

[Alex Fan]

Hi Spinnaker community,

Does Spinnaker have examples or guide to do pure event-driven simulation? along the line of (Brette 2006). 

Instead of synchronising every 1ms, could it wake up on spike only and use timr difference to analytically update its internal state? That is very well enough for pure lif event-based sim.

For more complicated nonlinear neuron model and non-instantenous synapse, it would need to estimate the next threshold crossing time and queue that event. Some global controller fetch the earliest spiking event to process.

I found https://ieeexplore.ieee.org/document/7169171 seems to attempt doing this, but cannot find more examples.

Cheers,

Alex

[Steve Furber]

Hi Alex,

This is clearly doable on SpiNNaker, but apart from the example you cite below I’m not aware of it being done.

Even with LIF models, pure event-based operation is only straightforward if the synapse uses the voltage step model.

Then all post-synaptic spikes happen at the same time as a pre-synaptic spike, and the closed-form solution of the

LIF equation can be used. If the synaptic model is current-based, such as the double-exponential model, then the

post-synaptic spike will happen some time after the last pre-synaptic spike, so a timer event will have to be used to

generate the post-synaptic spike at the right time. There is no event queuing taking place except to add delays to

pre-syaptic spikes to model the conductance delays between neurons.

As I say, this is all doable, but I’m not sure that we have many examples or guides to help you here.

Best wishes,

—Steve

Steve Furber,                                                     
Professor Emeritus,         http://apt.cs.manchester.ac.uk/people/sfurber/
Department of Computer Science,              mobile:  (+44) 780 169 7951
The University of Manchester, Oxford Road, Manchester M13 9PL, UK.

On 23 Dec 2025, at 17:24, Alex Fan <f1942...@gmail.com> wrote:

Hi Spinnaker community,

Does Spinnaker have examples or guide to do pure event-driven simulation? along the line of (Brette 2006) [pubmed.ncbi.nlm.nih.gov]. 

Instead of synchronising every 1ms, could it wake up on spike only and use timr difference to analytically update its internal state? That is very well enough for pure lif event-based sim.

For more complicated nonlinear neuron model and non-instantenous synapse, it would need to estimate the next threshold crossing time and queue that event. Some global controller fetch the earliest spiking event to process.

I found https://ieeexplore.ieee.org/document/7169171 [ieeexplore.ieee.org] seems to attempt doing this, but cannot find more examples.

Cheers,

Alex

-- 
You received this message because you are subscribed to the Google Groups "SpiNNaker Users Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to spinnakeruser...@googlegroups.com.
To view this discussion, visit https://groups.google.com/d/msgid/spinnakerusers/2955041f-7a7d-4850-b378-320c3e9c4c23n%40googlegroups.com [groups.google.com].

[Andrew Rowley]

Hi,

Yes, I have also heard of people looking at this, but not seen much beyond the paper linked in terms of successful implementation.  As Steve says, all the parts are likely there in the hardware and there will be quite a bit of software support for this, but it would be a case of putting together all the parts to make it really do what you want to.  It is worth also highlighting that the neuron model will affect this quite a bit too, since you would have to do all the neuron updates at the point of evaluation.  I think in the example that you mentioned, they used a simplified model with a linear decay (from what I can see), which of course makes it easy.  In the standard LIF case, decay is exponential, so you might then have to either calculate an exponential when you activate the neuron, or else run the update repeatedly for all the "missed" steps, this latter approach being much less precise and also not likely much better than just doing this on a timestep-by-timestep basis.

Hope that helps somewhat - happy to discuss this more if you want to go through any ideas.

Andrew 🙂

---

From: spinnak...@googlegroups.com <spinnak...@googlegroups.com> on behalf of Steve Furber <steve....@manchester.ac.uk>
Sent: 28 December 2025 14:00
To: Alex Fan <f1942...@gmail.com>
Cc: SpiNNaker Users Group <spinnak...@googlegroups.com>
Subject: Re: [SpiNNaker Mailing List] pure event-driven simulation

 

To view this discussion, visit https://groups.google.com/d/msgid/spinnakerusers/B958F038-EBA0-4E92-8D12-1795FD485643%40manchester.ac.uk [groups.google.com].

[Andrew Rowley]

---

From: ADB <a...@ecs.soton.ac.uk>
Sent: 05 January 2026 10:59
To: f1942...@gmail.com <f1942...@gmail.com>
Cc: Andrew Rowley <Andrew...@manchester.ac.uk>; Steve Furber <steve....@manchester.ac.uk>

Subject: Re: [SpiNNaker Mailing List] pure event-driven simulation

 

Hi Alex

The attached may be of interest....

Specifically in the ANSWER documentation, section 2.3, the type 2 model
does, I think, explicitly what you're talking about.

The performance of ANSWER, set alongside SPiNNaker and BRIAN, is presented
in the IEEE paper; specifically figure 10 might be helpful.

Regards

Andrew Brown

On 05/01/2026 08:00, Andrew Rowley wrote:

CAUTION: This e-mail originated outside the University of Southampton.

To view this discussion, visit https://groups.google.com/d/msgid/spinnakerusers/CWXP265MB4203740D395D5533F66E2CF3C586A%40CWXP265MB4203.GBRP265.PROD.OUTLOOK.COM [groups.google.com].

-- 
------------------------------------
Professor A.D. Brown
Department of Electronics and Computer Science
University of Southampton
SO17 1BJ, Hampshire, UK
T: The University no longer provides staff with telephones
M: +44 (7464) 981171