search

SpiNNakerManchester / sPyNNaker

The SpiNNaker implementation of the PyNN neural networking language
Apache License 2.0
105 stars 44 forks source link

Ring buffer sigma and spikes-per-second on a per synapse-type basis #136

Open alan-stokes opened 9 years ago

alan-stokes commented 9 years ago

different synapse types have different statistics and separate ring-buffer shifts to scale them with, so ideally, you should also be able to specify ring buffer sigma and spikes-per-second separately for different synapse types.

linked to #132

hopper333 commented 9 years ago

The ring-buffer scaling is a function of connectivity and expected input firing rates in a particular network configuration, so pretty well independent of synapse type.

M

Michael Hopkins, SpiNNaker project, APT group, School of Computer Science, University of Manchester, Manchester M13 9PL michael.hopkins@manchester.ac.ukmailto:simon.davidson@manchester.ac.uk

On 25 Aug 2015, at 10:11, Alan Stokes notifications@github.com<mailto:notifications@github.com> wrote:

different synapse types have different statistics and separate ring-buffer shifts to scale them with, so ideally, you should also be able to specify ring buffer sigma and spikes-per-second separately for different synapse types.

linked to #132https://github.com/SpiNNakerManchester/sPyNNaker/issues/132

— Reply to this email directly or view it on GitHubhttps://github.com/SpiNNakerManchester/sPyNNaker/issues/136.

rowleya commented 9 years ago

You can do the scaling per synapse type i.e. if you know the firing rate of all the populations exciting this population and a different firing rate for all the populations inhibiting this population, you can do two separate calculations. We already do the separate calculations but we use a firing rate and sigma value which is fixed over the entire simulation. If you know more about your network, the ability to change this per population and per synapse type could help you to get better overall results; how much is to be seen, but it isn’t too hard to enable this (Alan has already done it per-population, per synapse-type is a bit more fiddly but doable).

From: Michael Hopkins [mailto:notifications@github.com] Sent: 25 August 2015 11:45 To: SpiNNakerManchester/sPyNNaker sPyNNaker@noreply.github.com Subject: Re: [sPyNNaker] Ring buffer sigma and spikes-per-second on a per synapse-type basis (#136)

The ring-buffer scaling is a function of connectivity and expected input firing rates in a particular network configuration, so pretty well independent of synapse type.

M

Michael Hopkins, SpiNNaker project, APT group, School of Computer Science, University of Manchester, Manchester M13 9PL michael.hopkins@manchester.ac.ukmailto:simon.davidson@manchester.ac.ukmailto:michael.hopkins@manchester.ac.uk%3cmailto:simon.davidson@manchester.ac.uk

On 25 Aug 2015, at 10:11, Alan Stokes notifications@github.com<mailto:notifications@github.com<mailto:notifications@github.com%3cmailto:notifications@github.com>> wrote:

different synapse types have different statistics and separate ring-buffer shifts to scale them with, so ideally, you should also be able to specify ring buffer sigma and spikes-per-second separately for different synapse types.

linked to #132https://github.com/SpiNNakerManchester/sPyNNaker/issues/132

— Reply to this email directly or view it on GitHubhttps://github.com/SpiNNakerManchester/sPyNNaker/issues/136.

— Reply to this email directly or view it on GitHubhttps://github.com/SpiNNakerManchester/sPyNNaker/issues/136#issuecomment-134552799.

Christian-B commented 3 years ago

Is this still valid?

rowleya commented 3 years ago

I still don't think it is possible to do this. I don't know where it is relevant to do so.

hopper333 commented 3 years ago

Hi guys

I’m not sure what the status is with this now. Have we moved over to the atomic weight multiplied up as appropriate which means that the upper bound idea is not used in the same way now?

M

Michael Hopkins (Research Fellow)

SpiNNaker project, APT group, Department of Computer Science, The University of Manchester, Oxford Road, Manchester M13 9PL, UK @.**@.>

From: Andrew Rowley @.> Reply-To: SpiNNakerManchester/sPyNNaker @.> Date: Monday, 22 March 2021 at 15:18 To: SpiNNakerManchester/sPyNNaker @.> Cc: Michael Hopkins @.>, Comment @.***> Subject: Re: [SpiNNakerManchester/sPyNNaker] Ring buffer sigma and spikes-per-second on a per synapse-type basis (#136)

I still don't think it is possible to do this. I don't know where it is relevant to do so.

— You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/SpiNNakerManchester/sPyNNaker/issues/136#issuecomment-804144257, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ACLDOZF2TEHR5E3NQ2BCPQTTE5NS7ANCNFSM4BOGOHRA.

rowleya commented 3 years ago

We haven't completely moved to that yet, mainly due to issues with working out what the "atomic" weight will be! I think we will end up needing a combination of techniques e.g. maximum weight calculation, then minimum weight = max / 65536, and then also add in stochastic rounding to do things smaller than min weight. It will take a bit of effort to get these things done though!

hopper333 commented 3 years ago

OK thanks Andrew. The joys of fixed-point maths.

M

Michael Hopkins (Research Fellow)

SpiNNaker project, APT group, Department of Computer Science, The University of Manchester, Oxford Road, Manchester M13 9PL, UK @.**@.>

From: Andrew Rowley @.> Reply-To: SpiNNakerManchester/sPyNNaker @.> Date: Monday, 22 March 2021 at 15:52 To: SpiNNakerManchester/sPyNNaker @.> Cc: Michael Hopkins @.>, Comment @.***> Subject: Re: [SpiNNakerManchester/sPyNNaker] Ring buffer sigma and spikes-per-second on a per synapse-type basis (#136)

We haven't completely moved to that yet, mainly due to issues with working out what the "atomic" weight will be! I think we will end up needing a combination of techniques e.g. maximum weight calculation, then minimum weight = max / 65536, and then also add in stochastic rounding to do things smaller than min weight. It will take a bit of effort to get these things done though!

— You are receiving this because you commented. Reply to this email directly, view it on GitHubhttps://github.com/SpiNNakerManchester/sPyNNaker/issues/136#issuecomment-804177636, or unsubscribehttps://github.com/notifications/unsubscribe-auth/ACLDOZFFU5IQPFXEISRWVALTE5RS7ANCNFSM4BOGOHRA.

andrewgait commented 1 year ago

I would think this can be closed if either https://github.com/SpiNNakerManchester/sPyNNaker/pull/1386 or https://github.com/SpiNNakerManchester/sPyNNaker/pull/852 are merged in the future.