Hello everyone!
I am writing to you regarding two matters:
Reset Network/Kernel in nest2->nest 3 In the last developer conference, Daphne Cornelisse talked about that she used ResetNetwork() to solve her problem. ResetNetwork() is marked as deprecated. No one said anything, so I got confused why this is apparently the recommended way or at least approved. She showed me, that it works (in her case). It is deprecated, so there is probably some good reasoning behind it. The documentation says: "ResetNetwork is deprecated and will be removed in NEST 3, because this function is not fully able to reset network and simulator state. What are the edge cases where the use causes problems?
In this ticket, it is stated that the feature is just removed with any replacement. https://github.com/nest/nest-simulator/issues/525
Thus, in nest3 there is only ResetKernel(). This means that you have to rebuild the network for any application where you do multiple simulations with different input or parameter changes. I am using nest3 for reinforcement learning and in each training episode, I have to extract all the weights and save them, reset the kernel, reconstruct the net, then load all the weights. This adds a lot of overhead in performance and bloats my code. I basically have another front-end storing the net and talking to the nest back-end. Therefore, the update to nest3 is a downgrade for many applications. I don’t have a solution for this issue, but I want to spark some discussion as I learned that I am not the only nest user to stumble into this issue.
STDP Performance boost by manual computation in python In the paper "Demonstrating Advantages of Neuromorphic Computation: A Pilot Study“ by Wunderlich et al. (https://www.frontiersin.org/articles/10.3389/fnins.2019.00260/full) some performance improvement on STDP was reported.
"The synaptic weight updates in each iteration were restricted to those synapses which transmitted spikes, i.e., the synapses from the active input unit to all output units (32 out of the 1,024 synapses), as the correlation a+ of all other synapses is zero in a perfect simulation without fixed-pattern noise. This has the effect of reducing the overall time required to simulate one iteration[…]“
The provided source code (https://github.com/electronicvisions/model-sw-pong/blob/976e0778ca05cfd96c4c... https://github.com/electronicvisions/model-sw-pong/blob/976e0778ca05cfd96c4c1f5abf3b7a352b87cb72/pang.py#L231) indeed contains a manual computation of STDP. When using the nest library I don’t expect to do some computation in python to be faster. It appears to me that the nest implementation is computing STDP every time, even without spikes? Maybe someone can comment on this whether this can be improved in nest?
Kind regards, Benedikt S. Vogler
-- Benedikt S. Vogler benedikt.s.vogler@tum.de Student M.Sc. Robotics, Cognition, Intelligence
Dear Benedict,
Regarding your second question, from a quick glance at that paper it seems the authors do not state explicitly that their computation is faster than NEST. They probably mean that computing the weight change only when a presynaptic spike arrives is faster than evaluating their update code at each time step (which is not what NEST does, but what a straightforward simulation of the authors' hardware would do: see the line "in contrast to the implementation on BSS2"). As for why the authors chose to implement plasticity outside of NEST rather than pick some existing model of reward-modulated STDP, this is not explained in the paper, but one might suppose that the reason may have been to ensure that the plasticity is identical to their hardware setup.
Anyway, synaptic weight change computation in NEST is also triggered by a presynaptic spike (just because we only need to know the up-to-date weight value when a postsynaptic spike has to be transmitted). So, there does not seem to be anything that needs improvement in this matter. Please correct me if I get something wrong.
Sincerely yours, Alex Serenko Graduate student at Kurchatov Institute, Moscow, Russia
вт, 12 мая 2020 г. в 16:07, Benedikt S. Vogler benedikt.s.vogler@tum.de:
Dear Benedict,
When it comes to ResetNetwork(), there was never really a definite definition of what the function should do. What resetting the network truly means. Does it mean resetting all model parameters, all synapse parameters, both? Do we reset to default values, or to set initial values? The function was never complete, and the use of it could lead to strange results, because some parameters were reset while others were not.
To try to illustrate the complexity and what we had problems with when defining what reset network means, image you initialize your neurons with random membrane potential. Then you simulate for a while before resetting your network. Are we supposed to go back to the initial random potentials, or should you then get the default membrane potential for the model? Maybe it makes most sense that we get the initial random values. Then we need to store them for all neurons in the network, in case someone does a ResetNetwork(), so that we can apply them again. This can be quite complex, because you might set a lot of variables, and you might have a very large network.
It would be very helpful if you create an issue where you write exactly what you need to have reset for your application. What you envision a ResetNetwork to do. We might be able to restore some of the functionalities to make it easier for you.
Best regards, Stine Vennemo
________________________________ Fra: Alexey Serenko serenko@phystech.edu Sendt: 12. mai 2020 16:12 Til: NEST User Mailing List; benedikt.s.vogler@tum.de Emne: [NEST Users] Re: Reset Kernel + STDP Performance
Dear Benedict,
Regarding your second question, from a quick glance at that paper it seems the authors do not state explicitly that their computation is faster than NEST. They probably mean that computing the weight change only when a presynaptic spike arrives is faster than evaluating their update code at each time step (which is not what NEST does, but what a straightforward simulation of the authors' hardware would do: see the line "in contrast to the implementation on BSS2"). As for why the authors chose to implement plasticity outside of NEST rather than pick some existing model of reward-modulated STDP, this is not explained in the paper, but one might suppose that the reason may have been to ensure that the plasticity is identical to their hardware setup.
Anyway, synaptic weight change computation in NEST is also triggered by a presynaptic spike (just because we only need to know the up-to-date weight value when a postsynaptic spike has to be transmitted). So, there does not seem to be anything that needs improvement in this matter. Please correct me if I get something wrong.
Sincerely yours, Alex Serenko Graduate student at Kurchatov Institute, Moscow, Russia
вт, 12 мая 2020 г. в 16:07, Benedikt S. Vogler <benedikt.s.vogler@tum.demailto:benedikt.s.vogler@tum.de>: Hello everyone!
I am writing to you regarding two matters:
Reset Network/Kernel in nest2->nest 3 In the last developer conference, Daphne Cornelisse talked about that she used ResetNetwork() to solve her problem. ResetNetwork() is marked as deprecated. No one said anything, so I got confused why this is apparently the recommended way or at least approved. She showed me, that it works (in her case). It is deprecated, so there is probably some good reasoning behind it. The documentation says: "ResetNetwork is deprecated and will be removed in NEST 3, because this function is not fully able to reset network and simulator state. What are the edge cases where the use causes problems?
In this ticket, it is stated that the feature is just removed with any replacement. https://github.com/nest/nest-simulator/issues/525
Thus, in nest3 there is only ResetKernel(). This means that you have to rebuild the network for any application where you do multiple simulations with different input or parameter changes. I am using nest3 for reinforcement learning and in each training episode, I have to extract all the weights and save them, reset the kernel, reconstruct the net, then load all the weights. This adds a lot of overhead in performance and bloats my code. I basically have another front-end storing the net and talking to the nest back-end. Therefore, the update to nest3 is a downgrade for many applications. I don’t have a solution for this issue, but I want to spark some discussion as I learned that I am not the only nest user to stumble into this issue.
STDP Performance boost by manual computation in python In the paper "Demonstrating Advantages of Neuromorphic Computation: A Pilot Study“ by Wunderlich et al. (https://www.frontiersin.org/articles/10.3389/fnins.2019.00260/full) some performance improvement on STDP was reported.
"The synaptic weight updates in each iteration were restricted to those synapses which transmitted spikes, i.e., the synapses from the active input unit to all output units (32 out of the 1,024 synapses), as the correlation a+ of all other synapses is zero in a perfect simulation without fixed-pattern noise. This has the effect of reducing the overall time required to simulate one iteration[…]“
The provided source code (https://github.com/electronicvisions/model-sw-pong/blob/976e0778ca05cfd96c4c...) indeed contains a manual computation of STDP. When using the nest library I don’t expect to do some computation in python to be faster. It appears to me that the nest implementation is computing STDP every time, even without spikes? Maybe someone can comment on this whether this can be improved in nest?
Kind regards, Benedikt S. Vogler
-- Benedikt S. Vogler benedikt.s.vogler@tum.demailto:benedikt.s.vogler@tum.de Student M.Sc. Robotics, Cognition, Intelligence
_______________________________________________ NEST Users mailing list -- users@nest-simulator.orgmailto:users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.orgmailto:users-leave@nest-simulator.org
Dear Stine!
Thank you for your answer. We posted a proposal regarding the reset network functionality https://github.com/nest/nest-simulator/issues/1618 on the github issues.
Kind regards, Benedikt S. Vogler
-
Alexey Serenko -
Benedikt S. Vogler -
Stine Brekke Vennemo