`tsodyks_synapse` and `tsodyks2_synapse` models have different behavior for short term depression
Hello,
I have followed the example script comparing `tsodyks_synapse` and `tsodyks2_synapse` models, which can be found here: https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody...
Using the parameter set for depression instead of facilitation I found that `tsodyks2_synapse` model leads to weird peaks of synaptic currents (at least I would not expect that for short term depression) - during the second stimulation the weight rescaling seems to be shifted by one spike compared to tsodyks_synapse (as if parameter "u" would be updated after rescaling the weight instead of before).
Does anybody know why is there this difference between the two synapse models?
Thank you.
All the best, PH
Dear Pavel,
I can confirm you observation, and I share your surprise. Intuitively, one would expect that the synaptic state should return to the initial state after the long period of silence. However, after consulting the source papers (todyks_synapse: doi:10.1523/JNEUROSCI.20-01-j0003.2000, tsodyks2_synapse: doi:10.1016/S0893-6080(01)00144-7) and looking at the NEST code, I conclude that what the example shows is indeed the intended behaviour.
A key difference between tsodyks_synapse and todyks2_synapse is the order of update steps: upon arrival of the presynaptic spike, tsodyks_synapse first updates the synapse state (x,y,z,u; see lines 258-277 in tsodyks_synapse.h), and then reports the new weight (line 285 in tsodyks_synapse.h). In tsodyks2_synapse, this is reversed: first, the synaptic weight is computed based on the current (i.e., the old) synapse state (line 236 in tsodyks2_synapse.h), and then the synapse state is updated (lines 243,244 in tsodyks2_synapse.h). When the first spike of the second burst of presynaptic spikes arrives in the example (https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody...), the recovery value of tsodyks2_synapse still is where it was at the end of the first burst of spikes (before the break). This is exactly how the model is defined in (Maass and Markram, 2002; doi:10.1016/S0893-6080(01)00144-7; see equations (1-5)). It needs one more spike to really "recover".
I hope this clarifies things a bit.
Best Tom
On 12.12.24 12:05, Pavel Haman wrote:
Hello,
I have followed the example script comparing `tsodyks_synapse` and `tsodyks2_synapse` models, which can be found here: https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody...
Using the parameter set for depression instead of facilitation I found that `tsodyks2_synapse` model leads to weird peaks of synaptic currents (at least I would not expect that for short term depression) - during the second stimulation the weight rescaling seems to be shifted by one spike compared to tsodyks_synapse (as if parameter "u" would be updated after rescaling the weight instead of before).
Does anybody know why is there this difference between the two synapse models?
Thank you.
All the best, PH _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hi Tom,
Thanks for your input on this! I was just looking at the code too. It seems that the order of updates (first update adaptation variables, then send spike, versus first send spike, then update adaptation variables) was changed, in order to prevent the adaptation from kicking in for the very first spike to ever go through the synapse. See https://github.com/nest/nest-simulator/pull/1672/.) Notice that t_lastspike is initialised to zero, rather than some "special" value like -1. It is possible that this fix was erroneous, in that it ends up affecting not just the first-ever spike, but also subsequent spikes. I am investigating a bit more, but if you are convinced that the current behaviour is the correct one, then I can stop investigating.
Much obliged! Charl
On Mon, Dec 16, 2024, at 11:01, Tom Tetzlaff wrote:
Dear Pavel,
I can confirm you observation, and I share your surprise. Intuitively, one would expect that the synaptic state should return to the initial state after the long period of silence. However, after consulting the source papers (todyks_synapse: doi:10.1523/JNEUROSCI.20-01-j0003.2000, tsodyks2_synapse: doi:10.1016/S0893-6080(01)00144-7) and looking at the NEST code, I conclude that what the example shows is indeed the intended behaviour.
A key difference between tsodyks_synapse and todyks2_synapse is the order of update steps: upon arrival of the presynaptic spike, tsodyks_synapse first updates the synapse state (x,y,z,u; see lines 258-277 in tsodyks_synapse.h), and then reports the new weight (line 285 in tsodyks_synapse.h). In tsodyks2_synapse, this is reversed: first, the synaptic weight is computed based on the current (i.e., the old) synapse state (line 236 in tsodyks2_synapse.h), and then the synapse state is updated (lines 243,244 in tsodyks2_synapse.h). When the first spike of the second burst of presynaptic spikes arrives in the example (https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody...), the recovery value of tsodyks2_synapse still is where it was at the end of the first burst of spikes (before the break). This is exactly how the model is defined in (Maass and Markram, 2002; doi:10.1016/S0893-6080(01)00144-7; see equations (1-5)). It needs one more spike to really "recover".
I hope this clarifies things a bit.
Best Tom
On 12.12.24 12:05, Pavel Haman wrote:
Hello,
I have followed the example script comparing `tsodyks_synapse` and `tsodyks2_synapse` models, which can be found here: https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody...
Using the parameter set for depression instead of facilitation I found that `tsodyks2_synapse` model leads to weird peaks of synaptic currents (at least I would not expect that for short term depression) - during the second stimulation the weight rescaling seems to be shifted by one spike compared to tsodyks_synapse (as if parameter "u" would be updated after rescaling the weight instead of before).
Does anybody know why is there this difference between the two synapse models?
Thank you.
All the best, PH _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
--
Dr. Tom Tetzlaff Institute for Advanced Simulation (IAS-6) Computational and Systems Neuroscience Jülich Research Centre Jülich, Germany
https://www.fz-juelich.de/profile/tetzlaff_t
Forschungszentrum Jülich GmbH 52425 Jülich Sitz der Gesellschaft: Jülich Eingetragen im Handelsregister des Amtsgerichts Düren Nr. HR B 3498 Vorsitzender des Aufsichtsrats: MinDir Stefan Müller Geschäftsführung: Prof. Dr. Astrid Lambrecht (Vorsitzende), Karsten Beneke (stellv. Vorsitzender), Dr. Ir. Pieter Jansens
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Attachments:
- smime.p7s
A little follow-up: I made a pull request with a suggested fix for this issue: https://github.com/nest/nest-simulator/pull/3384
Tom, Pavel, would you mind having a look to see if this addresses the issue?
Again, much obliged! Charl
On Mon, Dec 16, 2024, at 11:14, Charl Linssen wrote:
Hi Tom,
Thanks for your input on this! I was just looking at the code too. It seems that the order of updates (first update adaptation variables, then send spike, versus first send spike, then update adaptation variables) was changed, in order to prevent the adaptation from kicking in for the very first spike to ever go through the synapse. See https://github.com/nest/nest-simulator/pull/1672/.) Notice that t_lastspike is initialised to zero, rather than some "special" value like -1. It is possible that this fix was erroneous, in that it ends up affecting not just the first-ever spike, but also subsequent spikes. I am investigating a bit more, but if you are convinced that the current behaviour is the correct one, then I can stop investigating.
Much obliged! Charl
On Mon, Dec 16, 2024, at 11:01, Tom Tetzlaff wrote:
Dear Pavel,
I can confirm you observation, and I share your surprise. Intuitively, one would expect that the synaptic state should return to the initial state after the long period of silence. However, after consulting the source papers (todyks_synapse: doi:10.1523/JNEUROSCI.20-01-j0003.2000, tsodyks2_synapse: doi:10.1016/S0893-6080(01)00144-7) and looking at the NEST code, I conclude that what the example shows is indeed the intended behaviour.
A key difference between tsodyks_synapse and todyks2_synapse is the order of update steps: upon arrival of the presynaptic spike, tsodyks_synapse first updates the synapse state (x,y,z,u; see lines 258-277 in tsodyks_synapse.h), and then reports the new weight (line 285 in tsodyks_synapse.h). In tsodyks2_synapse, this is reversed: first, the synaptic weight is computed based on the current (i.e., the old) synapse state (line 236 in tsodyks2_synapse.h), and then the synapse state is updated (lines 243,244 in tsodyks2_synapse.h). When the first spike of the second burst of presynaptic spikes arrives in the example (https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody...), the recovery value of tsodyks2_synapse still is where it was at the end of the first burst of spikes (before the break). This is exactly how the model is defined in (Maass and Markram, 2002; doi:10.1016/S0893-6080(01)00144-7; see equations (1-5)). It needs one more spike to really "recover".
I hope this clarifies things a bit.
Best Tom
On 12.12.24 12:05, Pavel Haman wrote:
Hello,
I have followed the example script comparing `tsodyks_synapse` and `tsodyks2_synapse` models, which can be found here: https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody...
Using the parameter set for depression instead of facilitation I found that `tsodyks2_synapse` model leads to weird peaks of synaptic currents (at least I would not expect that for short term depression) - during the second stimulation the weight rescaling seems to be shifted by one spike compared to tsodyks_synapse (as if parameter "u" would be updated after rescaling the weight instead of before).
Does anybody know why is there this difference between the two synapse models?
Thank you.
All the best, PH _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
--
Dr. Tom Tetzlaff Institute for Advanced Simulation (IAS-6) Computational and Systems Neuroscience Jülich Research Centre Jülich, Germany
https://www.fz-juelich.de/profile/tetzlaff_t
Forschungszentrum Jülich GmbH 52425 Jülich Sitz der Gesellschaft: Jülich Eingetragen im Handelsregister des Amtsgerichts Düren Nr. HR B 3498 Vorsitzender des Aufsichtsrats: MinDir Stefan Müller Geschäftsführung: Prof. Dr. Astrid Lambrecht (Vorsitzende), Karsten Beneke (stellv. Vorsitzender), Dr. Ir. Pieter Jansens
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Attachments:
- smime.p7s
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hello Tom and Charl,
thank you for your time, explanations and replies. You confirmed what I thought was happening.
I quickly checked the paper (Maass and Markram, 2002; doi:10.1016/S0893-6080(01)00144-7) and equations (1-5). I think that the model should not update weights with delay of one spike, since $\Delta t_n$ in the equations is actually $t_{n+1}-t_{n}$, i.e. referring to the new spike (it is explained in the sentence before equations), not to older spikes. (I believe here could be the issue, since interpreting $\Delta t_n$, as I would naturally do, as $t_{n}-t_{n-1}$ leads to this one spike delay of updates)
Effective weight for spike $n+1$, $A_{n+1}$ is given as $A_{n+1}= A u_{n+1} R_{n+1}$ and $u_{n+1}$ depends on the times of the new spike $t_{n+1}$ and previous spike $t_{n}$
I will check Charl's PR later this week.
All the best, Pavel
po 16. 12. 2024 v 22:19 odesílatel Charl Linssen nest-users@turingbirds.com napsal:
A little follow-up: I made a pull request with a suggested fix for this issue: https://github.com/nest/nest-simulator/pull/3384
Tom, Pavel, would you mind having a look to see if this addresses the issue?
Again, much obliged! Charl
On Mon, Dec 16, 2024, at 11:14, Charl Linssen wrote:
Hi Tom,
Thanks for your input on this! I was just looking at the code too. It seems that the order of updates (first update adaptation variables, then send spike, versus first send spike, then update adaptation variables) was changed, in order to prevent the adaptation from kicking in for the very first spike to ever go through the synapse. See https://github.com/nest/nest-simulator/pull/1672/.) Notice that t_lastspike is initialised to zero, rather than some "special" value like -1. It is possible that this fix was erroneous, in that it ends up affecting not just the first-ever spike, but also subsequent spikes. I am investigating a bit more, but if you are convinced that the current behaviour is the correct one, then I can stop investigating.
Much obliged! Charl
On Mon, Dec 16, 2024, at 11:01, Tom Tetzlaff wrote:
Dear Pavel,
I can confirm you observation, and I share your surprise. Intuitively, one would expect that the synaptic state should return to the initial state after the long period of silence. However, after consulting the source papers (todyks_synapse: doi:10.1523/JNEUROSCI.20-01-j0003.2000, tsodyks2_synapse: doi:10.1016/S0893-6080(01)00144-7) and looking at the NEST code, I conclude that what the example shows is indeed the
intended
behaviour.
A key difference between tsodyks_synapse and todyks2_synapse is the order of update steps: upon arrival of the presynaptic spike, tsodyks_synapse first updates the synapse state (x,y,z,u; see lines 258-277 in tsodyks_synapse.h), and then reports the new weight (line 285 in tsodyks_synapse.h). In tsodyks2_synapse, this is reversed: first, the synaptic weight is computed based on the current (i.e., the old) synapse state (line 236 in tsodyks2_synapse.h), and then the synapse state is updated (lines 243,244 in tsodyks2_synapse.h). When the first spike of the second burst of presynaptic spikes arrives in the example (
https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody...),
the recovery value of tsodyks2_synapse still is where it was at the end of the first burst of spikes (before the break). This is exactly how the model is defined in (Maass and Markram, 2002; doi:10.1016/S0893-6080(01)00144-7; see equations (1-5)). It needs one more spike to really "recover".
I hope this clarifies things a bit.
Best Tom
On 12.12.24 12:05, Pavel Haman wrote:
Hello,
I have followed the example script comparing `tsodyks_synapse` and
`tsodyks2_synapse` models, which can be found here:
https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody...
Using the parameter set for depression instead of facilitation I found
that `tsodyks2_synapse` model leads to weird peaks of synaptic currents (at least I would not expect that for short term depression) - during the second stimulation the weight rescaling seems to be shifted by one spike compared to tsodyks_synapse (as if parameter "u" would be updated after rescaling the weight instead of before).
Does anybody know why is there this difference between the two synapse
models?
Thank you.
All the best, PH _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
--
Dr. Tom Tetzlaff Institute for Advanced Simulation (IAS-6) Computational and Systems Neuroscience Jülich Research Centre Jülich, Germany
https://www.fz-juelich.de/profile/tetzlaff_t
Forschungszentrum Jülich GmbH 52425 Jülich Sitz der Gesellschaft: Jülich Eingetragen im Handelsregister des Amtsgerichts Düren Nr. HR B 3498 Vorsitzender des Aufsichtsrats: MinDir Stefan Müller Geschäftsführung: Prof. Dr. Astrid Lambrecht (Vorsitzende), Karsten Beneke (stellv. Vorsitzender), Dr. Ir. Pieter Jansens
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Attachments:
- smime.p7s
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hi Pavel,
you are right. Indeed, I assumed that $\Delta{}t_n=t_n-t_{n-1}$, rather than $\Delta{}t_n=t_{n+1}-t_n$, as used in the paper. Well spotted. This makes much more sense now. Sorry for the confusion. I agree that the current NEST implementation is hence wrong, obviously the result of an earlier "fix", as Charl explained. I had a quick look at Charl's PR and think this will solve the problem.
Best Tom
On 17.12.24 12:22, Pavel Haman wrote:
Hello Tom and Charl,
thank you for your time, explanations and replies. You confirmed what I thought was happening.
I quickly checked the paper (Maass and Markram, 2002; doi:10.1016/S0893-6080(01)00144-7) and equations (1-5). I think that the model should not update weights with delay of one spike, since $\Delta t_n$ in the equations is actually $t_{n+1}-t_{n}$, i.e. referring to the new spike (it is explained in the sentence before equations), not to older spikes. (I believe here could be the issue, since interpreting $\Delta t_n$, as I would naturally do, as $t_{n}-t_{n-1}$ leads to this one spike delay of updates)
Effective weight for spike $n+1$, $A_{n+1}$ is given as $A_{n+1}= A u_{n+1} R_{n+1}$ and $u_{n+1}$ depends on the times of the new spike $t_{n+1}$ and previous spike $t_{n}$
I will check Charl's PR later this week.
All the best, Pavel
po 16. 12. 2024 v 22:19 odesílatel Charl Linssen <nest-users@turingbirds.com mailto:nest-users@turingbirds.com> napsal:
A little follow-up: I made a pull request with a suggested fix for this issue: https://github.com/nest/nest-simulator/pull/3384 <https://github.com/nest/nest-simulator/pull/3384> Tom, Pavel, would you mind having a look to see if this addresses the issue? Again, much obliged! Charl On Mon, Dec 16, 2024, at 11:14, Charl Linssen wrote: > Hi Tom, > > Thanks for your input on this! I was just looking at the code too. It > seems that the order of updates (first update adaptation variables, > then send spike, versus first send spike, then update adaptation > variables) was changed, in order to prevent the adaptation from kicking > in for the very first spike to ever go through the synapse. See > https://github.com/nest/nest-simulator/pull/1672/ <https://github.com/nest/nest-simulator/pull/1672/>.) Notice that > t_lastspike is initialised to zero, rather than some "special" value > like -1. It is possible that this fix was erroneous, in that it ends up > affecting not just the first-ever spike, but also subsequent spikes. I > am investigating a bit more, but if you are convinced that the current > behaviour is the correct one, then I can stop investigating. > > Much obliged! > Charl > > > On Mon, Dec 16, 2024, at 11:01, Tom Tetzlaff wrote: >> Dear Pavel, >> >> I can confirm you observation, and I share your surprise. Intuitively, >> one would expect that the synaptic state should return to the initial >> state after the long period of silence. However, after consulting the >> source papers (todyks_synapse: doi:10.1523/JNEUROSCI.20-01-j0003.2000, >> tsodyks2_synapse: doi:10.1016/S0893-6080(01)00144-7) and looking at the >> NEST code, I conclude that what the example shows is indeed the intended >> behaviour. >> >> A key difference between tsodyks_synapse and todyks2_synapse is the >> order of update steps: upon arrival of the presynaptic spike, >> tsodyks_synapse first updates the synapse state (x,y,z,u; see lines >> 258-277 in tsodyks_synapse.h), and then reports the new weight (line >> 285 >> in tsodyks_synapse.h). In tsodyks2_synapse, this is reversed: first, >> the >> synaptic weight is computed based on the current (i.e., the old) >> synapse >> state (line 236 in tsodyks2_synapse.h), and then the synapse state is >> updated (lines 243,244 in tsodyks2_synapse.h). When the first spike of >> the second burst of presynaptic spikes arrives in the example >> (https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsodyks2_synapse.html <https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsodyks2_synapse.html>), >> the recovery value of tsodyks2_synapse still is where it was at the end >> of the first burst of spikes (before the break). This is exactly how >> the >> model is defined in (Maass and Markram, 2002; >> doi:10.1016/S0893-6080(01)00144-7; see equations (1-5)). It needs one >> more spike to really "recover". >> >> I hope this clarifies things a bit. >> >> Best >> Tom >> >> >> >> On 12.12.24 12:05, Pavel Haman wrote: >>> Hello, >>> >>> I have followed the example script comparing `tsodyks_synapse` and `tsodyks2_synapse` models, which can be found here: >>> https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsodyks2_synapse.html <https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsodyks2_synapse.html> >>> >>> Using the parameter set for depression instead of facilitation I found that `tsodyks2_synapse` model leads to weird peaks of synaptic currents (at least I would not expect that for short term depression) - during the second stimulation the weight rescaling seems to be shifted by one spike compared to tsodyks_synapse (as if parameter "u" would be updated after rescaling the weight instead of before). >>> >>> Does anybody know why is there this difference between the two synapse models? >>> >>> Thank you. >>> >>> All the best, >>> PH >>> _______________________________________________ >>> NEST Users mailing list -- users@nest-simulator.org <mailto:users@nest-simulator.org> >>> To unsubscribe send an email to users-leave@nest-simulator.org <mailto:users-leave@nest-simulator.org> >> >> -- >> --- >> >> Dr. Tom Tetzlaff >> Institute for Advanced Simulation (IAS-6) >> Computational and Systems Neuroscience >> Jülich Research Centre >> Jülich, Germany >> >> https://www.fz-juelich.de/profile/tetzlaff_t <https://www.fz-juelich.de/profile/tetzlaff_t> >> >> >> ----------------------------------------------------------------------- >> ----------------------------------------------------------------------- >> Forschungszentrum Jülich GmbH >> 52425 Jülich >> Sitz der Gesellschaft: Jülich >> Eingetragen im Handelsregister des Amtsgerichts Düren Nr. HR B 3498 >> Vorsitzender des Aufsichtsrats: MinDir Stefan Müller >> Geschäftsführung: Prof. Dr. Astrid Lambrecht (Vorsitzende), >> Karsten Beneke (stellv. Vorsitzender), Dr. Ir. Pieter Jansens >> ----------------------------------------------------------------------- >> ----------------------------------------------------------------------- >> >> _______________________________________________ >> NEST Users mailing list -- users@nest-simulator.org <mailto:users@nest-simulator.org> >> To unsubscribe send an email to users-leave@nest-simulator.org <mailto:users-leave@nest-simulator.org> >> >> Attachments: >> * smime.p7s > _______________________________________________ > NEST Users mailing list -- users@nest-simulator.org <mailto:users@nest-simulator.org> > To unsubscribe send an email to users-leave@nest-simulator.org <mailto:users-leave@nest-simulator.org> _______________________________________________ NEST Users mailing list -- users@nest-simulator.org <mailto:users@nest-simulator.org> To unsubscribe send an email to users-leave@nest-simulator.org <mailto:users-leave@nest-simulator.org>
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hello Tom and Charl,
I have checked Charl's PR - it solves the problem. After a brief silent period the synapse behaves the same way as tsodyks_synapse and as could be expected.
I have found one more issue playing with parameters, which I think is connected to the original issue ( https://github.com/nest/nest-simulator/issues/957) regarding the first spike.
When initiating the synapse, changing parameter 'u' (in dep_params dictionary) to anything different than 'U', e.g. raising it to 0.9 or lowering to 0.2 leads to a different first peak (and consequently the following peaks), compared to tsodyks_synapse model. The model obviously determines the first value u_1 (referring to the paper Maass and Markram, 2002; doi:10.1016/S0893-6080(01)00144-7) as 'u' parameter. Which I believe does not correspond with the paper, where it is set u_1 = U (eg. 2), so for the first spike it should be set as 'U'.
Currently it behaves such that the u_1 value is 'u' and it is ignoring any time passed since the start of the simulation (i.e. the synapse does not relax). This means that for this synapse 'u' does not represent the initial value of the dynamical variable which is in differential equations describing the synaptic behavior, but rather strictly the value of that variable at the time of the first spike.
I am not sure that is the intended behavior, I can just report that it was unexpected behavior for me (it does not depend whether the first spike is at time t=2 ms or t=500 ms).
Note, that the default value of 'u' for tsodyks2_synapse is 0.5 regardless of the value of 'U'.
All the best, Pavel
út 17. 12. 2024 v 14:14 odesílatel Tom Tetzlaff t.tetzlaff@fz-juelich.de napsal:
Hi Pavel,
you are right. Indeed, I assumed that $\Delta{}t_n=t_n-t_{n-1}$, rather than $\Delta{}t_n=t_{n+1}-t_n$, as used in the paper. Well spotted. This makes much more sense now. Sorry for the confusion. I agree that the current NEST implementation is hence wrong, obviously the result of an earlier "fix", as Charl explained. I had a quick look at Charl's PR and think this will solve the problem.
Best Tom
On 17.12.24 12:22, Pavel Haman wrote:
Hello Tom and Charl,
thank you for your time, explanations and replies. You confirmed what I thought was happening.
I quickly checked the paper (Maass and Markram, 2002; doi:10.1016/S0893-6080(01)00144-7) and equations (1-5). I think that the model should not update weights with delay of one spike, since $\Delta t_n$ in the equations is actually $t_{n+1}-t_{n}$, i.e. referring to the new spike (it is explained in the sentence before equations), not to older spikes. (I believe here could be the issue, since interpreting $\Delta t_n$, as I would naturally do, as $t_{n}-t_{n-1}$ leads to this one spike delay of updates)
Effective weight for spike $n+1$, $A_{n+1}$ is given as $A_{n+1}= A u_{n+1} R_{n+1}$ and $u_{n+1}$ depends on the times of the new spike $t_{n+1}$ and previous spike $t_{n}$
I will check Charl's PR later this week.
All the best, Pavel
po 16. 12. 2024 v 22:19 odesílatel Charl Linssen <nest-users@turingbirds.com mailto:nest-users@turingbirds.com> napsal:
A little follow-up: I made a pull request with a suggested fix for this issue: https://github.com/nest/nest-simulator/pull/3384 <https://github.com/nest/nest-simulator/pull/3384> Tom, Pavel, would you mind having a look to see if this addresses the issue? Again, much obliged! Charl On Mon, Dec 16, 2024, at 11:14, Charl Linssen wrote: > Hi Tom, > > Thanks for your input on this! I was just looking at the code too. It > seems that the order of updates (first update adaptationvariables,
> then send spike, versus first send spike, then update adaptation > variables) was changed, in order to prevent the adaptation from kicking > in for the very first spike to ever go through the synapse. See > https://github.com/nest/nest-simulator/pull/1672/ <https://github.com/nest/nest-simulator/pull/1672/>.) Notice that > t_lastspike is initialised to zero, rather than some "special"value
> like -1. It is possible that this fix was erroneous, in that it ends up > affecting not just the first-ever spike, but also subsequent spikes. I > am investigating a bit more, but if you are convinced that the current > behaviour is the correct one, then I can stop investigating. > > Much obliged! > Charl > > > On Mon, Dec 16, 2024, at 11:01, Tom Tetzlaff wrote: >> Dear Pavel, >> >> I can confirm you observation, and I share your surprise. Intuitively, >> one would expect that the synaptic state should return to the initial >> state after the long period of silence. However, after consulting the >> source papers (todyks_synapse: doi:10.1523/JNEUROSCI.20-01-j0003.2000, >> tsodyks2_synapse: doi:10.1016/S0893-6080(01)00144-7) and looking at the >> NEST code, I conclude that what the example shows is indeed the intended >> behaviour. >> >> A key difference between tsodyks_synapse and todyks2_synapse isthe
>> order of update steps: upon arrival of the presynaptic spike, >> tsodyks_synapse first updates the synapse state (x,y,z,u; seelines
>> 258-277 in tsodyks_synapse.h), and then reports the new weight (line >> 285 >> in tsodyks_synapse.h). In tsodyks2_synapse, this is reversed: first, >> the >> synaptic weight is computed based on the current (i.e., the old) >> synapse >> state (line 236 in tsodyks2_synapse.h), and then the synapse state is >> updated (lines 243,244 in tsodyks2_synapse.h). When the first spike of >> the second burst of presynaptic spikes arrives in the example >> (https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody... < https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody...
), >> the recovery value of tsodyks2_synapse still is where it was at the end >> of the first burst of spikes (before the break). This is exactly how >> the >> model is defined in (Maass and Markram, 2002; >> doi:10.1016/S0893-6080(01)00144-7; see equations (1-5)). It needs one >> more spike to really "recover". >> >> I hope this clarifies things a bit. >> >> Best >> Tom >> >> >> >> On 12.12.24 12:05, Pavel Haman wrote: >>> Hello, >>> >>> I have followed the example script comparing `tsodyks_synapse` and `tsodyks2_synapse` models, which can be found here: >>>
https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody... < https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody...
>>> >>> Using the parameter set for depression instead of facilitation I found that `tsodyks2_synapse` model leads to weird peaks of synaptic currents (at least I would not expect that for short term depression) - during the second stimulation the weight rescaling seems to be shifted by one spike compared to tsodyks_synapse (as if parameter "u" would be updated after rescaling the weight instead of before). >>> >>> Does anybody know why is there this difference between the two synapse models? >>> >>> Thank you. >>> >>> All the best, >>> PH >>> _______________________________________________ >>> NEST Users mailing list -- users@nest-simulator.org <mailto:users@nest-simulator.org> >>> To unsubscribe send an email to users-leave@nest-simulator.org <mailto:users-leave@nest-simulator.org> >> >> -- >> --- >> >> Dr. Tom Tetzlaff >> Institute for Advanced Simulation (IAS-6) >> Computational and Systems Neuroscience >> Jülich Research Centre >> Jülich, Germany >> >> https://www.fz-juelich.de/profile/tetzlaff_t <https://www.fz-juelich.de/profile/tetzlaff_t> >> >> >>
>>
>> Forschungszentrum Jülich GmbH >> 52425 Jülich >> Sitz der Gesellschaft: Jülich >> Eingetragen im Handelsregister des Amtsgerichts Düren Nr. HR B3498
>> Vorsitzender des Aufsichtsrats: MinDir Stefan Müller >> Geschäftsführung: Prof. Dr. Astrid Lambrecht (Vorsitzende), >> Karsten Beneke (stellv. Vorsitzender), Dr. Ir. Pieter Jansens >>
>>
>> >> _______________________________________________ >> NEST Users mailing list -- users@nest-simulator.org <mailto:users@nest-simulator.org> >> To unsubscribe send an email to users-leave@nest-simulator.org <mailto:users-leave@nest-simulator.org> >> >> Attachments: >> * smime.p7s > _______________________________________________ > NEST Users mailing list -- users@nest-simulator.org <mailto:users@nest-simulator.org> > To unsubscribe send an email to users-leave@nest-simulator.org <mailto:users-leave@nest-simulator.org> _______________________________________________ NEST Users mailing list -- users@nest-simulator.org <mailto:users@nest-simulator.org> To unsubscribe send an email to users-leave@nest-simulator.org <mailto:users-leave@nest-simulator.org>
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
--
Dr. Tom Tetzlaff Institute for Advanced Simulation (IAS-6) Computational and Systems Neuroscience Jülich Research Centre Jülich, Germany
https://www.fz-juelich.de/profile/tetzlaff_t
Forschungszentrum Jülich GmbH 52425 Jülich Sitz der Gesellschaft: Jülich Eingetragen im Handelsregister des Amtsgerichts Düren Nr. HR B 3498 Vorsitzender des Aufsichtsrats: MinDir Stefan Müller Geschäftsführung: Prof. Dr. Astrid Lambrecht (Vorsitzende), Karsten Beneke (stellv. Vorsitzender), Dr. Ir. Pieter Jansens
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Hi Charl,
this is interesting. What I meant was that the tsodyks2 synapse is --as far as I can tell-- implemented in NEST exactly as it is described in the original publication (Maass and Markram, 2002), in particular the order of reporting the weight (sending the spike) and updating the synapse state. In this sense, the NEST implementation does the right thing. If this order has been changed in the past in NEST, then this was good in this respect. Whether it is perhaps more meaningful to swap this order is a different question. Changing this would lead to a different model.
Best Tom
On 16.12.24 11:14, Charl Linssen wrote:
Hi Tom,
Thanks for your input on this! I was just looking at the code too. It seems that the order of updates (first update adaptation variables, then send spike, versus first send spike, then update adaptation variables) was changed, in order to prevent the adaptation from kicking in for the very first spike to ever go through the synapse. See https://github.com/nest/nest-simulator/pull/1672/.) Notice that t_lastspike is initialised to zero, rather than some "special" value like -1. It is possible that this fix was erroneous, in that it ends up affecting not just the first-ever spike, but also subsequent spikes. I am investigating a bit more, but if you are convinced that the current behaviour is the correct one, then I can stop investigating.
Much obliged! Charl
On Mon, Dec 16, 2024, at 11:01, Tom Tetzlaff wrote:
Dear Pavel,
I can confirm you observation, and I share your surprise. Intuitively, one would expect that the synaptic state should return to the initial state after the long period of silence. However, after consulting the source papers (todyks_synapse: doi:10.1523/JNEUROSCI.20-01-j0003.2000, tsodyks2_synapse: doi:10.1016/S0893-6080(01)00144-7) and looking at the NEST code, I conclude that what the example shows is indeed the intended behaviour.
A key difference between tsodyks_synapse and todyks2_synapse is the order of update steps: upon arrival of the presynaptic spike, tsodyks_synapse first updates the synapse state (x,y,z,u; see lines 258-277 in tsodyks_synapse.h), and then reports the new weight (line 285 in tsodyks_synapse.h). In tsodyks2_synapse, this is reversed: first, the synaptic weight is computed based on the current (i.e., the old) synapse state (line 236 in tsodyks2_synapse.h), and then the synapse state is updated (lines 243,244 in tsodyks2_synapse.h). When the first spike of the second burst of presynaptic spikes arrives in the example (https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody...), the recovery value of tsodyks2_synapse still is where it was at the end of the first burst of spikes (before the break). This is exactly how the model is defined in (Maass and Markram, 2002; doi:10.1016/S0893-6080(01)00144-7; see equations (1-5)). It needs one more spike to really "recover".
I hope this clarifies things a bit.
Best Tom
On 12.12.24 12:05, Pavel Haman wrote:
Hello,
I have followed the example script comparing `tsodyks_synapse` and `tsodyks2_synapse` models, which can be found here: https://nest-simulator.readthedocs.io/en/stable/auto_examples/evaluate_tsody...
Using the parameter set for depression instead of facilitation I found that `tsodyks2_synapse` model leads to weird peaks of synaptic currents (at least I would not expect that for short term depression) - during the second stimulation the weight rescaling seems to be shifted by one spike compared to tsodyks_synapse (as if parameter "u" would be updated after rescaling the weight instead of before).
Does anybody know why is there this difference between the two synapse models?
Thank you.
All the best, PH _______________________________________________ NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
--
Dr. Tom Tetzlaff Institute for Advanced Simulation (IAS-6) Computational and Systems Neuroscience Jülich Research Centre Jülich, Germany
https://www.fz-juelich.de/profile/tetzlaff_t
Forschungszentrum Jülich GmbH 52425 Jülich Sitz der Gesellschaft: Jülich Eingetragen im Handelsregister des Amtsgerichts Düren Nr. HR B 3498 Vorsitzender des Aufsichtsrats: MinDir Stefan Müller Geschäftsführung: Prof. Dr. Astrid Lambrecht (Vorsitzende), Karsten Beneke (stellv. Vorsitzender), Dr. Ir. Pieter Jansens
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
Attachments:
- smime.p7s
NEST Users mailing list -- users@nest-simulator.org To unsubscribe send an email to users-leave@nest-simulator.org
-
Charl Linssen -
Pavel Haman -
Tom Tetzlaff