Add symmetric stdp synapse to NEST #218

sanjayankur31 · 2016-02-01T16:13:34Z

Reference:
Vogels et al. (2011) Inhibitory Plasticity Balances Excitation and
Inhibition in Sensory Pathways and Memory Networks. Science Vol. 334,
Issue 6062, pp. 1569-1573 DOI: 10.1126/science.1211095

http://www.sciencemag.org/content/334/6062/1569.abstract
Also adds a python test file to the manualtests folders.

sanjayankur31 · 2016-02-01T16:23:27Z

The test python file in manualtests generates this graph:
https://github.com/sanjayankur31/nest-symmetric-stdp-test/blob/master/results/gnuplot.png

Things I'm not sure about:

I'm not so sure of the case where spikes are co-incident. Even when I didn't have the conditional, I didn't get any facilitation. I'm wondering if this has something with the way nest does things?
Do I need to implement the hpc variant too? And a hom variant?

tammoippen · 2016-02-02T13:32:15Z

models/stdp_symmetric_connection.h

@@ -0,0 +1,300 @@
+/*
+ *  stdp_connection.h


In the copyright notice, the correct file name should be stated: stdp_symmetric_connection.h. This is why regressiontests/ticket-659-copyright.py fails.

tammoippen · 2016-02-02T13:36:19Z

Also the formatting of some of the file seems to be off, e.g. see TravisCI output. Please make sure, that you use clang-format version 3.6 for the formatting.

Reference: Vogels et al. (2011) Inhibitory Plasticity Balances Excitation and Inhibition in Sensory Pathways and Memory Networks. Science Vol. 334, Issue 6062, pp. 1569-1573 DOI: 10.1126/science.1211095 http://www.sciencemag.org/content/334/6062/1569.abstract Also adds a python test file to the manualtests folders.

The version in Fedora is 3.7.x which was causing the tests to fail.

sanjayankur31 · 2016-02-02T16:25:28Z

@tammoippen - Thank you for notes on the test failures. The new set of commits seem to be OK.

jakobj · 2016-02-15T11:12:14Z

@flinz @suku248 would be good second reviewers. @sanjayankur31 : a non-manual test would be a great addition.

sanjayankur31 · 2016-02-15T12:00:12Z

Thanks @jakobj - I'll look into writing a non-manual test.

seeholza · 2016-02-15T14:16:41Z

models/stdp_symmetric_connection.h

+  facilitate_( double_t w, double_t kplus )
+  {
+    double_t norm_w = ( w / Wmax_ ) + ( lambda_ * eta_ * kplus );
+    return norm_w < 1.0 ? norm_w * Wmax_ : Wmax_;


I believe there is a small error, caused by the weight maximization. Let me demonstrate here, but the same holds for depression.

For given values of w and kplus, you are assigning the new w to be (assuming you don't hit the norm_w < 1.0):

w_returned = [( w_old / Wmax_ ) + ( lambda_ * eta_ * kplus )] * Wmax_ = w_old + lambda_ * eta_ * kplus * Wmax

This makes the weight updates scale additionally with Wmax, which I believe should not be the case. For facilitation this should rather be:

w_returned = w_old + lambda_ * eta_ * kplus

You can change the lines to something akin to this implementation.

Just as a remark: Looking at the corresponding line in stdp_connection for the additive case (mu_plus=0) I think we get the same scaling with Wmax. @abigailm @mhelias am I getting something wrong or is this an intended behavior?

I wasn't sure about this either, but decided to go with the way stdp_connection is implemented (as you observe above).

seeholza · 2016-02-15T15:03:46Z

Hey @sanjayankur31, I had a quick look at your code and the paper.

Before you go on diggin into the details below, here is a general point to consider: The Vogels learning rule is indeed not implementable by the current stdp_connection (there is a constant depression term, independent of k_minus), so this class of learning rules is a sensible addition. However, for the sake of keeping NEST more general, and this class more reusable, it could be beneficial to instead make this a STDP connection (using both tau_plus and tau_minus) with optional constant terms for both depression and facilitation (alpha_fac and alpha_dep). For tau_plus=tau_minus and alpha_fac=0 the current model would be contained as a special case.
@heplesser @jougs what's your opinion on this?

Besides, here are comments on the current class:

Naming: I think the class should be named stdp_connection_symmetric in accordance with, e.g. stdp_connection_hom. One could think about naming this stdp_connection_vogels, since it is a little more specialized than being simply symmetric: it has a constant depression term (see also the general remark above).
Redundancy of lambda: I think the parameter lambda might be redundant. You already introduce the parameter eta (the learning rate, as in the paper) and both only occur as lambda * eta. In this case, to adhere to paper notation, remove lambda from the implementation.
Update description: You should mention in the model description that the essential difference to stdp_connection is the constant depression of synapses, independent of the actual spike times. I think this could be helpful for people browsing models, that are not familiar with the Vogels model.
Expose alpha instead of rho_0: The following might be a matter of discussion, since its rather a question of how general we want NEST models to be implemented.
The analysis in the paper and identification of rho_0 (named kappa in your implementation) with a target firing rate relies on the feedforward setup of inhibition (Fig. S1).
I would personally prefer to use for the model implementation only the more general parameter alpha (Eq. 4 in the supplement) and then mention that for alpha = 2 * tau * rho_0 one would recover the exact implementation of Vogels' feedforward scenario.
This has the advantage that the inhibitory offset alphais directly exposed and controllable, since it does not change with both tau and kappa.
Testing: As you already mentioned above, you should include a proper automated test for the synapse. You can simply adapt for example this test for the triplet rule to check for correct weight changes against the actual equations of Vogels Eqs. S4 & S5.

sanjayankur31 · 2016-02-15T15:14:18Z

Hi @flinz - thank you very much for your comments. I'm travelling over the next week or so and will address them as soon as get back.

sanjayankur31 · 2016-02-15T15:18:47Z

One question before I make any more changes - is the suggested practice in nest to simply add more commits on top of the current ones - thus maintaining the complete review and code history, or should I squash commits up to make it all tidy (and then maybe force push)?

jougs · 2016-02-15T15:31:31Z

@sanjayankur31: we think it's nicer to just add commits to your branch in order to have a nice and transparent history of code review process.

@flinz: many thanks for your nice and solid review and the quick reaction.

@heplesser: could you make sure the copyright transfer is handled properly? Thanks!

sanjayankur31 · 2016-02-15T19:00:17Z

I've fixed some of the minor issues and pushed commits. I'll fix the major issues as soon as I can.

As requested, I've replaced kappa with alpha, and added a comment in the documentation explaining the relation between the two.

weidel-p · 2016-03-02T16:15:27Z

testsuite/manualtests/test_stdp_symmetric.py

+
+    # set up the synapse
+    syn_spec_synapse = {'weight': weight_pre, 'Wmax': 100.,
+                        'kappa': 0.1, 'eta': 0.001,


When I run this test, I get Unused dictionary items: kappa.
Looking into the default parameter reveals that kappa is indeed not a parameter.

kappa was replaced by alpha after the review. The manual test file hasn't been updated since it'll be replaced by an automated test (I'm working on the latter)

sanjayankur31 · 2016-04-06T10:39:46Z

TODO - add source files to models/CMakeLists.txt

heplesser · 2016-06-10T11:06:45Z

@sanjayankur31 Would you pull the newest changes from master so that Travis can check this PR properly now that #391 is merged?

This ensures that the synapse can be both excitatory and inhibitory.

sanjayankur31 · 2016-06-20T11:08:28Z

Hiya, I've updated the source to use an implementation similar to the triplet synapse in #284. So this is finally ready for a complete review and merge :)

heplesser · 2016-06-20T11:53:53Z

models/modelsmodule.cpp

@@ -112,6 +112,7 @@
 #include "stdp_triplet_connection.h"
 #include "stdp_dopa_connection.h"
 #include "stdp_pl_connection_hom.h"
+#include "vogels_sprekeler_connection.h"


Could you put this in alphabetical order?

heplesser · 2016-06-20T11:57:14Z

@sanjayankur31 The code looks clean except for two little details, see inline comments.

@flinz @suku248 Would you have a final look before we merge?

suku248 · 2016-07-05T15:02:48Z

Seems fine to me 👍

heplesser · 2016-07-05T15:09:49Z

@sanjayankur31 And merging :).

seeholza · 2016-08-09T12:19:18Z

@heplesser sorry I was on holidays. Too late for the party 👍

Add three types of nearest-neighbour STDP: symmetric (but not to be confused with PR nest#218), presynaptic-centered, and restricted symmetric. These are worth implementing because described in a highly- cited review [Morrison A., Diesmann M., and Gerstner W. (2008) Phenomenological models of synaptic plasticity based on spike timing, Biol. Cybern. 98, 459--478]. In these models a spike is taken into account in the weight change rule not with all preceding spikes, but only with some of the nearest, see the three pairing schemes in Fig. 7 in [Morrison et al., 2008]. The implementation relies on two additional variables - presynaptic and postsynaptic traces, like in stdp_synapse. However, their update rules differ from those of the latter. That's why the archiving_node was modified to add one more state variable, nearest_neighbor_Kminus. It is not stored in the node, just computed on the fly when requested. To return nearest_neighbor_Kminus I changed the archiving_node.get_K_values() function signature, because, as fas as I can see, this function is not currently used anywhere. The tests for all three models reside in a single file (but as three separate tests), and are the following: generate two Poisson spike sequences as presynaptic and postsynaptic, feed them to NEST to get the synaptic weight change, then reproduce the weight change independently outside of NEST and check whether the expected weight equals the obtained one.

tammoippen reviewed Feb 2, 2016
View reviewed changes

sanjayankur31 added 2 commits February 2, 2016 14:41

Format files using clang-format 3.6.0

92b55b6

The version in Fedora is 3.7.x which was causing the tests to fail.

sanjayankur31 force-pushed the symmetric_stdp_synapse branch from b33fc84 to 92b55b6 Compare February 2, 2016 14:42

seeholza reviewed Feb 15, 2016
View reviewed changes

sanjayankur31 added 4 commits February 15, 2016 16:16

Rename to stdp_connection_symmetric

0c4350e

Remove lambda

1a446ca

Update model description

4817c20

Remove pointless comment

08f73f3

sanjayankur31 added 6 commits February 16, 2016 12:10

Correct file name in header

181efc0

Add Kplus to dictionary

aed19ed

Add bad property check for Kplus

9ddb61a

Replace kappa with alpha

8c4a16c

As requested, I've replaced kappa with alpha, and added a comment in the documentation explaining the relation between the two.

Remove normalised calculations

6bc3208

Intermediate - wip

022cbd0

weidel-p reviewed Mar 2, 2016
View reviewed changes

Remove manual test

a99072f

sanjayankur31 added 7 commits April 10, 2016 12:25

Merge branch 'master' into symmetric_stdp_synapse

93839a7

Add vogels sprekeler source file to CMakelists.txt

13a865e

Merge branch 'master' into symmetric_stdp_synapse

6badc11

Reformat with clang-format

58fcecb

Merge branch 'master' into symmetric_stdp_synapse

d8a8b7d

Pep8 related corrections

09faf76

Merge branch 'master' into symmetric_stdp_synapse

a5933c7

sanjayankur31 added 3 commits June 10, 2016 14:16

Merge branch 'master' into symmetric_stdp_synapse

a8f29ea

Merge branch 'master' into symmetric_stdp_synapse

280695c

Update file to mimic triplet synapse source

00e27b1

This ensures that the synapse can be both excitatory and inhibitory.

heplesser reviewed Jun 20, 2016
View reviewed changes

heplesser added the T: Enhancement New functionality, model or documentation label Jun 20, 2016

heplesser assigned suku248 Jun 20, 2016

sanjayankur31 added 3 commits June 20, 2016 13:08

Correct alphabetical order

b568dc1

Remove commented lines

4291efb

Merge branch 'master' into symmetric_stdp_synapse

2389d55

heplesser merged commit 62c019b into nest:master Jul 5, 2016

sanjayankur31 deleted the symmetric_stdp_synapse branch July 5, 2016 15:30

aserenko mentioned this pull request Dec 6, 2017

Add STDP synapses with nearest-neighbor spike pairing schemes #865

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add symmetric stdp synapse to NEST #218

Add symmetric stdp synapse to NEST #218

sanjayankur31 commented Feb 1, 2016

sanjayankur31 commented Feb 1, 2016

tammoippen Feb 2, 2016

tammoippen commented Feb 2, 2016

sanjayankur31 commented Feb 2, 2016

jakobj commented Feb 15, 2016

sanjayankur31 commented Feb 15, 2016

seeholza Feb 15, 2016

sanjayankur31 Feb 15, 2016

seeholza commented Feb 15, 2016

sanjayankur31 commented Feb 15, 2016

sanjayankur31 commented Feb 15, 2016

jougs commented Feb 15, 2016

sanjayankur31 commented Feb 15, 2016

weidel-p Mar 2, 2016

sanjayankur31 Mar 2, 2016

sanjayankur31 commented Apr 6, 2016

heplesser commented Jun 10, 2016

sanjayankur31 commented Jun 20, 2016

heplesser Jun 20, 2016

heplesser commented Jun 20, 2016

suku248 commented Jul 5, 2016

heplesser commented Jul 5, 2016

seeholza commented Aug 9, 2016

Add symmetric stdp synapse to NEST #218

Add symmetric stdp synapse to NEST #218

Conversation

sanjayankur31 commented Feb 1, 2016

sanjayankur31 commented Feb 1, 2016

tammoippen Feb 2, 2016

Choose a reason for hiding this comment

tammoippen commented Feb 2, 2016

sanjayankur31 commented Feb 2, 2016

jakobj commented Feb 15, 2016

sanjayankur31 commented Feb 15, 2016

seeholza Feb 15, 2016

Choose a reason for hiding this comment

sanjayankur31 Feb 15, 2016

Choose a reason for hiding this comment

seeholza commented Feb 15, 2016

sanjayankur31 commented Feb 15, 2016

sanjayankur31 commented Feb 15, 2016

jougs commented Feb 15, 2016

sanjayankur31 commented Feb 15, 2016

weidel-p Mar 2, 2016

Choose a reason for hiding this comment

sanjayankur31 Mar 2, 2016

Choose a reason for hiding this comment

sanjayankur31 commented Apr 6, 2016

heplesser commented Jun 10, 2016

sanjayankur31 commented Jun 20, 2016

heplesser Jun 20, 2016

Choose a reason for hiding this comment

heplesser commented Jun 20, 2016

suku248 commented Jul 5, 2016

heplesser commented Jul 5, 2016

seeholza commented Aug 9, 2016