New method Connections::synapseCompetiton by ctrl-z-9000-times · Pull Request #466 · htm-community/htm.core

ctrl-z-9000-times · 2019-05-11T17:38:38Z

As discussed on the forum: https://discourse.numenta.org/t/synapse-competition/5957

breznak

I really like the new research this brings in.
And the throughout tests 👍
A few nits to the implementation.

But I don't see this actually used anywhere. Can you add a bool toggle to the SP,TM to actually make use of the new feature? Ideally show it in the MNIST exeriment?

src/nupic/algorithms/Connections.cpp

breznak · 2019-05-12T13:39:15Z

src/nupic/algorithms/Connections.cpp

+void Connections::synapseCompetition(
+                    const Segment    segment,
+                    const SynapseIdx minimumSynapses,
+                    const SynapseIdx maximumSynapses)


do we also have something like Connections' minConnectedSynapsesPerSegment ? If so, please also check against that

We have a method Connections::raisePermanencesToThreshold. What about it?

not a method, constants. Whether connections have in itself defined some limits, in the meaning of what minimumSynapses are.

ctrl-z-9000-times · 2019-05-12T15:18:08Z

But I don't see this actually used anywhere. Can you add a bool toggle to the SP,TM to actually make use of the new feature? Ideally show it in the MNIST exeriment?

I have integrated it into the MNIST experiment on the branch columnPooler2. The awkward part is that nothing in that branch is ready for public consumption.

breznak · 2019-05-13T13:42:36Z

I have integrated it into the MNIST experiment on the branch columnPooler2. The awkward part is that nothing in that branch is ready for public consumption.

yes, and as this is useful general-purpose and could conceptually be useful in SP/TM as well. Where should it be called?

Can we call "synaptic competition" (maybe call it synapticInhibition, as it's more descriptive and inline with the inhibition at cell level?) at Connections::updateSynapsesPermanence() (esp if said method is reworked to accept a vector (now its called in for-loop for each synapse)
and make SP/TM have a flag to toggle it, or I'd even say make it the defaults.

ctrl-z-9000-times · 2019-05-13T14:30:46Z

could be useful in TM

I don't know about the TM, I haven't tested this at all. The TM uses different methods to control the number of synapses in its potential pool; it add synapses when it wants more, and it does not add more synapses than it needs.

maybe call it synapticInhibition, as it's more descriptive and inline with the inhibition at cell level?

I don't think that inhibitory cells have anything to do with controlling the number of synapses. I also think that competition is more descriptive - it states what it does, not how it does it.

breznak · 2019-05-13T14:48:42Z

I don't know about the TM, I haven't tested this at all. The TM uses different methods to control the number of synapses in its potential pool;

can you add your experiment as an example? If we can visualize synapse utilization in SP on the MNIST task, we can compare with using this feature and potentialy make it default.

it states what it does, not how it does it.

that's right 👍

ctrl-z-9000-times · 2019-05-13T15:12:01Z

If we can visualize synapse utilization in SP

This is done by saving the connections to file (which the MNIST example already does), and calling
$ python3 -m nupic.examples.connections_analysis $CONNECTIONS

breznak · 2019-05-13T15:17:55Z

$ python3 -m nupic.examples.connections_analysis $CONNECTIONS

oh, I forgot about that script! Will try it out

sort directly by permanences instead of indirectly by synapses

breznak

A really nice feature! and great tests 👍

a number of, but small, nits; soon we're ready to go
please link the doc/motivation you've posted somewhere
can you add the python script with histogram if permanences?
and most importantly, ok if this is to be used for ColumnPooler, but I'd love to see at least experiment using it with current SP,TM. (if works well either as default, or with a toggle)

Thanks 🆒

src/htm/algorithms/Connections.cpp

breznak · 2019-07-03T02:10:55Z

src/htm/algorithms/Connections.cpp

+  // Sort the potential pool by permanence values, and look for the synapse with
+  // the N'th greatest permanence, where N is the desired number of connected
+  // synapses.  Then calculate how much to change the N'th synapses permance by
+  // such that it becomes a connected synapse.  After that there will be exactly


how much to change the N'th synapses permance by such that it becomes a connected synapse

Unrelated to acceptance of this PR, but: what is the biological mechanism for this? I am doing some work on at SP now and realized I "dont like" raisePermanencesToThreshold. The biological synapse cannot have this, and grow bit by bit (synPermActiveInc).

@ctrl-z-9000-times ping on this PR, I'm looking forward to have it merge, when you have time to integrated the feedback 👼

Still thinking about this PR:

please make it used in SP, TM.

consider the following:

instead of bumpSegment that puts all misperforming synapses in line (to min/max-desired)

have them die out.

if too few synapses on segment, create a random new. Or create each time one is removed/dead.

src/htm/algorithms/Connections.cpp

src/test/unit/algorithms/ConnectionsTest.cpp

breznak · 2019-07-03T02:48:26Z

src/test/unit/algorithms/ConnectionsTest.cpp

+  emptySegment.ncon = 0;
+  emptySegment.min  = 3;
+  emptySegment.max  = 100;
+  emptySegment.dont_check = true;


can we create a new synapse on an empty seg, if needed, rather than failing?

The synapse competition method does not have access to the current inputs. The call would need to specify which presynapses to use.

src/test/unit/algorithms/ConnectionsTest.cpp

ctrl-z-9000-times · 2019-07-12T00:01:54Z

Breznak,

I did everything except to actually use it in the SpatialPooler.
I'd like to merge this anyways since it's ready.

breznak · 2019-07-12T00:36:59Z

I'd like to merge this anyways since it's ready.

the new test is feiling with Debug.

except to actually use it in the SpatialPooler.

that's a pity, but fine, I'll try this later.

this could be used for SP,TM, right?
called on each syn update (would be probably expensive), or just with some probability?
where should this be called? I'd say randomly on updateSynapsePermanence?

ctrl-z-9000-times · 2019-07-12T01:59:23Z

Thanks for approving!

this could be used for SP,TM, right?

Yes, the SP could use this. The TM should not need it, I don't think this is applicable to the TM's segments.

called on each syn update (would be probably expensive), or just with some probability?

Your right: this does not need to be called too often.
Maybe only on random cycles, or if iteration % 10 == 0

where should this be called? I'd say randomly on updateSynapsePermanence?

Call after each call to adaptSegment

breznak

Thank you @ctrl-z-9000-times for this new functionality! Looking forward for seeing it used.

New method Connections::synapseCompetiton

c6f66b0

ctrl-z-9000-times added feature new feature, extending API research new functionality of HTM theory, research idea labels May 11, 2019

ctrl-z-9000-times self-assigned this May 11, 2019

ctrl-z-9000-times and others added 2 commits May 11, 2019 13:38

Merge branch 'master' into syn-comp

3dc97ce

SynapseCompetition fix unit tests

675198f

ctrl-z-9000-times closed this May 11, 2019

ctrl-z-9000-times reopened this May 11, 2019

breznak requested changes May 12, 2019

View reviewed changes

ctrl-z-9000-times added the in_progress label May 12, 2019

ctrl-z-9000-times added 3 commits June 17, 2019 11:46

Merge branch 'master' into syn-comp

8cac9a2

Merge branch 'master' into syn-comp

35bbbf3

Connection::synapseCompetition

ed2745e

sort directly by permanences instead of indirectly by synapses

ctrl-z-9000-times added ready and removed in_progress labels Jul 3, 2019

ctrl-z-9000-times requested a review from breznak July 3, 2019 01:03

breznak requested changes Jul 3, 2019

View reviewed changes

This was referenced Jul 3, 2019

TM Connections integration #537

Merged

Boosting cleanup #544

Open

ctrl-z-9000-times and others added 3 commits July 11, 2019 19:39

Synapse Competition: Review feedback

3f78ee9

Merge branch 'master' into syn-comp

3a4202e

synapse competition fixes

a023977

breznak previously approved these changes Jul 12, 2019

View reviewed changes

breznak mentioned this pull request Jul 12, 2019

SP, TM use synapseCompetition() #558

Open

Merge branch 'master' into syn-comp

73a95eb

Synapse Competition: Fix assert statement.

ac7a1d6

ctrl-z-9000-times dismissed breznak’s stale review via ac7a1d6 July 12, 2019 03:22

breznak approved these changes Jul 12, 2019

View reviewed changes

breznak merged commit 09d5391 into master Jul 12, 2019

breznak deleted the syn-comp branch July 12, 2019 06:34

This was referenced Jul 24, 2019

Connections: apply synapsesCompetition() in adaptSegment() #584

Open

Model constantly grows in size, implement removing synapses, segments; death #461

Open

Conversation

ctrl-z-9000-times commented May 11, 2019

Uh oh!

breznak left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

breznak May 12, 2019

Choose a reason for hiding this comment

Uh oh!

ctrl-z-9000-times May 12, 2019

Choose a reason for hiding this comment

Uh oh!

breznak May 13, 2019

Choose a reason for hiding this comment

Uh oh!

ctrl-z-9000-times commented May 12, 2019

Uh oh!

breznak commented May 13, 2019

Uh oh!

ctrl-z-9000-times commented May 13, 2019

Uh oh!

breznak commented May 13, 2019

Uh oh!

ctrl-z-9000-times commented May 13, 2019

Uh oh!

breznak commented May 13, 2019

Uh oh!

breznak left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

breznak Jul 3, 2019

Choose a reason for hiding this comment

Uh oh!

breznak Jul 10, 2019

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

breznak Jul 3, 2019

Choose a reason for hiding this comment

Uh oh!

ctrl-z-9000-times Jul 11, 2019

Choose a reason for hiding this comment

Uh oh!

Uh oh!

ctrl-z-9000-times commented Jul 12, 2019

Uh oh!

breznak commented Jul 12, 2019

Uh oh!

ctrl-z-9000-times commented Jul 12, 2019

Uh oh!

breznak left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants