Add association matrix learning rule (AML).

[jgosmann]

Adds the association matrix learning rule (AML). This learning rule allows to learn associations from cue vectors to target vectors in a one-shot fashion without catastrophic forgetting. An implementation is provided for the reference simulator and Nengo OCL. As far as I can tell, Nengo OCL does not provide a possibility to register custom plan_* functions. Thus, the OCL implementation adds a new AmlSimulator class inheriting nengo_ocl.Simulator.

I added a test. To run it with the OCL implemenation:

pytest nengo_extras/tests/test_learning_rules.py -k aml --plots --simulator nengo_extras.ocl.AmlSimulator --slow

Demonstration:

More details on the learning rule will be in my forthcoming PhD thesis.

[Seanny123]

FYI, static tests are failing due to flake8 errors.

[jgosmann]

It'd be nice to get this merge and a release. I am planning a tutorial on the learning rule during the summer school.

[tbekolay]

I'll do this tomorrow 👍

[tbekolay]

Test runs successfully for me, so looks good! I'll add the learning rule to the learning_rules.rst file in the docs as part of the merge.

I did have a small issue with the test taking forever, but it turned out it was because of how the plt fixture works when not actually doing plotting. The iteration over fig.axes would never stop, so I changed it to explicitly iterating over a list of the three axes.

[jgosmann]

The iteration over fig.axes would never stop, so I changed it to explicitly iterating over a list of the three axes.

Maybe we also should fix the fixture? (But I'm fine with explicitly iterating over the axes here to have backwards compatibility.)

[tbekolay]

Maybe we also should fix the fixture? (But I'm fine with explicitly iterating over the axes here to have backwards compatibility.)

Yeah, I'll make an issue for it in Nengo.

[tbekolay]

nengo/nengo#1441 made, merging!