Randall O'Reilly

Issues for using MPI in GUI interactive mode

1

mpi depends entirely on each `proc` executing the same sequence of `AllGather` etc calls at the same times. If any node doesn't, everything just waits and then probably timeouts with...

Tests to add

1

Starting with a big list here and then can make separate tickets if that is better, later.. # core mechanisms: * Layer with sub-pools -- make sure inhibition is computed...

Data parallelism for GPU (and CPU)

1

We should get major efficiencies and speedup by running multiple "data" pathways through the same synaptic weights and network architecture. In effect, it is like "shared weights" for multiple copies...

Previously, we only learned for subPools that got above the gating threshold: ```Go prjnparams.go:369: if subPool.Gated.IsTrue() { dtr = rn.SpkMax * sn.CaSpkD // we will get the credit later at...

To make the code more self-documenting, we can switch to functional names for all layers, e.g., in the BOA model, but also maintain a biological `map[string]string` map that translates into...

currently just keeps hammering away at well-learned cases -- weights get too extreme. simple rescorla-wagner filter on rew should deal with that.

separate predictive learning loops from attentional ones

1

The demands of each are not compatible, and there are two separate pulvinar maps of visual areas, in addition to possible core vs. matrix distinction. Need to do more research,...

Available neurophys evidence suggests that 6CT neurons show a spectrum of different temporal delays, and in general such temporal spectral coding models are efficient at covering a range of different...

no need to keep that separately in deep at this point..