Randall O'Reilly
Randall O'Reilly
even when using the same `mat32.FastExp` implementation, the Gnmda values diverge between GPU and CPU. Run the `TestStdGPUnData1Debug` test in examples/boa (set `TEST_DEBUG=true` to enable) to see these divergences. I...
It would probably be useful to use float16 for the synaptic variables. There is a nice existing Go package: https://github.com/x448/float16 vulkan 1.2 now includes it as a fully supported option:...
AxonPrjns has no useful methods other than Add, and everything has been moved to the layer (SendPrjnNameTry etc) which return emer.Prjn interface and are not super useful for actually operating...
There is a very strange bug in GPU NewState: layerparams.go: ```Go // NewStateNeuron handles all initialization at start of new input pattern. // Should already have presented the external input...
There are various things that could be done on the GPU, which are currently only on CPU, but are low-cost and thus low-pri: * PlusPhasePost (matrix gating, corsim stat, act...
* Sender only spikes * Sender spikes and recv Spiked window -- 10 msec
Due to extra complexity in supporting synaptic-level random variables, this has not been implemented for a while and is now officially gone. Can reinstate at some later point as we...
It is not useful in the scale of networks we're currently using. Will likely become important as we scale up, but it does add a bit of complexity and cost...
Early in learning, the first curiosity / novelty VSMtxGo pool can drive gating even with no BLA novelty pool activity, because it is disinhibited by the curiosity drive and it...