Alejandro de la Vega

That's probably right.

Haven't tried it, but we are looking for a way to link back to Neuroscout (and were really just goign to abuse the `paper_url`)

Sounds like good options to me. We'll probably go with 1) (currently being used to link to OpenNeuro raw dataset), but maybe 2) can be opened as a separate issue.

I would suggest we should not merge it with BEP 16, but I also don't think we can make this decision in haste, as it has potential implications. Alternatively, this...

+1 for `node`. I think `model` is a very loaded word with more conceptual implications, whereas `node` can be more generic.

@JulioAPeraza awesome! q: why is it only scalable to a small number of CPUs? I don't quite understand that. but given that limitation, I agree with your reasoning and think...

Ah, disk i/o would certainly explain it. I wonder using "threads" would also help: https://joblib.readthedocs.io/en/latest/parallel.html#thread-based-parallelism-vs-process-based-parallelism

Do you know if explicitly setting `parallel=True` would help the numba optimized `_update_peak_assignments`? https://numba.readthedocs.io/en/stable/user/parallel.html It looks like the main concern would be the same matrices are being access by the...

Otherwise, I took a closer look at the GCLDA algorithm, and this all looks reasonable to me. I may have some more nit-picky comments later, but perhaps @jdkent and @tsalo...

By default, when running `Nodes`, the transformation history is now stored within the outputs. For example: ``` specs = root_node.run(group_by=root_node.group_by, force_dense=False) specs[0].coll_hist >>> {'run': [, , , ]} ``` The...