bartvm / blocks-paper Goto Github PK

I don't know the details of what needs to be said about this but I recall him asking that everyone mention this. This is in addition to acknowledging CIFAR as a source of funding.

Just opening a few tickets to organise discussion points.

First one: Should we submit the paper somewhere, or just stick with arXiv? Someone mentioned JMLR MLOSS to me once, which I think might not be a bad idea, but I'm not too familiar with it: http://jmlr.csail.mit.edu/mloss/. Any other options?

I re-organised the Blocks section a bit according to comments made by @rizar, but that meant I had to trim the Fuel part from @vdumoulin a bit to stay within the 4-page limit. Considering the NIPS deadline is this Friday, I was hoping to put it in arXiv at the start of the week (Monday or Tuesday?) so please have a read and make changes/comments (including things like the author list, title, etc.). I might make a few more edits tomorrow, but overall I think I'm okay with it right now.

Seems like Blocks mostly consists of 4 clearly separate components:

• blocks-graph: slightly more used-friendly interface to Theano computation graph, annotations and management of them, e.g. filtering, graph transformations
• blocks-bricks: the bricks and Selector as a tool to handle brick hierarchies (which is in its infancy so far, but I think will be more relevant as models become more and more complicated over years)
• blocks-algorithms: well, algorithms
• blocks-main-loop: the main loop and the extensions, the top-most level

The are of course also various utilities, but this contribute not so much to the big picture.
And yeah, I find it really nice how clearly these components are separated from each other!

I could make an SVG figure that illustrates this structure, which can be used in the paper and later in the documentation.

I'm thinking definitely before the NIPS deadline so that people have something to cite by then.

Are you sure it's possible? I mean, probably some of us are also have certain other duties related to NIPS rush. Also, there are few major things in Blocks which I think should be fixed beforehand, e.g. Model and Dump. And if we do it by the camera-ready deadline, we are still good, aren't we?