Thanks for sharing! Skimmed it a little and bookmarked to read later when I have time. This looks awesome.
There’s a lot of awesome work going on in federated learning / privacy preserving ML. I heard a talk by someone at Google at NeurIPS 2019 about how they use it for their keyboard prediction (among other things) and thought it was super exciting. Got me down the rabbithole of reading papers in this space, although I haven’t had a chance to apply it in more practical settings (yet).
Glad you’re interested! Google’s keyboard prediction work is perhaps the most prominent deployment of this stuff so far, and the papers are great. Exactly how that feature works as a piece of software is a little opaque IMO, and I’m a big fan of this blog post as a more accessible real world example: https://florian.github.io/federated-learning-firefox/.
Yeah I’ve read that post. It made me wonder why search engines haven’t yet tried to do local learning for recommendations, or what that might look like if implemented…
Do you know if anyone is distilling information in this space (academic / practical applications)? Writing about it is on my TODO list and I finally have time to start consolidating information / putting together resources, but thought I’d see if others already do this. I’m aware of a few yearly reviews done by prominent blogs but not sure if there are niche writers who I haven’t found yet!
The report I wrote was an attempt at that, although it’s pretty high level, and it’s now two years out of date. (I posted it here because it had just been unpaywalled.) The report was the basis for a StrangeLoop talk last year https://www.youtube.com/watch?v=VUINeZUAlx8 or https://mike.place/talks/fl/ (slides) . The last slide has a bunch of references, but they are all >= 1 year old.
Other than that, the recent reviews I’ve seen are all quite academic (no blog posts that I know of, although I’m sure they exist). Probably the most useful academic review I’ve seen is this one https://arxiv.org/abs/1912.04977. But because it reviews open problems (rather than solved problems!) it necessarily doesn’t have much to say about the real world practicalities. https://arxiv.org/abs/1902.01046 may also be useful, but is a little vague and very Googly.
Thanks for sharing! Skimmed it a little and bookmarked to read later when I have time. This looks awesome.
There’s a lot of awesome work going on in federated learning / privacy preserving ML. I heard a talk by someone at Google at NeurIPS 2019 about how they use it for their keyboard prediction (among other things) and thought it was super exciting. Got me down the rabbithole of reading papers in this space, although I haven’t had a chance to apply it in more practical settings (yet).
Glad you’re interested! Google’s keyboard prediction work is perhaps the most prominent deployment of this stuff so far, and the papers are great. Exactly how that feature works as a piece of software is a little opaque IMO, and I’m a big fan of this blog post as a more accessible real world example: https://florian.github.io/federated-learning-firefox/.
Yeah I’ve read that post. It made me wonder why search engines haven’t yet tried to do local learning for recommendations, or what that might look like if implemented…
Do you know if anyone is distilling information in this space (academic / practical applications)? Writing about it is on my TODO list and I finally have time to start consolidating information / putting together resources, but thought I’d see if others already do this. I’m aware of a few yearly reviews done by prominent blogs but not sure if there are niche writers who I haven’t found yet!
The report I wrote was an attempt at that, although it’s pretty high level, and it’s now two years out of date. (I posted it here because it had just been unpaywalled.) The report was the basis for a StrangeLoop talk last year https://www.youtube.com/watch?v=VUINeZUAlx8 or https://mike.place/talks/fl/ (slides) . The last slide has a bunch of references, but they are all >= 1 year old.
Other than that, the recent reviews I’ve seen are all quite academic (no blog posts that I know of, although I’m sure they exist). Probably the most useful academic review I’ve seen is this one https://arxiv.org/abs/1912.04977. But because it reviews open problems (rather than solved problems!) it necessarily doesn’t have much to say about the real world practicalities. https://arxiv.org/abs/1902.01046 may also be useful, but is a little vague and very Googly.