We've been using Torch for a while at work, really appreciate Facebook doing this. Some of us had a hard time adjusting to Lua, but it was well worth it.
Facebook open-sources deep-learning modules
21–30 of 76 posts
Re: Facebook open-sources deep-learning modules
#22Earlier quoted context omitted.
> Ads certainly don't feel like much of the FB experience, nowhere near the level of Google, or even Twitter. Facebook creates an artificial distinction between labeled ads (which are, of course, obvious ads) and posts with paid reach (which are also ads, but presented as normal content), exactly for the purpose of maximizing the quantity of ads while minimizing the impression of content being dominated by ads.
Right, so ads don't feel like much of the FB experience. My point is that FB seems much, much more focused on creating an engaging user experience than it does on ads, so it seems wrong to say that their deep learning efforts have ads as their one main goal.
> Right, so ads don't feel like much of the FB experience.
Right, but the issue was the claim that with FB everything is about ads, not the UX feels like it is about ads. Making ads not feel like as much of the FB experience as they are is, itself, about ads -- and, particularly, is about maximizing the degree to which the FB experience is actually about ads.
Re: Facebook open-sources deep-learning modules
#23Earlier quoted context omitted.
Basically everything in FB has one main goal: Ads Secondary goals might be: Spam detection and extracting info from the graph/pictures, etc FB identify people from your Social Graph in Pictures, for example
"Basically everything in FB has one main goal: Ads" Is this actually true, or a glib answer? Ads certainly don't feel like much of the FB experience, nowhere near the level of Google, or even Twitter. They're negligible, really. Whereas shaping the feed and identifying people in pics, etc, is central to the FB experience.
Categorization is critical to boosting precision in information retrieval tasks. Deep networks for cropped object recognition are an important source of (inferred) content categories (especially when other context, e.g. photo captions, are not useful). While a good image categorizer would be very useful for ad targeting and auctions, the technology can also boost precision for news feed, graph search, etc., which improves user experience.
Re: Facebook open-sources deep-learning modules
#24This is a recruiting move. Facebook is building its DL team, and to do that, they need to grow the community using their tools.
I see no issue with open sourcing tools to recruit developers. Honestly, that's one of the best methods I can think of.
I'm not against it. It's a great way to recruit devs. Just wanted to point out that they would likely be recruiting from the community.
Re: Facebook open-sources deep-learning modules
#25Too bad it is Torch. Working in the lua environment is not enjoyable at all. Every error becomes a long procedure of looking deep into the source of the framework since there is no type information or stack traces to go off whatsoever. You are constantly guessing about the shape of the data or what parameters are supposed to look like.
try http://deeplearning4j.org/
they just partnered with canonical: http://insights.ubuntu.com/?p=9373
Re: Facebook open-sources deep-learning modules
#26Earlier quoted context omitted.
Right, so ads don't feel like much of the FB experience. My point is that FB seems much, much more focused on creating an engaging user experience than it does on ads, so it seems wrong to say that their deep learning efforts have ads as their one main goal.
> Right, so ads don't feel like much of the FB experience. Right, but the issue was the claim that with FB everything is about ads, not the UX feels like it is about ads. Making ads not feel like as much of the FB experience as they are is, itself, about ads -- and, particularly, is about maximizing the degree to which the FB experience is actually about ads.
"Making ads not feel like as much of the FB experience as they are is, itself, about ads..."
Well, we just disagree here. I see it from exactly the other end. That the main FB goal is to engage users, and minimizing the (necessary, but grudgingly so) ads is about that.
Re: Facebook open-sources deep-learning modules
#27Sadly, this has facebook's new, horrible patent clause.
For those who aren't aware, it says
1. If facebook sues you, and you counterclaim over patents (whether about software or not), you will lose rights under all these patent grants.
So essentially you can't defend yourself.
This is different than the typical apache style patent grant, which instead would say "if you sue me over patents in apache licensed software x, you lose rights to software x" (IE it's limited to software, and limited to the thing you sued over)
2. It terminates if you challenge the validity of any facebook patent in any way. So no shitty software patent busting!
Re: Facebook open-sources deep-learning modules
#28Paper with proper summary: http://arxiv.org/abs/1412.7580
This paper describes one part of the fbcunn release (the fast convolution layers implemented via FFT, with the source available at https://github.com/facebook/fbcunn/tree/master/src/cuda/fft ). There's a lot more in fbcunn if you want to check it out.
In my experience, the fully connected layers are the bottleneck. The other issue was the alternating compute-heavy convolution and the IO-heavy pooling. I'm curious how this FFT implementation stacks up against cuDNN (what's the speedup like for just the convolutional layers? and then what's the overall speedup like?).
Re: Facebook open-sources deep-learning modules
#29Earlier quoted context omitted.
Right, so ads don't feel like much of the FB experience. My point is that FB seems much, much more focused on creating an engaging user experience than it does on ads, so it seems wrong to say that their deep learning efforts have ads as their one main goal.
> Right, so ads don't feel like much of the FB experience. Right, but the issue was the claim that with FB everything is about ads, not the UX feels like it is about ads. Making ads not feel like as much of the FB experience as they are is, itself, about ads -- and, particularly, is about maximizing the degree to which the FB experience is actually about ads.
I think the answer is both-and vs. either-or.
FB's bait is user engagement. So, applying machine learning to the goal of engaging users makes sense.
But, to monetize that user-engagement they, of course, use ads. Applying machine learning directly there makes sense too.
Beyond that, it's just semantics. Yes, one could accurately say that everything FB does is about ads because, ultimately, they are a for-profit company which derives its revenue from advertising. In that way, the ultimate goal of all of its activity is aimed at generating ad revenue.
But, that doesn't mean that every single activity or bit of technology that supports that activity is directly tied to driving ad revenue. That is to say, that if FB featured no advertising but, instead, charged users, they would still likely deploy machine-learning to user-engagement and, thus, customer-retention.
Re: Facebook open-sources deep-learning modules
#30Too bad it is Torch. Working in the lua environment is not enjoyable at all. Every error becomes a long procedure of looking deep into the source of the framework since there is no type information or stack traces to go off whatsoever. You are constantly guessing about the shape of the data or what parameters are supposed to look like.
there's clear and well stack-traces at the lua level. there's great debugger (in fblualib, called fb.debugger). there's an inbuilt profiler into LuaJIT. too bad you've had a bad experience.
Thanks I am always willing to give it another shot with more tools.