iskander

44 ms·

541.
by iskander 14y ago
Just a small clarificatation: the 2.x version of LuaJIT only compiles floating point operations to SSE2 ( http://lua-users.org/lists/lua-l/2010-03/msg00142.html ). I'm not sure, however, whether it does vectorization or just operates on one
542.
by iskander 14y ago
>Academia is already facing difficulties in computer science because much of the state-of-the-art research is being done by private companies...As a consequence, there are a number of areas in computer science where the leading academic
543.
Which NP-complete problems are actually easy?(reddit.com)
1 points by iskander 14y ago|0 comments
544.
by iskander 14y ago
Huh, your extensive research doesn't square with much of what I learned as a neuroscience undergrad. We read a few papers that demonstrated non-trivial physiological responses due to 'sham' treatments. In particular, placebo treatments have
545.
by iskander 14y ago
Some related studies: 'Partial antagonism of placebo analgesia by naloxone' ( http://www.ncbi.nlm.nih.gov/pubmed/6308540 ) -- no actual morphine was ever adminstered, only a regimen of saline injections, whose placebo effect is apparently a
546.
by iskander 14y ago
I was hoping for "cartesian product analysis" (how can we trim down the exponential explosion of types?), fearing "certified public accountant" and was completely surprised.
547.
by iskander 14y ago
They have some very prominent machine learning folks there and I went to the Machine Learning Summer School they hosted a few years back. Seemed legitimate enough to me!
548.
by iskander 14y ago
Is anyone from the States still in Jalalabad or is the security situation too risky? Also, how did you guys get started?
549.
by iskander 14y ago
I agree that this was an amazing piece. However... >Is it Islam? The last time Afghanistan had any internal stability was under a fundamentalist Islamic government...so probably not? Anyway, prescribing solutions from thousands of miles
550.
by iskander 14y ago
Huh, I guess these comments reinforce just how diverse opinions can get-- I really enjoyed both the style and content of the story and it didn't occur to me that anyone wouldn't.
551.
by iskander 14y ago
For us crypto-ignoramuses, what's wrong with what you quoted?
552.
by iskander 14y ago
I think Steinbeck's original phrasing is better: "Socialism never took root in America because the poor see themselves not as an exploited proletariat but as temporarily embarrassed millionaires."
553.
by iskander 14y ago
Which of Ciresan's projects are you referring to? Everything I've seen by him uses convolutional layers of some sort.
554.
by iskander 14y ago
I think culturally, Hinton made a big splash and got people to pay attention to learning hierarchies and SGD-like training algorithms. Algorithmically, though, SGD is both ancient and still the dominant deep learning training technique (tho
555.
by iskander 14y ago
Boosting selects successive weak learners for the same classification problem, but under a changing distribution/weighting of the input space. Deep learning stacks complex models to create increasingly abstract representations. All I can re
556.
by iskander 14y ago
You can't train fully connected deep models with backprop, or at least not easily or well. An alternative solution to this problem is spatial weight pooling (Yann's convolutional networks) which play well with SGD.
557.
by iskander 14y ago
There is no "a neural net". Which model do you mean? Certainly not the deep belief networks under discussion in the article. edit: I'm not sure I was clear enough-- the term "neural network" is a misnomer that encompasses extremely differen
558.
by iskander 14y ago
Are you talking about Hinton's "A Fast Learning Algorithm for Deep Belief Nets"? Before that was published, Hinton's lab and their spiritual allies were training large restricted boltzmann machines via truncated sampling for decades. And Ya
559.
by iskander 14y ago
Very cool! I'm impressed at all the energy and fast pace of development coming out of Travis Oliphant's new company. Also, I'm curious about how this will interact with numba. Blaze seems to have a much richer type system and lazily constru
560.
by iskander 14y ago
The deconvolved images have some pretty nasty ringing. If you use an algorithm which places a realistic prior on the pixel gradients you will probably get cleaner results. For example, check out: Fast image deconvolution using hyper-laplaci
561.
by iskander 14y ago
Same thing in Seattle! I went to an Apple Store there with a friend and we found an Microsoft store across the parking lot. The exterior and some of the decor were similar, but the Microsoft products were much more haphazard (different bran
562.
by iskander 14y ago
It's not that you can't solve the halting problem--- it's that you can't always solve the halting problem. Sometimes (such as when you're building things that can explode or kill people) you'll settle for an overly conservative approximat
563.
by iskander 14y ago
Think Coffee on Mercer is a nice work spot when it's not swamped by NYU undergrads (later at night tends to be less crowded). Also, if you're anywhere near Fort Greene, Smooch is a pretty awesome coffee shop.
564.
by iskander 14y ago
It's a closely guarded secret but at the very least we know that every page is annotated with many different signals, which are combined with magical secret sauce.
565.
by iskander 14y ago
Your legal name is Jonathon but friends know you as Johnny. Should you be reported to facebook HQ for not using your legal name? What if your nickname distorts its origins further, such as "J", which easily turns into "Jay". Have we know le
566.
by iskander 14y ago
I've been off facebook for half a year but I'm still tempted to return. There isn't a good alternative for sharing photos with a group of friends, though hopefully something will emerge.
567.
by iskander 14y ago
I mean the actual weights of the neural network. After the first iteration, each worker starts with the same weights but performs SGD updates on different subsets of the data. After each epoch, you re-seed the workers with the best weights.
568.
by iskander 14y ago
It seems like a pretty straightforward MapReduce, though I agree that if you're doing searches repeatedly or your data is small enough you should use a database. (map = search a partition and return top k results, reduce = combine multiple
569.
by iskander 14y ago
My officemate was playing with training neural networks in parallel and he found that iterating what you were doing ended up with faster training times and lower error in the long run. 1. Seed every worker with a random initial paramet
570.
by iskander 14y ago
>So computing a stochastic gradient over each example in parallel, and them aggregating them, isn't necessarily a bad idea. Wait a minute, I don't think that's what the paper is suggesting. The communication costs of aggregating every m

More ›