Want to read Slashdot from your mobile device? Point it at m.slashdot.org and keep reading!


Forgot your password?
Slashdot Deals: Deal of the Day - 6 month subscription of Pandora One at 46% off. ×

Comment Looking forward to the next year! (Score 1) 106

This may read like I'm a Julia fan-boy ... I guess I am.

I found out about Julia from the Machine Learning course from Coursera. Not directly, for at that time it was Octave; the advice given there was "trust me, for machine learning, this syntax is better." Indeed for many machine learning algorithms, the basis of understanding it, is vector and matrix operations. The innovation of Matlab which both Octave, which is essentially a gnu, open-source implementation of Matlab, and Julia is making vector valued variables first class (e.g. M*X, M^-1 where M is a matrix and X is a vector) makes things succinct and clear -- btw M^-1 is a representation of the inverse of M, an O^3 order algorithm in 4 characters?

Now yes, Python has numpy, which is close syntactically, but there are yet other comparisons were is not quite so easy, and Julia has an advantage here in that it's so new that devs are still tolerant of syntax changes -- for instance the behavior of {} was changed between Julia 0.3 and 0.4. And so if there's something new on the horizon that needs a re-org, Julia is better able to handle it.

The other thing of course which Julia and Python and R communities are attempting to do is to figure out the best way to extract the optimizations available from LLVM, and owing to it's close ties to and ability to modify to conform to changes of LLVM, Julia also has an advantage. As I've posted before, expect Julia to be able to scale almost linearly on the Xenon Phi (Knight's Landing+) for HPC linear algebra oriented applications -- expect this by Julia 0.5.

Comment Re:Oh, goody (Score 1) 91

I think we can agree to try not to be an idiot, but really that's as much as I'm willing to agree to. The problem with arbitrary rules is they are easily gamed or abused. Remember when the Ferguson police got the FAA to declare a no fly zone -- the reason was bullshit of course, but oops too late sorry about that. Lifting the argument up even beyond that, there are legitimate circumstances for the government to restrict behavior and there are other circumstances where they would like to, but shouldn't be able to. That tension is lost when there's no opportunity to mis-behave, and that includes individuals that are not being thoughtful.

Comment Re:News for (computer architecture) nerds... (Score 1) 179

The specific application? I'm not sure, but as for the nature of the application that could make use of this is a combination of traditional mechanisms and neural networks. Thus AI couched in big data. Spark of course would be part of it. There are a number of NN frameworks that make use of cudNN. The first because so much has already been written assuming a standard kind of architecture; think OpenBlas. Newer supervised learning results revolve around neural networks and GPUs are best for that for even though the TFLOP comparison might be similar, the real limitation on NN is memory bandwidth, and GPU win here. Plus all the additional circuitry dedicated to double precision is wasted on NN, but valuable for say PCA. Trivia question; what's the register size on the Phi? It's 512 bits, which can be manipulated to vectorize, but how to do that? Only a compiler can manage that, and both gcc and LLVM do so, at a higher layer and longer term what language is best suited for this? I'm betting on Julia since the language is oriented to these problems and is centered on LLVM but mostly because of the people.

Comment Re:How about a Beowolf cluster of these (Score 1) 84

Beowolf? Don't think so. If anything then maybe Spark with the map or reduce step being executed on the GPU, or better yet Tensor Flow. But as pointed out elsewhere, this chip is not even for that, especially next year when the new cards with NVlink blow away 980/Titan-X stuff of this year. No this thing is for drones, AR, or image recognition on embedded anything where power consumption and latency are the overwhelming factors. Otherwise graphics cards will outperform, or if latency is not a factor, then the whole thing can be offloaded to AWS or similar. Also, 16 bit (half-precision) floats normally bad for numerics are fine for neural networks with the bonus advantage of effectively doubling the memory bandwidth and problem size which are the current limitations.

Comment Re:Real Reason For Caps (Score 1) 264

I can tell you how bad it is, at least in the case of AT&T and Time Warner. In general, it's not bad so long as your problem is a common one.

However, the situation changes whenever you operate outside of the norm, or lets' say 1.5sigma for stats nerds. At that point it's as bad as it gets. I've had call centers just simply hang up on me twice when faced with core-networking problems. I believe the business decision in these situations the ISP would rather lose you as a customer forever rather than pay to fix your uncommon problem -- in fact I've been told this directly by a AT&T tech, but he sounded bitter. In other words, these companies are fundamentally set up to be unable to accommodate the %1 problem.

Is this a case of the former or the latter? I would say it's interesting because it sits right on the margin between the 2, and Comcast is trying to push the envelope towards the latter.

Comment Re:Time to bring out the thirty-year-old cliches (Score 1) 568

This metaphor only tells part of the story. Woodpecker's would only destroy buildings 1.0 (0.9 of course would collapse all by itself), 2.0 would have swivel guns to shoot the woodpeckers before they could "attack" the house, but would of course shoot some people too, but usually not their owner. House 3.0 would be pretty good, but by then people would have grown bored of houses and decided to live in mine shafts. The first groundhog would tend to destroy mine shaft 1.0 (mine shaft 0.9 would of course cave in by itself), etc.

Comment Re:Err... (Score 1) 161

The torrent part I agree with: torrenting can be very demanding on the networks, and torrents are not used in applications that require real time. They'll be fine if their file transfers take a little bit longer.

10% longer ---> 10% more time spent --> 0.1*C*X lost efficiency from loss of relevancy. Here's the problem, you're uninformed and are suggesting unfair solutions because of it. Now imagine the problem magnified a million-fold (a million people all with their own perspective) some of which aren't even trying to be fair. In the face of that, how does one suggest prioritization rules when the ones with most-likely adopted opinions are the ones with the most money?

Comment But is this enough to change policy? (Score 2) 38

Is seems pretty clear now that an outright ban on neonicotinoids is what is is called for here. The overall effect in bee population in countries that use is versus ones that don't combined with the detectable presence in honey and now this show that even controlled use is has too many unexpected side effects. The mere benefit of improved pest control efficiency is not worth this danger.

Comment The problem of anonymity (Score 1) 184

These broad statistics like average rating and stars are really a throwback to a time when processors were too slow for techniques like cooperative filtering , which in this context, essentially uses people to recommend movies and movies to recommend people back and forth for a more accurate estimator. The system is simultaneously able to find people that like the same kind of movies based on a common set of features and movies that have those features. The idea being that if you're a person that has preferences similar to someone that has rates a film highly then you will tend to also rate it highly. But the problem for simple rating sites is that these advanced techniques require you to establish an identity to differentiate yourself from "the average" and contribute by rating, which can be tedious while systems like Netflix can probably do more by inferring you rating by repeat viewing, tendency to scan, etc. However, once identity is established, then even fraudulent types of behavior can be dealt with. It's surprising that imdb, rotten tomatoes, fandango, etc cling to such inaccuracies, but I could find no reference to such techniques, though admittedly I didn't spend a long time looking.

Comment Re:And this matters *why*? (Score 1) 77

As for me, 60/6 currently expected to upgrade to 300/20 on Nov 3; Time Warner. Google fiber later. For now ASUS 802.11ac hand-me-down is sufficient (not always downstairs). What's going on with your pathetic bandwidth, do you live in some rural area or something? LEO satellite will take care of that... not this year.

Comment Cue the Kenny Loggins Music!! (Score 1) 29

Wherein the Maverick AI enabled drone shows real promise until he loses his SSD backup best friend in a freak accident sending him into a deep depression and bought of soul searching. But he pulls his shit together in the end and saves the day.

Yea, and let's throw some lasers in there too for good measure, and maybe aliens, not sure.

Nah, aliens.

"You can't get very far in this world without your dossier being there first." -- Arthur Miller