Neural networks do benefit from parallelism, and i'm sure GPUs will help them run a bit faster, but it's not enough...

I'm convinced by Simon Knowle's analysis of learning and inference compute patterns: which if you accept - focusing on making NNs run faster on CPUs and GPUs appears to be an approach with severely limited potential. This isn't about some basic hardware optimisations gained by turning something into an ASIC, it's because design features of CPUs and GPUs actively work against NN compute patterns.

These issues are explained well in his keynote at 18:30, he goes on to explain how inference compute patterns (quite literally in the case of computer vision) want to do the complete opposite of what a GPU is designed to do in a deterministic way. Here is a list of the pattern characteristics:

1. Large, irregular, sparse data-structures (graphs) -> SIMD vectors can't help.
2. Huge aggregate compute, huge available parallelism.
3. Little compute per communicated bit -> opposite of GPU trend.
4. No data access locality -> caches are harmful.

I haven't bought a new car in my life and probably never will, the vast majority are not the kind of people who have cash to splash on new and shiny cars... Currently that's all that electric cars are.

Apologies, i forgot the NSA (especially when combined with FBI) only posses the ability to interpret everything literally... you see the contents of the communication you intercepted (aka my post) was a peculiar type of human fiction known as a "joke", what this means is that i did not actually train and send an army of electro sensitive zombies to defeat the NSA at sugar grove.

Bwahahahaha, my evil plan... "Release the brainwashed electro-sensitive zombies!", all but invisible to the NSA, no known modern technology can track them...

Let see how long they can stand the onslaught of pseudo-science arguments until they cave and abandon post.

:P thanks for that clarification i just couldn't make the connection before.

Zero doesn't make a lot of sense if for instance you are dividing something by a dynamically changing denominator that hits zero at some point... the result would change from a very large number suddenly to 0.

Divide by zero is infinity so using the largest supported number type seems reasonable for the calculation of real numbers.

You are missing one major variable.

You're bagel shop analogy has no rate, if instead each customer is only allowed to buy one bagel per minute and the bagel shop has sold the unlimited bagels package to 100 people... then they know that the maximum capacity they must be able to provide is 100 bagels per minute.

Of course It's not reasonable to expect an ISP would ever buy their maximum possible required capacity because they would quickly be out-competed, but if their peak load requires more capacity then using sneaky tactics to limit heavy users rates is not an honest strategy, in which case they are overselling. Basing your capacity requirements on peak usage statistics, combined with 1/n allocation in the event of capacity exhaustion is fair.

Big ISPs wouldn't do this because they'd rather sell an "unlimited package" that is only unlimited to people who never use more than 1 GiB per week. They will happily disappoint the 10% who try to use their service to the extent it was advertised to keep those 90% happy.

I'd agree that it's important to provide a base level of service to everyone no matter what happens to the capacity... But does "eveyone" include the heavy users? i fail to see anything more fair than 1/n customers division, the heavy users gets as much bandwidth as the lightweight users in this scenario... how is that not fair.

http://thumbs2.ebaystatic.com/...

*slow clap* for people who have nothing interesting to add.

poor argument

Then bathrooms and changing rooms are all sexist.

Segregation of sex is sometimes just the practical and appropriate choice, especially in cases of defining characteristics and basic interaction of sexes... Is segregation appropriate and necessary in this situation? that depends on whether less extreme options exist. In certain institutions rules for this kind of interaction between sexes (which tends to introduce extreme bias) calls for strict disciplinary action.

Suggesting segregation of gender to prevent inappropriate relations in a workplace is not "sexist", it doesn't mean it's not an extreme option either though.

You're right, he does say that, and that is out of line when taken literally... Honestly though first time i read it i took it based on context; which is a two way relationship... you can interpret the second part of that statement and i interpreted it to mean that both sexes can't take criticism very well from someone whom they are intimately involved with in "the lab", which seems quite likely.

Acknowledging the consequences of gender is not sexist.

Too many stupid people confuse issues that are intrinsic to gender with sexism... The fact that humans have gender creates problems, some are specific to one gender and some apply to both: Women need bras... that's not sexist, that is being a woman.

In this case; Straight people of opposite genders have higher probability of being attracted to each other, this creates issues in the workplace - not exactly shocking is it.

Feminazis what us to be asexual or something, i say go fuck yourselves, and i'll fuck someone else. good day.

... Why do you own an iPhone.

The browser has direct control of globally determining what is allowed in that page. You cannot click to play HTML5 adverts because the whole point is that they are part of the page... there is no rule for determining how ads are embedded into a page, so there is no reliable objective way of telling if that content is an advert or not, for instance attempting to implement a generic-global click to play WebGL content is just as likely to target your primary content as it is an ad.

It's reasonable to argue however that flash "is not an ad" because using click to play on flash kills everything that is flash... It's not inconceivable that the canvas 2d and 3d contexts could be limited in some way, the problem is that a lot of use of those contexts are not only legitimate and widely used but are also lightly used and not CPU or GPU hungry. In which case there perhaps a "click to give full resources" as opposed to a "click to do anything at all" would be more appropriate.

I'm not saying HTML5 is evil, i really like the web actively moving away from plugins and toward better more refined specifications to write content properly, i'm saying advertisers are evil, they fill pages we want to view with alien content that can eat up resources we don't want to give it... flash was actually a solution to isolate that (don't get me wrong i hate flash, but i was happy for content i hate to stay in flash).

I actually quite like that most of the highly animated CPU hogging adverts are written in flash, because i can easily disable all of them.

What concerns me is when those advertisers are finally forced to start writing them in javascript + Canvas / SVG / WebGL... yes it's possible to write efficient animated HTML5 content, request animation frame etc... but that's not forced, you think advertisers give a shit about that stuff? they will use everything at their disposal once flash is considered completely obsolete. Look forward to unsandboxed memory leaks and poorly optimised animation directly in your page... yay