mdda

Wow, I'm surprised to see a comment so critical of Venture Capitalists on HN. But I guess it's fair: The VC industry as a whole badly underperforms the S&P index on a risk-adjusted basis. Or aren't those the asset managers you were talking about?

(I'm not totally clear whether it was meant to be 1penny for each speck avoided or 1penny vs all specks.). But two follow-up questions :

a) would you press the 7bn specks button for an individual 'reward' of $70million?

b) would you press the N-specks button if given a choice between that and the 1-torture button?

Agreed - what I don't understand is why most of the discussion here is about high valuations. The big reveal of the article is that the valuations are known by investors to be illusory - and (colloquially) the VCs are ripping the faces off the entrepreneurs.

(responding here because your comment addresses the actual new content of the paper).

Four things pop out at me from the paper:

1) The whitening (per batch) & rescaling (overall) is a neat new idea. But (as referred to in their p5 comment about the bias term being subsumed) this also points to the idea that the (Wu+b) transformation probably has a better-for-learning 'factorization', since their un-scale/re-scale operation on (Wu+b) is mainly taking out such a factor (while also putting in the minibatch accumulation change).

2) The idea that this could replace Dropout as the go-to trick for speeding up learning is pretty worrying (IMHO), since the gains from Dropout seem to be in a 'meta network' direction, rather than a data-dependency direction. Both approaches seem well worth understanding more thoroughly, even though the 'do what works' ML approach might favour leaving Dropout behind.

3) The publication of this paper, so closely behind the new ReLu+ results from Microsoft, seems too coincidental. One has to wonder what other results each of the companies has in their back-pockets so that they can repeatedly steal the crown from each other.

4) For me, the application to MNIST is attention-grabbing enough. While I appreciate that playing with Inception (etc) sexes-up the paper a lot, it raises the hurdle for others who may not have that quantity of hardware to contribute to the (more interesting) project of improving the learning rates of all projects (which is quite possible to do on the MNIST dataset, except that it's pretty much 'solved' with the error cases being pretty questionable for humans too).

I'm going to assume your comment wasn't as mean as it appears in writing.

Why am I here? Because I started my first business in college, then went into finance to get some 'real world' experience, and then went on to start several other businesses (both tech and finance). I've employed quite a few people over the years, had some good times, some bad. You know : The basic entrepreneurial story. I started programming with 6502 machine code, C, C++, Perl, and more recently Python, Scala and GPUs. Will likely start something to do with deep-learning/NLP in the New Year. HN feels like home. That's why I'm here.

As to the content of my comment : It was intentionally low. I wanted to see how many negatives a valid comment about actually being in the finance industry would generate (-2 after a couple of hours, constant since). And, besides, any person in the finance business would immediately smile at the clear-cut upside/downside evaluation that is so common in the business. I just said it out loud.

Best wishes to you.

It's also interesting that they might not have got a warrant to look at his garden for a month. They weren't looking for any particular behaviour - just something that would allow them to barge into his house and discover evidence of drug-related wrong-doing. If they had spotted (say) public drunkenness, they might also have argued that his house be searched. Putting up a camera to spot 'any tiny thing that the person does wrong' would seem to be very flimsy warrant-wise (I would hope).

And reading the first paragraph explains why left-wing/right-wing is not so clear-cut :

"We like free enterprise and tend to favour deregulation and privatisation. But we also like gay marriage, want to legalise drugs and disapprove of monarchy. "

According to the Technical Detail tab : "5V2A DC input : This is for 5V power input, with an inner diameter of 0.8mm, and an outer diameter of 2.5mm. The ODROID-C1 consumes less than 0.5A in most cases, but it can climb to 2A if many passive USB peripherals are attached directly to the main board. "

What's the upside to admitting on HN to having front office experience?

I don't see that Skype get a license to the content, only that you should have the rights to distribute it (http://www.skype.com/en/legal/tou-usa/#5), and that they have the right to "take it down" if they feel like it.

The reason that the emoticons are so well targeted is not for the users themselves. If someone includes one of the emoticons, it makes machine-identification of sentiment trivial (compared to reading and interpreting the text).

(not sure whether this is something that everyone knows already : Perhaps I only just got the news... (Presuming this is actually true, of course))

Agreed - the title should be changed. Perhaps the submitter saw the Nvidia cards in there and read this as an OpenCL vs CUDA kind of thing, not seeing the principal cool-factor of OpenCL : that OpenCL interoperates across all these devices (GPUs and CPUs - despite Nvidia making it difficult to find in their SDK)

Excellent! The article manages to make two opposing arguments simultaneously (same problem can be seen in Firefox too).

The most remarkable take-away from this whole genre of word-embedding is that just by doing 'dumb averages' of word contexts and then optimizing the 'vector[word]' on the input (and output sides), you end up with a SEMANTIC understanding of the English language in the word vectors.

This paper is the latest in the series (across multiple researchers), and seems to boil the task down to its bare minimum : Just a raw least-squares optimization works. And instead of the 'linguistic knowledge' being smuggled into the problem set-up increasing (initially, people used tree-embeddings, and WordNet bootstrapping, in the 2003 papers), this is getting rid of almost all structure. And ending up with better results.

So, instead of semantics being a naturally very deep problem, apparently common sense understanding can be derived from surface statistics. IMHO, more people should be excited about this (from an AI standpoint).

And I just became aware of the NYC Halal carts run by the NYPD (in the talk by Steve Rambam at HopeX recently) [0].

[0] http://youtu.be/dNZrq2iK87k?t=2h3m9s

LTV normally incorporates churn as part of the calculation. If it didn't then unless you have a time-value-of-money discount in there, the LTV of a customer would be infinite. Perhaps you're positing that the LTV refers to the actual physical Lifetime of a customer - but, statistically, that just turns into a churnrate too (sad to say, even 100% of a consumer's lifetime is more than a 0.5% churn rate spread across the whole population).

This is a great article for the corporate titans of yesteryear, which puts today's 'innovation malaise' into perspective (as well as putting the IT boom in a historical context). For instance :

""" Perhaps the greatest shock to our Rip Van Winkle economist, however, would be the fact that, with the exception of the plastics industry, the main engines of growth in the past 50 years were already mature or rapidly maturing industries, based on well-known technologies, back in 1913. """

and

""" Within the next ten years, information will become very much cheaper. An hour of computer time today costs several hundred dollars at a minimum; I have seen figures that put the cost at about a dollar an hour in 1973 or so. Maybe it won’t come down that steeply, but come down it will. """

Overall, a very interesting read.

OTOH, there's nothing to stop a separate cipher key for the first text, and the author making sure that the key words chosen (there's always a choice, since only the first letters are used) map (roughly) to an interesting sequence in the Declaration of Independence - as a gotcha.

Another great comeback for the annals of HN (like https://news.ycombinator.com/item?id=35083)

Not in his individual Wikipedia article (yet) : He name is the 'C' in VC-dimension : http://en.wikipedia.org/wiki/VC_dimension

If I were the UK government, I would have paid for a last minute "YES" poll : Just to jolt the NO voters into going to vote (because otherwise, they may have not bothered, seeing that the polls were all pointing their way anyway).

If you are male, then this financier's philanthropy may help you in the future : http://en.wikipedia.org/wiki/Michael_Milken (Prostate Cancer)

Here's how to kill two birds with one (gamified) stone : If you host your static blog on GitHub Pages, you'll get a 'contribution' point for every day you post (or even commit) to it. Thus building up your GitHub streak, and post count/size consistently.

If you click on the links, there are machine & human-readable bibliographic entries on the same page.

The irony(?) for me is that I'm a UK ex-pat (who has paid the UK license fee in the UK this year) living in Singapore. But my Singaporean ISP "helpfully" provides VPN connections so that I'm a local in foreign countries (so that I can use Pandora, for instance 'in New York'). So... I'm overseas, but I share all the problems of the UK residents, who are, understandably, also miffed.

PS: Having looked at the video (via the wayback machine) - is it the Sleeper that I'm reminded of? I'm getting deja vu with some 70s SciFi spoof...

karpathy is being too modest here. He's also created his own convnet.js [1], which you can play with online [2] (complete with relevant, working demos), etc.

[1] https://github.com/karpathy/convnetjs [2] http://cs.stanford.edu/people/karpathy/convnetjs/

The patent was drafted very generally, but seems to focus on the volume (rather than elasticity) of the internal components. OTOH, it would be great if it were even more generally drafted, since its priority date is in 1995. 1995+17 = 2012, so the world is now wide open for innovation in this area, unimpeded by patents (IANAL).

There are some good TLDRs here. But the actual article includes a link to a UI with the actual ImageNet decision task created by the author. You can compete against the winning ConvNet too..

Willow Garage somehow pulling the rug from under Unbounded Robotics is a backwards move for this industry, which has always struggled to get decent forward momentum (being innovative electro-mechanical hardware, looking for early adopters, etc)

On a possibly related/speculative note, it seemed to me (when looking over the list of robotics companies acquired by Google at the end of 2013) that Google has been organizing a 'getting the band back together' thing. Obvious connections included Redwood Robotics and Industrial Perception, both also Willow Garage spin-offs. That, and one of the original founders of Willow Garage (Scott Hassan) was ex-early-Googler himself.

And Scala [5] (with a Spark-specific fork too, iSpark [6])

[5] https://github.com/mattpap/IScala [6] https://github.com/tribbloid/ISpark