aab0

The Democratic party targeting university students, and the Republican party trying voter-suppression efforts aimed at students (forbidding voting at the university where one lives, requiring photo IDs & banning college IDs as photo IDs) are nothing new.

That raises as many questions as it answers. Why FPGAs and not GPUs which can run just about any deep neural network but usually faster and more efficiently?

Technically, it's called 'domain drop catching': https://en.wikipedia.org/wiki/Domain_drop_catching Common service, so I wonder how park.io is so successful.

The disparity doesn't go away if you use victimization surveys.

This is not necessarily that surprising. Hinton's 'dark knowledge' (not cited in the paper) already showed that a remarkable amount of information is hidden in the classification probabilities emitted by a model, and that one neural net can learn a lot from and reverse-engineer another neural net given just its precise predictions.

I quipped the other day after reading it, 'In retrospect, training the first RL agents on Doom may not have been the best idea.'

All the anecdotes I've seen of being addicted to nicotine gum or patches have been of former smokers who were using them to try to quit. That's not necessarily nicotine being addictive so much as their pre-existing tobacco addiction latching onto a replacement.

Probably. You can add in 'speaker' as a bit of metadata to the samples (this is what is meant by 'conditioning on') and teach it to speak like different people, so if you have a diverse sample of speakers and you add in 'accent' as another variable, it might well learn to disentangle individual speakers from their accents and then you can control generated accents by changing the metadata.

"The researchers found that attempting to move at a superhuman pace (eg one action every frame), resulted in a subpar performance."

Moving at extremely fine-grained timesteps can make learning much more difficult, because now a reward arrives millions of timesteps delayed rather than hundreds or thousands. It's like trying to teach a NN to compose piano music by starting down at the 1ms raw audio level. This is part of why audio synthesis was so difficult up until recently with DeepMind's WaveNet. In theory, being able to move every frame should enable extremely superhuman performance, but in practice, you can't learn your way there. So often people will chunk data to make it easier to learn the higher-level concepts: operate on words, rather than characters, for example.

They're already sold as pets. Reportedly they're pretty good pets if you don't mind the price tag.

Just follow the WP links.

"Researchers, interested organizations or individuals are welcome to use our scientific data library for legitimate research endeavours. This data is available free of charge, however we do ask that a memorandum of understanding is executed for access privileges to our data. Please send us a message using the Contact Us form and state the nature of your request."

http://www.haidasalmonrestoration.com/index.php/science/scie...

For a 4x boost, I imagine Tensorflow and Torch will get support shortly after the GPUs start shipping in real quantities.

The argument is also bogus from the pro-Russian perspective: if this is 'legalized cheating' and so awful and hypocritical, why weren't the Russians doing it, much less resorting to stealing urine samples? They have doctors who can write certificates too.

It's not a true power-law (assuming it's not a lognormal). It's truncated by the fact that anyone considering a startup would never have earned more than the world GDP, or more specifically, the peak tech market cap to date (Apple's $775b). Once it's truncated at a finite value, the moments become meaningful.

Sounds like he's already been hit where it hurts. If his fortune is ~$3b and the stockpile is ~$6b, then he's losing a ton of money every day to interest and opportunity cost and shipping it back to Vietnam and however much the price has fallen since he bought it all, which could be a large fraction. Given how his partners have the knives out for him, he may not even have that $3b. One can stockpile, but as always, how does one bury the body?

90 minutes of computation for 1 second of audio on DeepMind's GPUs: https://twitter.com/hardmaru/status/773968758519902208

That's a lot of cranks of Moore's law. Better hope for considerable algorithmic improvements. (Raw is probably overkill anyway.)

This Bloomberg article adds nothing, and is less informative & interesting (doesn't have the audio samples for starters) than the original DeepMind blog post, and shouldn't be on the front page.

Hail Helix!

> Karpathy is pretty famous for his blogging and open software.

Yes, he's famous for that, but that didn't earn him his PhD nor would it have gotten him tenure. As he remarks about his teaching of his course, him doing a good job actively worked against him because... it's not writing a sexy new paper or networking.

> He also recommend you to release code in this article, so I'm not really sure what you are getting at.

Releasing code isn't the same thing as creating polished end-user applicable stuff on the level of char-rnn. In ML, you're increasingly expected to at least chuck over the wall a barebones implementation to demonstrate it works at all, but there is no expectation that it will be generalized, well-written, or polished, or maintained, and typically they are not. (Most ML releases I've looked at are kind of horrifying from a software engineering perspective. Just thinking about improved-gan makes me shudder.)

All valid advice, but god is some of it depressing. Papers can be evaluated by flipping through and looking for pretty graphs and equations. Incremental, replication, or comparison work is discouraged. Never include the dead ends or what didn't work. Get into only the elite colleges and betwork at conferences as much as possible - it's not what you know but who you know. Hype up and make your paper as sexy and short as possible. Tell a story. Good teaching, blogging, and sharing software probably hurts you.

This is a longitudinal study, not a cross-sectional one; attrition can be measured. You should also provide some evidence for the idea that hordes of high-IQ people are going homeless and committing suicide, as opposed to occasional ones with mental illness (especially schizophrenia), given that all the existing evidence tends to imply the opposite, if you don't want to come off as a Taleb-citing crank.

I don't think the students at Johns Hopkins are the usual sort of 20-29yos who would knock up a teen girl. And inasmuch as SMPY has been running for so many decades and has thousands of participants, if 10 or 20% of the participants were becoming teen moms, someone probably would've noticed by now.

"Even the quote you cite suggests that the total amount of variance accounted for was about 19%, which isn't too shabby by any means, but is also very crude when you think about implications for real-world consequences."

This is 19% after range restriction. They're making the same point that the graph of quartiles does: even after you set an extremely high bar, differences in the test score are still predicting quite a bit of variance despite all the other possible diluting factors like geography/family/test-error/personality/interest/wealth/opportunity... If anything, it shows that institutions aren't being 'crude' enough - if they were using the test scores optimally and extracting all of the signal, the variance would be 0%.

"How many "pioneering mathematicians" are not one-percenters? If the answer is "99%", then there's no bias in favor of those who do well on the SAT. The article doesn't give that information."

The article doesn't give that information, but unless you believe that being in the top 1% is actively harmful to becoming a pioneering mathematician, the odds ratios of they give for doctorates gives you a good idea, since no one will be a pioneering mathematician without earning a doctorate in math these days. Something like 25% of them have doctorates, compared to the general population which is more like ~1.7%, so simply going by proportion and ignoring the curves in https://my.vanderbilt.edu/smpy/files/2013/02/Ferriman_20101.... and the extreme tail behaviors of these things, at least 0.25*0.01 / 0.017 = 15% of pioneering mathematicians will have been 1%ers. (Take into account the tail, and it'll go up quite a bit.)

"Gifted children are "gifted" with a laser focus on the unusual things that they find fun - math, music, what-have-you. They spend countless hours playing with numbers or with music, while little Johnny is playing with a ball."

If you read Ericsson's papers, he does in fact specifically deny any role to talent and genes, by name, unequivocally. It's not a strawman, it's what he really believes. Deliberate practice is also falsified by Hambrick's meta-analyses: it does not explain much of difference in performance compared to talent. They are not gifted with 'a laser focus', they are gifted with things like intelligence.

You forgot the rest of that quote: "...based on the same method of intensive analysis—brute force, basically—that Deep Blue employed for chess."

I've wondered the same thing. I haven't noted any large improvements in Google Translate despite the RNNs demonstrating big gains for at least 3 years now. However, in all the public comments by Googlers I've read, I'm not sure they have rolled out RNN to Google Translate publically yet - this is the same company that created its own ASICs to save electricity on big deployments because Nvidia isn't moving fast enough.

"Cucumber grading has been automated for years. Look on YouTube for automated cucumber grading systems.[1][2] There are many competing vendors. The commercial machines process their video locally and don't need "the cloud". They're also much faster."

I think OP in a way demonstrates something characteristic of Japanese agriculture:

"There are also some automatic sorters on the market, but they have limitations in terms of performance and cost, and small farms don't tend to use them."

Japanese agriculture is notoriously unproductive compared to other countries' yields and inputs, and a major reason is the lack of scale and mom-and-pop farms. If this farm were bigger or needed to save labor costs more, it could afford the already existing solutions and would sort cucumbers into more standard grades than their own ad hoc system. But they're not, so instead they do it by hand.

It's interesting trying to compare this to Long Now. The article mentions his 'Double Negative' from ~1970 and that it's already badly decayed (a look on Google Maps supports this) after barely 40 years. It doesn't give much detail about the City, but the illustration (photo?) suggests that its most prominent features are mounds of unreinforced concrete. The famous Sandia study on long-term building that Long Now and Steward Brand often cites finds that unreinforced concrete does very poorly over the long run, and that to last, concrete needs to be reinforced and mixed extremely dryly with tons of manual labor (one reason Roman concrete has lasted so long); Heizer is doing this with a skeleton crew and must take shortcuts.

So it sounds like the depressions are going to fill in quickly on a 10,000 year timescale and the exposed concrete is going to break down on the order of centuries. It might last a relatively long time in a buried form like Mayan pyramids, but not as an artwork with integrity, and I doubt much will be left at all by 10,000 years.

Academic journals* rarely host large datasets (it's fairly unusual to even see some PDF supplements with vital summaries or results), and if you want to share data >10MB, you're mostly stuck doing it yourself through your own website or some institution's, or not sharing it at all.

Also, academic datasets aren't free of copyright concerns. Consider the famous Imagenet dataset for image classification. It's made of a million images pulled from Google Images. Did they get each photographs' creators' permission for such unlimited redistribution? Of course not. But there's no way the 'implied license' of posting a photo online extends that far... Like so much of the Internet, it's only possible in the absence of enforcement of copyright law.

* which is particularly frustrating because academic publishers make such enormous profits and hosting large datasets is exactly the sort of thing they should be doing if they were remotely interested in supporting science rather than making more money