The administration wants public feedback to help shape the National Institute of Standards and Technology-led effort.
The Trump administration wants the public to weigh in on standards and tools needed to advance intelligent technology.
The White House Office of Science and Technology Policy seeks insight into developing technical standards around artificial intelligence, according to a request for informationlaunched Wednesday. The National Institute of Standards and Technology will coordinate the RFI and all AI-standards related endeavors, as directed by the February executive orderon AI leadership.
Deputy Assistant to the President for Technology Policy Michael Kratsios said in a statement that the RFI is a direct deliverable set forth by the president’s American AI Initiative.
“The information we receive will be critical to Federal engagement in the development of technical standards for AI and strengthening the public’s trust and confidence in the technology,” Kratsios said.
The executive order on AI directs NIST to issue a set of standards and tools that will guide the government in its adoption of the nascent tech and this RFI marks the beginning of the agency’s development of those standards. NIST said it aims to gain input “through an open process” that envelops both this new RFI and other opportunities, including a public workshop.
Through the comments received from the RFI, NIST ultimately aims to better understand the present state, plans, challenges and opportunities related to the development and availability of AI technical standards and related tools. The agency is also interested in gauging the priority areas for federal involvement in activities related to AI standards and the present and future roles agencies can play in helping develop AI standards and tools to meet America’s needs.
Some of the major areas about which NIST is seeking information include technical standards and guidance needed to advance transparency, privacy and other issues around the trustworthiness of AI tech; the urgency of U.S. need for AI standards; the degree of federal agencies’ current and needed involvement to address the governments’ needs; roadmaps and other documents about plans to develop AI and further information around AI technical standards and tools that have already been developed, as well as information on the organizations that have done so.
The document encourages respondents to define “tools” and “standards” as they wish.
The agency also defines AI technologies and systems broadly, noting in the RFI that they “are considered to be comprised of software and/or hardware that can learn to solve complex problems, make predictions or solve tasks that require human-like sensing (such as vision, speech, and touch), perception, cognition, planning, learning, communication, or physical action.”
“Examples are wide-ranging and expanding rapidly,” it said.
Comments in response to the notice must be sent to NIST via mail or email by May 31. The agency plans to post submissions on its website in the future
Physicists are confident that a quantum computer will soon outperform the world’s most powerful supercomputer. To prove it, they have developed a test that will pit one against the other.
Twice a year, the TOP500 project publishes a ranking of the world’s most powerful computers. The list is eagerly awaited and hugely influential. Global superpowers compete to dominate the rankings, and at the time of writing China looms largest, with 229 devices on the list.
The US has just 121, but this includes the world’s most powerful: the Summit supercomputer at Oak Ridge National Laboratory in Tennessee, which was clocked at 143 petaflops (143 thousand million million floating point operations per second).
The ranking is determined by a benchmarking program called Linpack, which is a collection of Fortran subroutines that solve a range of linear equations. The time taken to solve the equations is a measure of the computer’s speed.
There is no shortage of controversy over this choice of benchmark. Computer architectures are usually optimized to solve specific problems, and many of these are very different from the Linpack challenge. Quantum computers, for example, are entirely unsuited to solving these kinds of problems.
Today we get an answer thanks to the work of Benjamin Villalonga at the Quantum Artificial Intelligence Lab at NASA Ames Research Center in Mountain View, California, and a group of colleagues who have developed a benchmarking test that works on both classical and quantum devices. In this way, it is possible to compare their performance.
What’s more, the team has used the new test to put the Summit, the world’s most powerful supercomputer, through its paces running at 281 petaflops. The result is the benchmark that quantum computers must beat to finally establish their supremacy in the rankings.
Finding a good measure of quantum computing power is no easy task. For a start, computer scientists have long known that quantum computers can outperform their classical counterparts in only a limited number of highly specialized tasks. And even then, no quantum computer is currently powerful enough to perform any of them particularly well because, for example, they are incapable of error correction.
So Villalonga and co looked for a much more fundamental test of quantum computing power that would work equally well for today’s primitive devices and tomorrow’s more advanced quantum machines, and could also be simulated on classical machines.
Their chosen problem is to simulate the evolution of quantum chaos using random quantum circuits. Simple quantum computers can do this because the process does not require powerful error correction, and it is relatively straightforward to filter out results that have been overwhelmed by noise.
It is also straightforward for classical machines to simulate quantum chaos. But the classical computing power required to do this rises exponentially with the number of qubits involved.
Two years ago, physicists determined that quantum computers with at least 50 qubits should achieve quantum supremacy over a classical supercomputer at that time.
But the goalposts are constantly moving as supercomputers are upgraded. For example, Summit is capable of significantly more petaflops now than in the last ranking in November, when it tipped the scales at 143 petaflops. Indeed, Oak Ridge National Labs this week unveiled plans to build a 1.5-exaflop machine by 2021. So being able to continually benchmark these machines against the emerging quantum computers is increasingly important.
Researchers at NASA and Google have created an algorithm called qFlex that simulate random quantum circuits on a classical machine. Last year, they showed that qFlex could simulate and benchmark the performance of a Google quantum computer called Bristlecone, which has 72 qubits. To do this, they used a supercomputer at NASA Ames with 20 petaflops of number-crunching power.
Now they’ve shown that the Summit supercomputer can simulate the performance of a much larger quantum device. “On Summit, we were able to achieve a sustained performance of 281 Pflop/s (single precision) over the entire supercomputer, simulating circuits of 49 and 121 qubits,” they say.
This 121 qubits is beyond the capability of any existing quantum computer. So classical computers remain a hair’s breadth ahead in the rankings.
But this is a race they are destined to lose. Plans are already afoot to build quantum computers with 100+ qubits within the next few years. And as quantum capabilities accelerate, the challenge of building ever more powerful classical machines is already coming up against the buffers.
The limiting factor for new machines is no longer the hardware but the power available to keep them humming. The Summit machine already requires a 14-megawatt power supply. That’s enough to light up an entire a medium-sized town. “To scale such a system by 10x would require 140 MW of power, which would be prohibitively expensive,” say Villalonga and co.
By contrast, quantum computers are frugal. Their main power requirement is the cooling for superconducting components. So a 72-qubit computer like Google’s Bristlecone, for example, requires about 14 kw. “Even as qubit systems scale up, this amount is unlikely to significantly grow,” say Villalonga and co.
So in the efficiency rankings, quantum computers are destined to wipe the floor with their classical counterparts sooner rather than later.
One way or another, quantum supremacy is coming. If this work is anything to go by, the benchmark that will prove it is likely to be qFlex.
Ref: arxiv.org/abs/1905.00444 : Establishing the Quantum Supremacy Frontier with a 281 Pflop/s Simulation
He revealed details about the company’s new rocket, engine, and lunar lander at a private event in Washington, DC, May 9.
Behind curtain number one … Jeff Bezos unveiled Blue Moon, the company’s lunar lander that has been in the works for the past three years. It will be able to land a 6.5-metric-ton payload on the moon’s surface. Watch Blue Origin’s video rendering of what Blue Moon’s lunar landing could look like here.
Who’s hitching a ride? The company announced a number of customers that will fly on Blue Moon, including Airbus, MIT, Johns Hopkins, and Arizona State University.
How are they getting there?New Glenn. This is by far the biggest rocket Blue Origin has ever built. The size will allow the massive Blue Moon lander to fit inside. It will have fewer weather constraints and will be rated to carry humans from the start. As with New Shepard, the company’s suborbital rocket, as well as SpaceX’s rockets, the first stage will be reusable, landing again after completing its mission. The first launch is targeted for 2021.
The new engine: Blue Origin also announced that the new BE-7 engine, which packs 10,000 pounds of thrust, will undergo its first hot fire test this year. This engine will propel Blue Moon
Facebook has been beset by scandals over the last year and many believe that nothing will change until its founder and CEO is gone.
A petition has been launched with one simple objective: to force Mark Zuckerberg to resign as CEO of Facebook.
The campaign group behind it, Fight for the Future, says that although there’s no “silver bullet” to “fix” Facebook, the company cannot address its underlying problems while Zuckerberg remains in charge.
The petition is highly unlikely to succeed, of course. It’s hard to imagine Zuckerberg stepping down voluntarily. And there’s not much Facebook’s board can do either, even if they wanted to. Zuckerberg controls about 60% of all voting shares in Facebook. He’s pretty much untouchable, both as CEO and as board chairman. Despite near-weekly scandals, the company is still growing, and it’s one of the most profitable business ventures in human history.
(Another potential solution, as described in a piece in the New York Timeswritten by one of Facebook’s cofounders, is to break the company up and implement new data privacy regulations in the US.)
Need a reminder as to why everyone is so angry with Facebook and Mark Zuckerberg anyway? Here’s a handy cut-out-and-keep list of just some of the most significant scandals involving the tech giant over the last year or so. (Not to mention all the wider problems of fake news or echo chambers or the decimation of the media. Or dodgy PR practices.)
The high-impact one
Back in March 2018, a whistleblower revealed that political consultancy Cambridge Analytica had collected private information from more than 87 million Facebook profiles without the users’ consent. Facebook let third parties scrape data from applications: in Cambridge Analytica’s case, a personality quiz developed by a Cambridge University academic, Aleksandr Kogan. Mark Zuckerberg responded by admitting “we made mistakes” and promising to restrict data sharing with third-party apps in the future.
In September 2018, Facebook admitted that 50 million users had had their personal information exposed by a hack on its systems. The number was later revised down to 30 million, which still makes it the biggest breach in Facebook’s history.
In March 2019 it turned out Facebook had been storing up to 600 million users’ passwords insecurely since 2012. Just days later, we learned that half a billion Facebook records had been left exposed on the public internet.
The discriminatory advertising practices
Facebook’s ad-serving algorithm automatically discriminates by gender and race, even when no one tells it to. Advertisers can also explicitly discriminateagainst certain areas when showing housing ads on Facebook, even though it’s illegal. Facebook has known about this problem since 2016. It still hasn’t fixed it.
The dodgy data deals
Facebook gave over 150 companies more intrusive access to users’ data than previously revealed, via special partnerships. We learned a bit more about this, and other dodgy data practices, in a cache of documents seized by the UK Parliament in November 2018. Facebook expects to be fined up to $5 billion for this and other instances of malpractice.
The vehicle for hate speech
The Christchurch, New Zealand, shooter used Facebook to live-stream his murder of 50 people. The broadcast was up for 20 minutes before any action was taken. We’re still waiting to hear what, if anything, Facebook will do about this issue (for example, it could choose to end its “Facebook Live” feature). It’s well established now that Facebook can help to fuel violence in the real world. But any response from Facebook has been piecemeal. It’s also a reminder of just how much power we’ve given Facebook (and its low-paid moderators) to decide what is and isn’t acceptable.
We’ve been wasting our processing power to train neural networks that are ten times too big.
Neural networks are the core software of deep learning. Even though they’re so widespread, however, they’re really poorly understood. Researchers have observed their emergent properties without actually understanding why they work the way they do.
Now a new paper out of MIT has taken a major step toward answering this question. And in the process the researchers have made a simple but dramatic discovery: we’ve been using neural networks far bigger than we actually need. In some cases they’re 10—even 100—times bigger, so training them costs us orders of magnitude more time and computational power than necessary.
Put another way, within every neural network exists a far smaller one that can be trained to achieve the same performance as its oversize parent. This isn’t just exciting news for AI researchers. The finding has the potential to unlock new applications—some of which we can’t yet fathom—that could improve our day-to-day lives. More on that later.
But first, let’s dive into how neural networks work to understand why this is possible.
A diagram of a neural network learning to recognize a lion.
JEFF CLUNE/SCREENSHOT
How neural networks work
You may have seen neural networks depicted in diagrams like the one above: they’re composed of stacked layers of simple computational nodes that are connected in order to compute patterns in data.
The connections are what’s important. Before a neural network is trained, these connections are assigned random values between 0 and 1 that represent their intensity. (This is called the “initialization” process.) During training, as the network is fed a series of, say, animal photos, it tweaks and tunes those intensities—sort of like the way your brain strengthens or weakens different neuron connections as you accumulate experience and knowledge. After training, the final connection intensities are then used in perpetuity to recognize animals in new photos.
While the mechanics of neural networks are well understood, the reason they work the way they do has remained a mystery. Through lots of experimentation, however, researchers have observed two properties of neural networks that have proved useful.
Observation #1. When a network is initialized before the training process, there’s always some likelihood that the randomly assigned connection strengths end up in an untrainable configuration. In other words, no matter how many animal photos you feed the neural network, it won’t achieve a decent performance, and you just have to reinitialize it to a new configuration. The larger the network (the more layers and nodes it has), the less likely that is. Whereas a tiny neural network may be trainable in only one of every five initializations, a larger network may be trainable in four of every five. Again, why this happens had been a mystery, but that’s why researchers typically use very large networks for their deep-learning tasks. They want to increase their chances of achieving a successful model.
Observation #2. The consequence is that a neural network usually starts off bigger than it needs to be. Once it’s done training, typically only a fraction of its connections remain strong, while the others end up pretty weak—so weak that you can actually delete, or “prune,” them without affecting the network’s performance.
For many years now, researchers have exploited this second observation to shrink their networks after training to lower the time and computational costs involved in running them. But no one thought it was possible to shrink their networks before training. It was assumed that you had to start with an oversize network and the training process had to run its course in order to separate the relevant connections from the irrelevant ones.
Jonathan Frankle, the MIT PhD student who coauthored the paper, questioned that assumption. “If you need way fewer connections than what you started with,” he says, “why can’t we just train the smaller network without the extra connections?” Turns out you can.
Michael Carbin (left) and Jonathan Frankle (right), the authors of the paper.
JASON DORFMAN, MIT CSAIL
The lottery ticket hypothesis
The discovery hinges on the reality that the random connection strengths assigned during initialization aren’t, in fact, random in their consequences: they predispose different parts of the network to fail or succeed before training even happens. Put another way, the initial configuration influences which final configuration the network will arrive at.
By focusing on this idea, the researchers found that if you prune an oversize network after training, you can actually reuse the resultant smaller network to train on new data and preserve high performance—as long as you reset each connection within this downsized network back to its initial strength.
From this finding, Frankle and his coauthor Michael Carbin, an assistant professor at MIT, propose what they call the “lottery ticket hypothesis.” When you randomly initialize a neural network’s connection strengths, it’s almost like buying a bag of lottery tickets. Within your bag, you hope, is a winning ticket—i.e., an initial configuration that will be easy to train and result in a successful model.
This also explains why observation #1 holds true. Starting with a larger network is like buying more lottery tickets. You’re not increasing the amount of power that you’re throwing at your deep-learning problem; you’re simply increasing the likelihood that you will have a winning configuration. Once you find the winning configuration, you should be able to reuse it again and again, rather than continue to replay the lottery.
Next steps
This raises a lot of questions. First, how do you find the winning ticket? In their paper, Frankle and Carbin took a brute-force approach of training and pruning an oversize network with one data set to extract the winning ticket for another data set. In theory, there should be much more efficient ways of finding—or even designing—a winning configuration from the start.
Second, what are the training limits of a winning configuration? Presumably, different kinds of data and different deep-learning tasks would require different configurations.
Third, what is the smallest possible neural network that you can get away with while still achieving high performance? Frankle found that through an iterative training and pruning process, he was able to consistently reduce the starting network to between 10% and 20% of its original size. But he thinks there’s a chance for it to be even smaller.
Already, many research teams within the AI community have begun to conduct follow-up work. A researcher at Princeton recently teased the results of a forthcoming paper addressing the second question. A team at Uber also published a new paper on several experiments investigating the nature of the metaphorical lottery tickets. Most surprising, they found that once a winning configuration has been found, it already achieves significantly better performance than the original untrained oversize network before any training whatsoever. In other words, the act of pruning a network to extract a winning configuration is itself an important method of training.
Neural network nirvana
Frankle imagines a future where the research community will have an open-source database of all the different configurations they’ve found, with descriptions for what tasks they’re good for. He jokingly calls this “neural network nirvana.” He believes it would dramatically accelerate and democratize AI research by lowering the cost and speed of training, and by allowing people without giant data servers to do this work directly on small laptops or even mobile phones.
It could also change the nature of AI applications. If you can train a neural network locally on a device instead of in the cloud, you can improve the speed of the training process and the security of the data. Imagine a machine-learning-based medical device, for example, that could improve itself through use without needing to send patient data to Google’s or Amazon’s servers.
“We’re constantly bumping up against the edge of what we can train,” says Jason Yosinski, a founding member of Uber AI Labs who coauthored the follow-up Uber paper, “meaning the biggest networks you can fit on a GPU or the longest we can tolerate waiting before we get a result back.” If researchers could figure out how to identify winning configurations from the get-go, it would reduce the size of neural networks by a factor of 10, even 100. The ceiling of possibility would dramatically increase, opening a new world of potential uses.
