Tag Archives: Deep learning
Making sure artificial intelligence does what we want and behaves in predictable ways will be crucial as the technology becomes increasingly ubiquitous. It’s an area frequently neglected in the race to develop products, but DeepMind has now outlined its research agenda to tackle the problem.
AI safety, as the field is known, has been gaining prominence in recent years. That’s probably at least partly down to the overzealous warnings of a coming AI apocalypse from well-meaning, but underqualified pundits like Elon Musk and Stephen Hawking. But it’s also recognition of the fact that AI technology is quickly pervading all aspects of our lives, making decisions on everything from what movies we watch to whether we get a mortgage.
That’s why DeepMind hired a bevy of researchers who specialize in foreseeing the unforeseen consequences of the way we built AI back in 2016. And now the team has spelled out the three key domains they think require research if we’re going to build autonomous machines that do what we want.
In a new blog designed to provide updates on the team’s work, they introduce the ideas of specification, robustness, and assurance, which they say will act as the cornerstones of their future research. Specification involves making sure AI systems do what their operator intends; robustness means a system can cope with changes to its environment and attempts to throw it off course; and assurance involves our ability to understand what systems are doing and how to control them.
A classic thought experiment designed to illustrate how we could lose control of an AI system can help illustrate the problem of specification. Philosopher Nick Bostrom’s posited a hypothetical machine charged with making as many paperclips as possible. Because the creators fail to add what they might assume are obvious additional goals like not harming people, the AI wipes out humanity so we can’t switch it off before turning all matter in the universe into paperclips.
Obviously the example is extreme, but it shows how a poorly-specified goal can lead to unexpected and disastrous outcomes. Properly codifying the desires of the designer is no easy feat, though; often there are not neat ways to encompass both the explicit and implicit goals in ways that are understandable to the machine and don’t leave room for ambiguities, meaning we often rely on incomplete approximations.
The researchers note recent research by OpenAI in which an AI was trained to play a boat-racing game called CoastRunners. The game rewards players for hitting targets laid out along the race route. The AI worked out that it could get a higher score by repeatedly knocking over regenerating targets rather than actually completing the course. The blog post includes a link to a spreadsheet detailing scores of such examples.
Another key concern for AI designers is making their creation robust to the unpredictability of the real world. Despite their superhuman abilities on certain tasks, most cutting-edge AI systems are remarkably brittle. They tend to be trained on highly-curated datasets and so can fail when faced with unfamiliar input. This can happen by accident or by design—researchers have come up with numerous ways to trick image recognition algorithms into misclassifying things, including thinking a 3D printed tortoise was actually a gun.
Building systems that can deal with every possible encounter may not be feasible, so a big part of making AIs more robust may be getting them to avoid risks and ensuring they can recover from errors, or that they have failsafes to ensure errors don’t lead to catastrophic failure.
And finally, we need to have ways to make sure we can tell whether an AI is performing the way we expect it to. A key part of assurance is being able to effectively monitor systems and interpret what they’re doing—if we’re basing medical treatments or sentencing decisions on the output of an AI, we’d like to see the reasoning. That’s a major outstanding problem for popular deep learning approaches, which are largely indecipherable black boxes.
The other half of assurance is the ability to intervene if a machine isn’t behaving the way we’d like. But designing a reliable off switch is tough, because most learning systems have a strong incentive to prevent anyone from interfering with their goals.
The authors don’t pretend to have all the answers, but they hope the framework they’ve come up with can help guide others working on AI safety. While it may be some time before AI is truly in a position to do us harm, hopefully early efforts like these will mean it’s built on a solid foundation that ensures it is aligned with our goals.
Image Credit: cono0430 / Shutterstock.com Continue reading
In 2017, artificial intelligence attracted $12 billion of VC investment. We are only beginning to discover the usefulness of AI applications. Amazon recently unveiled a brick-and-mortar grocery store that has successfully supplanted cashiers and checkout lines with computer vision, sensors, and deep learning. Between the investment, the press coverage, and the dramatic innovation, “AI” has become a hot buzzword. But does it even exist yet?
At the World Economic Forum Dr. Kai-Fu Lee, a Taiwanese venture capitalist and the founding president of Google China, remarked, “I think it’s tempting for every entrepreneur to package his or her company as an AI company, and it’s tempting for every VC to want to say ‘I’m an AI investor.’” He then observed that some of these AI bubbles could burst by the end of 2018, referring specifically to “the startups that made up a story that isn’t fulfillable, and fooled VCs into investing because they don’t know better.”
However, Dr. Lee firmly believes AI will continue to progress and will take many jobs away from workers. So, what is the difference between legitimate AI, with all of its pros and cons, and a made-up story?
If you parse through just a few stories that are allegedly about AI, you’ll quickly discover significant variation in how people define it, with a blurred line between emulated intelligence and machine learning applications.
I spoke to experts in the field of AI to try to find consensus, but the very question opens up more questions. For instance, when is it important to be accurate to a term’s original definition, and when does that commitment to accuracy amount to the splitting of hairs? It isn’t obvious, and hype is oftentimes the enemy of nuance. Additionally, there is now a vested interest in that hype—$12 billion, to be precise.
This conversation is also relevant because world-renowned thought leaders have been publicly debating the dangers posed by AI. Facebook CEO Mark Zuckerberg suggested that naysayers who attempt to “drum up these doomsday scenarios” are being negative and irresponsible. On Twitter, business magnate and OpenAI co-founder Elon Musk countered that Zuckerberg’s understanding of the subject is limited. In February, Elon Musk engaged again in a similar exchange with Harvard professor Steven Pinker. Musk tweeted that Pinker doesn’t understand the difference between functional/narrow AI and general AI.
Given the fears surrounding this technology, it’s important for the public to clearly understand the distinctions between different levels of AI so that they can realistically assess the potential threats and benefits.
As Smart As a Human?
Erik Cambria, an expert in the field of natural language processing, told me, “Nobody is doing AI today and everybody is saying that they do AI because it’s a cool and sexy buzzword. It was the same with ‘big data’ a few years ago.”
Cambria mentioned that AI, as a term, originally referenced the emulation of human intelligence. “And there is nothing today that is even barely as intelligent as the most stupid human being on Earth. So, in a strict sense, no one is doing AI yet, for the simple fact that we don’t know how the human brain works,” he said.
He added that the term “AI” is often used in reference to powerful tools for data classification. These tools are impressive, but they’re on a totally different spectrum than human cognition. Additionally, Cambria has noticed people claiming that neural networks are part of the new wave of AI. This is bizarre to him because that technology already existed fifty years ago.
However, technologists no longer need to perform the feature extraction by themselves. They also have access to greater computing power. All of these advancements are welcomed, but it is perhaps dishonest to suggest that machines have emulated the intricacies of our cognitive processes.
“Companies are just looking at tricks to create a behavior that looks like intelligence but that is not real intelligence, it’s just a mirror of intelligence. These are expert systems that are maybe very good in a specific domain, but very stupid in other domains,” he said.
This mimicry of intelligence has inspired the public imagination. Domain-specific systems have delivered value in a wide range of industries. But those benefits have not lifted the cloud of confusion.
Assisted, Augmented, or Autonomous
When it comes to matters of scientific integrity, the issue of accurate definitions isn’t a peripheral matter. In a 1974 commencement address at the California Institute of Technology, Richard Feynman famously said, “The first principle is that you must not fool yourself—and you are the easiest person to fool.” In that same speech, Feynman also said, “You should not fool the layman when you’re talking as a scientist.” He opined that scientists should bend over backwards to show how they could be wrong. “If you’re representing yourself as a scientist, then you should explain to the layman what you’re doing—and if they don’t want to support you under those circumstances, then that’s their decision.”
In the case of AI, this might mean that professional scientists have an obligation to clearly state that they are developing extremely powerful, controversial, profitable, and even dangerous tools, which do not constitute intelligence in any familiar or comprehensive sense.
The term “AI” may have become overhyped and confused, but there are already some efforts underway to provide clarity. A recent PwC report drew a distinction between “assisted intelligence,” “augmented intelligence,” and “autonomous intelligence.” Assisted intelligence is demonstrated by the GPS navigation programs prevalent in cars today. Augmented intelligence “enables people and organizations to do things they couldn’t otherwise do.” And autonomous intelligence “establishes machines that act on their own,” such as autonomous vehicles.
Roman Yampolskiy is an AI safety researcher who wrote the book “Artificial Superintelligence: A Futuristic Approach.” I asked him whether the broad and differing meanings might present difficulties for legislators attempting to regulate AI.
Yampolskiy explained, “Intelligence (artificial or natural) comes on a continuum and so do potential problems with such technology. We typically refer to AI which one day will have the full spectrum of human capabilities as artificial general intelligence (AGI) to avoid some confusion. Beyond that point it becomes superintelligence. What we have today and what is frequently used in business is narrow AI. Regulating anything is hard, technology is no exception. The problem is not with terminology but with complexity of such systems even at the current level.”
When asked if people should fear AI systems, Dr. Yampolskiy commented, “Since capability comes on a continuum, so do problems associated with each level of capability.” He mentioned that accidents are already reported with AI-enabled products, and as the technology advances further, the impact could spread beyond privacy concerns or technological unemployment. These concerns about the real-world effects of AI will likely take precedence over dictionary-minded quibbles. However, the issue is also about honesty versus deception.
Is This Buzzword All Buzzed Out?
Finally, I directed my questions towards a company that is actively marketing an “AI Virtual Assistant.” Carl Landers, the CMO at Conversica, acknowledged that there are a multitude of explanations for what AI is and isn’t.
He said, “My definition of AI is technology innovation that helps solve a business problem. I’m really not interested in talking about the theoretical ‘can we get machines to think like humans?’ It’s a nice conversation, but I’m trying to solve a practical business problem.”
I asked him if AI is a buzzword that inspires publicity and attracts clients. According to Landers, this was certainly true three years ago, but those effects have already started to wane. Many companies now claim to have AI in their products, so it’s less of a differentiator. However, there is still a specific intention behind the word. Landers hopes to convey that previously impossible things are now possible. “There’s something new here that you haven’t seen before, that you haven’t heard of before,” he said.
According to Brian Decker, founder of Encom Lab, machine learning algorithms only work to satisfy their preexisting programming, not out of an interior drive for better understanding. Therefore, he views AI as an entirely semantic argument.
Decker stated, “A marketing exec will claim a photodiode controlled porch light has AI because it ‘knows when it is dark outside,’ while a good hardware engineer will point out that not one bit in a register in the entire history of computing has ever changed unless directed to do so according to the logic of preexisting programming.”
Although it’s important for everyone to be on the same page regarding specifics and underlying meaning, AI-powered products are already powering past these debates by creating immediate value for humans. And ultimately, humans care more about value than they do about semantic distinctions. In an interview with Quartz, Kai-Fu Lee revealed that algorithmic trading systems have already given him an 8X return over his private banking investments. “I don’t trade with humans anymore,” he said.
Image Credit: vrender / Shutterstock.com Continue reading
New planets found in distant corners of the galaxy. Climate models that may improve our understanding of sea level rise. The emergence of new antimalarial drugs. These scientific advances and discoveries have been in the news in recent months.
While representing wildly divergent disciplines, from astronomy to biotechnology, they all have one thing in common: Artificial intelligence played a key role in their scientific discovery.
One of the more recent and famous examples came out of NASA at the end of 2017. The US space agency had announced an eighth planet discovered in the Kepler-90 system. Scientists had trained a neural network—a computer with a “brain” modeled on the human mind—to re-examine data from Kepler, a space-borne telescope with a four-year mission to seek out new life and new civilizations. Or, more precisely, to find habitable planets where life might just exist.
The researchers trained the artificial neural network on a set of 15,000 previously vetted signals until it could identify true planets and false positives 96 percent of the time. It then went to work on weaker signals from nearly 700 star systems with known planets.
The machine detected Kepler 90i—a hot, rocky planet that orbits its sun about every two Earth weeks—through a nearly imperceptible change in brightness captured when a planet passes a star. It also found a sixth Earth-sized planet in the Kepler-80 system.
AI Handles Big Data
The application of AI to science is being driven by three great advances in technology, according to Ross King from the Manchester Institute of Biotechnology at the University of Manchester, leader of a team that developed an artificially intelligent “scientist” called Eve.
Those three advances include much faster computers, big datasets, and improved AI methods, King said. “These advances increasingly give AI superhuman reasoning abilities,” he told Singularity Hub by email.
AI systems can flawlessly remember vast numbers of facts and extract information effortlessly from millions of scientific papers, not to mention exhibit flawless logical reasoning and near-optimal probabilistic reasoning, King says.
AI systems also beat humans when it comes to dealing with huge, diverse amounts of data.
That’s partly what attracted a team of glaciologists to turn to machine learning to untangle the factors involved in how heat from Earth’s interior might influence the ice sheet that blankets Greenland.
Algorithms juggled 22 geologic variables—such as bedrock topography, crustal thickness, magnetic anomalies, rock types, and proximity to features like trenches, ridges, young rifts, and volcanoes—to predict geothermal heat flux under the ice sheet throughout Greenland.
The machine learning model, for example, predicts elevated heat flux upstream of Jakobshavn Glacier, the fastest-moving glacier in the world.
“The major advantage is that we can incorporate so many different types of data,” explains Leigh Stearns, associate professor of geology at Kansas University, whose research takes her to the polar regions to understand how and why Earth’s great ice sheets are changing, questions directly related to future sea level rise.
“All of the other models just rely on one parameter to determine heat flux, but the [machine learning] approach incorporates all of them,” Stearns told Singularity Hub in an email. “Interestingly, we found that there is not just one parameter…that determines the heat flux, but a combination of many factors.”
The research was published last month in Geophysical Research Letters.
Stearns says her team hopes to apply high-powered machine learning to characterize glacier behavior over both short and long-term timescales, thanks to the large amounts of data that she and others have collected over the last 20 years.
Emergence of Robot Scientists
While Stearns sees machine learning as another tool to augment her research, King believes artificial intelligence can play a much bigger role in scientific discoveries in the future.
“I am interested in developing AI systems that autonomously do science—robot scientists,” he said. Such systems, King explained, would automatically originate hypotheses to explain observations, devise experiments to test those hypotheses, physically run the experiments using laboratory robotics, and even interpret the results. The conclusions would then influence the next cycle of hypotheses and experiments.
His AI scientist Eve recently helped researchers discover that triclosan, an ingredient commonly found in toothpaste, could be used as an antimalarial drug against certain strains that have developed a resistance to other common drug therapies. The research was published in the journal Scientific Reports.
Automation using artificial intelligence for drug discovery has become a growing area of research, as the machines can work orders of magnitude faster than any human. AI is also being applied in related areas, such as synthetic biology for the rapid design and manufacture of microorganisms for industrial uses.
King argues that machines are better suited to unravel the complexities of biological systems, with even the most “simple” organisms are host to thousands of genes, proteins, and small molecules that interact in complicated ways.
“Robot scientists and semi-automated AI tools are essential for the future of biology, as there are simply not enough human biologists to do the necessary work,” he said.
Creating Shockwaves in Science
The use of machine learning, neural networks, and other AI methods can often get better results in a fraction of the time it would normally take to crunch data.
For instance, scientists at the National Center for Supercomputing Applications, located at the University of Illinois at Urbana-Champaign, have a deep learning system for the rapid detection and characterization of gravitational waves. Gravitational waves are disturbances in spacetime, emanating from big, high-energy cosmic events, such as the massive explosion of a star known as a supernova. The “Holy Grail” of this type of research is to detect gravitational waves from the Big Bang.
Dubbed Deep Filtering, the method allows real-time processing of data from LIGO, a gravitational wave observatory comprised of two enormous laser interferometers located thousands of miles apart in California and Louisiana. The research was published in Physics Letters B. You can watch a trippy visualization of the results below.
In a more down-to-earth example, scientists published a paper last month in Science Advances on the development of a neural network called ConvNetQuake to detect and locate minor earthquakes from ground motion measurements called seismograms.
ConvNetQuake uncovered 17 times more earthquakes than traditional methods. Scientists say the new method is particularly useful in monitoring small-scale seismic activity, which has become more frequent, possibly due to fracking activities that involve injecting wastewater deep underground. You can learn more about ConvNetQuake in this video:
King says he believes that in the long term there will be no limit to what AI can accomplish in science. He and his team, including Eve, are currently working on developing cancer therapies under a grant from DARPA.
“Robot scientists are getting smarter and smarter; human scientists are not,” he says. “Indeed, there is arguably a case that human scientists are less good. I don’t see any scientist alive today of the stature of a Newton or Einstein—despite the vast number of living scientists. The Physics Nobel [laureate] Frank Wilczek is on record as saying (10 years ago) that in 100 years’ time the best physicist will be a machine. I agree.”
Image Credit: Romaset / Shutterstock.com Continue reading
Move over, deep learning. Neuromorphic computing—the next big thing in artificial intelligence—is on fire.
Just last week, two studies individually unveiled computer chips modeled after information processing in the human brain.
The first, published in Nature Materials, found a perfect solution to deal with unpredictability at synapses—the gap between two neurons that transmit and store information. The second, published in Science Advances, further amped up the system’s computational power, filling synapses with nanoclusters of supermagnetic material to bolster information encoding.
The result? Brain-like hardware systems that compute faster—and more efficiently—than the human brain.
“Ultimately we want a chip as big as a fingernail to replace one big supercomputer,” said Dr. Jeehwan Kim, who led the first study at MIT in Cambridge, Massachusetts.
Experts are hopeful.
“The field’s full of hype, and it’s nice to see quality work presented in an objective way,” said Dr. Carver Mead, an engineer at the California Institute of Technology in Pasadena not involved in the work.
Software to Hardware
The human brain is the ultimate computational wizard. With roughly 100 billion neurons densely packed into the size of a small football, the brain can deftly handle complex computation at lightning speed using very little energy.
AI experts have taken note. The past few years saw brain-inspired algorithms that can identify faces, falsify voices, and play a variety of games at—and often above—human capability.
But software is only part of the equation. Our current computers, with their transistors and binary digital systems, aren’t equipped to run these powerful algorithms.
That’s where neuromorphic computing comes in. The idea is simple: fabricate a computer chip that mimics the brain at the hardware level. Here, data is both processed and stored within the chip in an analog manner. Each artificial synapse can accumulate and integrate small bits of information from multiple sources and fire only when it reaches a threshold—much like its biological counterpart.
Experts believe the speed and efficiency gains will be enormous.
For one, the chips will no longer have to transfer data between the central processing unit (CPU) and storage blocks, which wastes both time and energy. For another, like biological neural networks, neuromorphic devices can support neurons that run millions of streams of parallel computation.
Optimism aside, reproducing the biological synapse in hardware form hasn’t been as easy as anticipated.
Neuromorphic chips exist in many forms, but often look like a nanoscale metal sandwich. The “bread” pieces are generally made of conductive plates surrounding a switching medium—a conductive material of sorts that acts like the gap in a biological synapse.
When a voltage is applied, as in the case of data input, ions move within the switching medium, which then creates conductive streams to stimulate the downstream plate. This change in conductivity mimics the way biological neurons change their “weight,” or the strength of connectivity between two adjacent neurons.
But so far, neuromorphic synapses have been rather unpredictable. According to Kim, that’s because the switching medium is often comprised of material that can’t channel ions to exact locations on the downstream plate.
“Once you apply some voltage to represent some data with your artificial neuron, you have to erase and be able to write it again in the exact same way,” explains Kim. “But in an amorphous solid, when you write again, the ions go in different directions because there are lots of defects.”
In his new study, Kim and colleagues swapped the jelly-like switching medium for silicon, a material with only a single line of defects that acts like a channel to guide ions.
The chip starts with a thin wafer of silicon etched with a honeycomb-like pattern. On top is a layer of silicon germanium—something often present in transistors—in the same pattern. This creates a funnel-like dislocation, a kind of Grand Canal that perfectly shuttles ions across the artificial synapse.
The researchers then made a neuromorphic chip containing these synapses and shot an electrical zap through them. Incredibly, the synapses’ response varied by only four percent—much higher than any neuromorphic device made with an amorphous switching medium.
In a computer simulation, the team built a multi-layer artificial neural network using parameters measured from their device. After tens of thousands of training examples, their neural network correctly recognized samples 95 percent of the time, just 2 percent lower than state-of-the-art software algorithms.
The upside? The neuromorphic chip requires much less space than the hardware that runs deep learning algorithms. Forget supercomputers—these chips could one day run complex computations right on our handheld devices.
A Magnetic Boost
Meanwhile, in Boulder, Colorado, Dr. Michael Schneider at the National Institute of Standards and Technology also realized that the standard switching medium had to go.
“There must be a better way to do this, because nature has figured out a better way to do this,” he says.
His solution? Nanoclusters of magnetic manganese.
Schneider’s chip contained two slices of superconducting electrodes made out of niobium, which channel electricity with no resistance. When researchers applied different magnetic fields to the synapse, they could control the alignment of the manganese “filling.”
The switch gave the chip a double boost. For one, by aligning the switching medium, the team could predict the ion flow and boost uniformity. For another, the magnetic manganese itself adds computational power. The chip can now encode data in both the level of electrical input and the direction of the magnetisms without bulking up the synapse.
It seriously worked. At one billion times per second, the chips fired several orders of magnitude faster than human neurons. Plus, the chips required just one ten-thousandth of the energy used by their biological counterparts, all the while synthesizing input from nine different sources in an analog manner.
The Road Ahead
These studies show that we may be nearing a benchmark where artificial synapses match—or even outperform—their human inspiration.
But to Dr. Steven Furber, an expert in neuromorphic computing, we still have a ways before the chips go mainstream.
Many of the special materials used in these chips require specific temperatures, he says. Magnetic manganese chips, for example, require temperatures around absolute zero to operate, meaning they come with the need for giant cooling tanks filled with liquid helium—obviously not practical for everyday use.
Another is scalability. Millions of synapses are necessary before a neuromorphic device can be used to tackle everyday problems such as facial recognition. So far, no deal.
But these problems may in fact be a driving force for the entire field. Intense competition could push teams into exploring different ideas and solutions to similar problems, much like these two studies.
If so, future chips may come in diverse flavors. Similar to our vast array of deep learning algorithms and operating systems, the computer chips of the future may also vary depending on specific requirements and needs.
It is worth developing as many different technological approaches as possible, says Furber, especially as neuroscientists increasingly understand what makes our biological synapses—the ultimate inspiration—so amazingly efficient.
Image Credit: arakio / Shutterstock.com Continue reading