Tag Archives: Deep learning
When every other tech startup claims to use artificial intelligence, it can be tough to figure out if an AI service or product works as advertised. In the midst of the AI “gold rush,” how can you separate the nuggets from the fool’s gold?
There’s no shortage of cautionary tales involving overhyped AI claims. And applying AI technologies to health care, education, and law enforcement mean that getting it wrong can have real consequences for society—not just for investors who bet on the wrong unicorn.
So IEEE Spectrum asked experts to share their tips for how to identify AI hype in press releases, news articles, research papers, and IPO filings.
“It can be tricky, because I think the people who are out there selling the AI hype—selling this AI snake oil—are getting more sophisticated over time,” says Tim Hwang, director of the Harvard-MIT Ethics and Governance of AI Initiative.
The term “AI” is perhaps most frequently used to describe machine learning algorithms (and deep learning algorithms, which require even less human guidance) that analyze huge amounts of data and make predictions based on patterns that humans might miss. These popular forms of AI are mostly suited to specialized tasks, such as automatically recognizing certain objects within photos. For that reason, they are sometimes described as “weak” or “narrow” AI.
Some researchers and thought leaders like to talk about the idea of “artificial general intelligence” or “strong AI” that has human-level capacity and flexibility to handle many diverse intellectual tasks. But for now, this type of AI remains firmly in the realm of science fiction and is far from being realized in the real world.
“AI has no well-defined meaning and many so-called AI companies are simply trying to take advantage of the buzz around that term,” says Arvind Narayanan, a computer scientist at Princeton University. “Companies have even been caught claiming to use AI when, in fact, the task is done by human workers.”
Here are three ways to recognize AI hype.
Look for Buzzwords
One red flag is what Hwang calls the “hype salad.” This means stringing together the term “AI” with many other tech buzzwords such as “blockchain” or “Internet of Things.” That doesn’t automatically disqualify the technology, but spotting a high volume of buzzwords in a post, pitch, or presentation should raise questions about what exactly the company or individual has developed.
Other experts agree that strings of buzzwords can be a red flag. That’s especially true if the buzzwords are never really explained in technical detail, and are simply tossed around as vague, poorly-defined terms, says Marzyeh Ghassemi, a computer scientist and biomedical engineer at the University of Toronto in Canada.
“I think that if it looks like a Google search—picture ‘interpretable blockchain AI deep learning medicine’—it's probably not high-quality work,” Ghassemi says.
Hwang also suggests mentally replacing all mentions of “AI” in an article with the term “magical fairy dust.” It’s a way of seeing whether an individual or organization is treating the technology like magic. If so—that’s another good reason to ask more questions about what exactly the AI technology involves.
And even the visual imagery used to illustrate AI claims can indicate that an individual or organization is overselling the technology.
“I think that a lot of the people who work on machine learning on a day-to-day basis are pretty humble about the technology, because they’re largely confronted with how frequently it just breaks and doesn't work,” Hwang says. “And so I think that if you see a company or someone representing AI as a Terminator head, or a big glowing HAL eye or something like that, I think it’s also worth asking some questions.”
Interrogate the Data
It can be hard to evaluate AI claims without any relevant expertise, says Ghassemi at the University of Toronto. Even experts need to know the technical details of the AI algorithm in question and have some access to the training data that shaped the AI model’s predictions. Still, savvy readers with some basic knowledge of applied statistics can search for red flags.
To start, readers can look for possible bias in training data based on small sample sizes or a skewed population that fails to reflect the broader population, Ghassemi says. After all, an AI model trained only on health data from white men would not necessarily achieve similar results for other populations of patients.
“For me, a red flag is not demonstrating deep knowledge of how your labels are defined.”
—Marzyeh Ghassemi, University of Toronto
How machine learning and deep learning models perform also depends on how well humans labeled the sample datasets use to train these programs. This task can be straightforward when labeling photos of cats versus dogs, but gets more complicated when assigning disease diagnoses to certain patient cases.
Medical experts frequently disagree with each other on diagnoses—which is why many patients seek a second opinion. Not surprisingly, this ambiguity can also affect the diagnostic labels that experts assign in training datasets. “For me, a red flag is not demonstrating deep knowledge of how your labels are defined,” Ghassemi says.
Such training data can also reflect the cultural stereotypes and biases of the humans who labeled the data, says Narayanan at Princeton University. Like Ghassemi, he recommends taking a hard look at exactly what the AI has learned: “A good way to start critically evaluating AI claims is by asking questions about the training data.”
Another red flag is presenting an AI system’s performance through a single accuracy figure without much explanation, Narayanan says. Claiming that an AI model achieves “99 percent” accuracy doesn’t mean much without knowing the baseline for comparison—such as whether other systems have already achieved 99 percent accuracy—or how well that accuracy holds up in situations beyond the training dataset.
Narayanan also emphasized the need to ask questions about an AI model’s false positive rate—the rate of making wrong predictions about the presence of a given condition. Even if the false positive rate of a hypothetical AI service is just one percent, that could have major consequences if that service ends up screening millions of people for cancer.
Readers can also consider whether using AI in a given situation offers any meaningful improvement compared to traditional statistical methods, says Clayton Aldern, a data scientist and journalist who serves as managing director for Caldern LLC. He gave the hypothetical example of a “super-duper-fancy deep learning model” that achieves a prediction accuracy of 89 percent, compared to a “little polynomial regression model” that achieves 86 percent on the same dataset.
“We're talking about a three-percentage-point increase on something that you learned about in Algebra 1,” Aldern says. “So is it worth the hype?”
Don’t Ignore the Drawbacks
The hype surrounding AI isn’t just about the technical merits of services and products driven by machine learning. Overblown claims about the beneficial impacts of AI technology—or vague promises to address ethical issues related to deploying it—should also raise red flags.
“If a company promises to use its tech ethically, it is important to question if its business model aligns with that promise,” Narayanan says. “Even if employees have noble intentions, it is unrealistic to expect the company as a whole to resist financial imperatives.”
One example might be a company with a business model that depends on leveraging customers’ personal data. Such companies “tend to make empty promises when it comes to privacy,” Narayanan says. And, if companies hire workers to produce training data, it’s also worth asking whether the companies treat those workers ethically.
The transparency—or lack thereof—about any AI claim can also be telling. A company or research group can minimize concerns by publishing technical claims in peer-reviewed journals or allowing credible third parties to evaluate their AI without giving away big intellectual property secrets, Narayanan says. Excessive secrecy is a big red flag.
With these strategies, you don’t need to be a computer engineer or data scientist to start thinking critically about AI claims. And, Narayanan says, the world needs many people from different backgrounds for societies to fully consider the real-world implications of AI.
Editor’s Note: The original version of this story misspelled Clayton Aldern’s last name as Alderton. Continue reading
Discovering a new drug can take decades, billions of dollars, and untold man hours from some of the smartest people on the planet. Now a startup says it’s taken a significant step towards speeding the process up using AI.
The typical drug discovery process involves carrying out physical tests on enormous libraries of molecules, and even with the help of robotics it’s an arduous process. The idea of sidestepping this by using computers to virtually screen for promising candidates has been around for decades. But progress has been underwhelming, and it’s still not a major part of commercial pipelines.
Recent advances in deep learning, however, have reignited hopes for the field, and major pharma companies have started tying up with AI-powered drug discovery startups. And now Insilico Medicine has used AI to design a molecule that effectively targets a protein involved in fibrosis—the formation of excess fibrous tissue—in mice in just 46 days.
The platform the company has developed combines two of the hottest sub-fields of AI: the generative adversarial networks, or GANs, which power deepfakes, and reinforcement learning, which is at the heart of the most impressive game-playing AI advances of recent years.
In a paper in Nature Biotechnology, the company’s researchers describe how they trained their model on all the molecules already known to target this protein as well as many other active molecules from various datasets. The model was then used to generate 30,000 candidate molecules.
Unlike most previous efforts, they went a step further and selected the most promising molecules for testing in the lab. The 30,000 candidates were whittled down to just 6 using more conventional drug discovery approaches and were then synthesized in the lab. They were put through increasingly stringent tests, but the leading candidate was found to be effective at targeting the desired protein and behaved as one would hope a drug would.
The authors are clear that the results are just a proof-of-concept, which company CEO Alex Zhavoronkov told Wired stemmed from a challenge set by a pharma partner to design a drug as quickly as possible. But they say they were able to carry out the process faster than traditional methods for a fraction of the cost.
There are some caveats. For a start, the protein being targeted is already very well known and multiple effective drugs exist for it. That gave the company a wealth of data to train their model on, something that isn’t the case for many of the diseases where we urgently need new drugs.
The company’s platform also only targets the very initial stages of the drug discovery process. The authors concede in their paper that the molecules would still take considerable optimization in the lab before they’d be true contenders for clinical trials.
“And that is where you will start to begin to commence to spend the vast piles of money that you will eventually go through in trying to get a drug to market,” writes Derek Lowe in his blog In The Pipeline. The part of the discovery process that the platform tackles represents a tiny fraction of the total cost of drug development, he says.
Nonetheless, the research is a definite advance for virtual screening technology and an important marker of the potential of AI for designing new medicines. Zhavoronkov also told Wired that this research was done more than a year ago, and they’ve since adapted the platform to go after harder drug targets with less data.
And with big pharma companies desperate to slash their ballooning development costs and find treatments for a host of intractable diseases, they can use all the help they can get.
Image Credit: freestocks.org / Unsplash Continue reading
People say size doesn’t matter, but when it comes to AI the makers of the largest computer chip ever beg to differ. There are plenty of question marks about the gargantuan processor, but its unconventional design could herald an innovative new era in silicon design.
Computer chips specialized to run deep learning algorithms are a booming area of research as hardware limitations begin to slow progress, and both established players and startups are vying to build the successor to the GPU, the specialized graphics chip that has become the workhorse of the AI industry.
On Monday Californian startup Cerebras came out of stealth mode to unveil an AI-focused processor that turns conventional wisdom on its head. For decades chip makers have been focused on making their products ever-smaller, but the Wafer Scale Engine (WSE) is the size of an iPad and features 1.2 trillion transistors, 400,000 cores, and 18 gigabytes of on-chip memory.
The Cerebras Wafer-Scale Engine (WSE) is the largest chip ever built. It measures 46,225 square millimeters and includes 1.2 trillion transistors. Optimized for artificial intelligence compute, the WSE is shown here for comparison alongside the largest graphics processing unit. Image Credit: Used with permission from Cerebras Systems.
There is a method to the madness, though. Currently, getting enough cores to run really large-scale deep learning applications means connecting banks of GPUs together. But shuffling data between these chips is a major drain on speed and energy efficiency because the wires connecting them are relatively slow.
Building all 400,000 cores into the same chip should get round that bottleneck, but there are reasons it’s not been done before, and Cerebras has had to come up with some clever hacks to get around those obstacles.
Regular computer chips are manufactured using a process called photolithography to etch transistors onto the surface of a wafer of silicon. The wafers are inches across, so multiple chips are built onto them at once and then split up afterwards. But at 8.5 inches across, the WSE uses the entire wafer for a single chip.
The problem is that while for standard chip-making processes any imperfections in manufacturing will at most lead to a few processors out of several hundred having to be ditched, for Cerebras it would mean scrapping the entire wafer. To get around this the company built in redundant circuits so that even if there are a few defects, the chip can route around them.
The other big issue with a giant chip is the enormous amount of heat the processors can kick off—so the company has had to design a proprietary water-cooling system. That, along with the fact that no one makes connections and packaging for giant chips, means the WSE won’t be sold as a stand-alone component, but as part of a pre-packaged server incorporating the cooling technology.
There are no details on costs or performance so far, but some customers have already been testing prototypes, and according to Cerebras results have been promising. CEO and co-founder Andrew Feldman told Fortune that early tests show they are reducing training time from months to minutes.
We’ll have to wait until the first systems ship to customers in September to see if those claims stand up. But Feldman told ZDNet that the design of their chip should help spur greater innovation in the way engineers design neural networks. Many cornerstones of this process—for instance, tackling data in batches rather than individual data points—are guided more by the hardware limitations of GPUs than by machine learning theory, but their chip will do away with many of those obstacles.
Whether that turns out to be the case or not, the WSE might be the first indication of an innovative new era in silicon design. When Google announced it’s AI-focused Tensor Processing Unit in 2016 it was a wake-up call for chipmakers that we need some out-of-the-box thinking to square the slowing of Moore’s Law with skyrocketing demand for computing power.
It’s not just tech giants’ AI server farms driving innovation. At the other end of the spectrum, the desire to embed intelligence in everyday objects and mobile devices is pushing demand for AI chips that can run on tiny amounts of power and squeeze into the smallest form factors.
These trends have spawned renewed interest in everything from brain-inspired neuromorphic chips to optical processors, but the WSE also shows that there might be mileage in simply taking a sideways look at some of the other design decisions chipmakers have made in the past rather than just pumping ever more transistors onto a chip.
This gigantic chip might be the first exhibit in a weird and wonderful new menagerie of exotic, AI-inspired silicon.
Image Credit: Used with permission from Cerebras Systems. Continue reading
When I lived in Beijing back in the 90s, a man walking his bike was nothing to look at. But today, I did a serious double-take at a video of a bike walking his man.
The bike itself looks overloaded but otherwise completely normal. Underneath its simplicity, however, is a hybrid computer chip that combines brain-inspired circuits with machine learning processes into a computing behemoth. Thanks to its smart chip, the bike self-balances as it gingerly rolls down a paved track before smoothly gaining speed into a jogging pace while navigating dexterously around obstacles. It can even respond to simple voice commands such as “speed up,” “left,” or “straight.”
Far from a circus trick, the bike is a real-world demo of the AI community’s latest attempt at fashioning specialized hardware to keep up with the challenges of machine learning algorithms. The Tianjic (天机*) chip isn’t just your standard neuromorphic chip. Rather, it has the architecture of a brain-like chip, but can also run deep learning algorithms—a match made in heaven that basically mashes together neuro-inspired hardware and software.
The study shows that China is readily nipping at the heels of Google, Facebook, NVIDIA, and other tech behemoths investing in developing new AI chip designs—hell, with billions in government investment it may have already had a head start. A sweeping AI plan from 2017 looks to catch up with the US on AI technology and application by 2020. By 2030, China’s aiming to be the global leader—and a champion for building general AI that matches humans in intellectual competence.
The country’s ambition is reflected in the team’s parting words.
“Our study is expected to stimulate AGI [artificial general intelligence] development by paving the way to more generalized hardware platforms,” said the authors, led by Dr. Luping Shi at Tsinghua University.
A Hardware Conundrum
Shi’s autonomous bike isn’t the first robotic two-wheeler. Back in 2015, the famed research nonprofit SRI International in Menlo Park, California teamed up with Yamaha to engineer MOTOBOT, a humanoid robot capable of driving a motorcycle. Powered by state-of-the-art robotic hardware and machine learning, MOTOBOT eventually raced MotoGPTM world champion Valentino Rossi in a nail-biting match-off.
However, the technological core of MOTOBOT and Shi’s bike vastly differ, and that difference reflects two pathways towards more powerful AI. One, exemplified by MOTOBOT, is software—developing brain-like algorithms with increasingly efficient architecture, efficacy, and speed. That sounds great, but deep neural nets demand so many computational resources that general-purpose chips can’t keep up.
As Shi told China Science Daily: “CPUs and other chips are driven by miniaturization technologies based on physics. Transistors might shrink to nanoscale-level in 10, 20 years. But what then?” As more transistors are squeezed onto these chips, efficient cooling becomes a limiting factor in computational speed. Tax them too much, and they melt.
For AI processes to continue, we need better hardware. An increasingly popular idea is to build neuromorphic chips, which resemble the brain from the ground up. IBM’s TrueNorth, for example, contains a massively parallel architecture nothing like the traditional Von Neumann structure of classic CPUs and GPUs. Similar to biological brains, TrueNorth’s memory is stored within “synapses” between physical “neurons” etched onto the chip, which dramatically cuts down on energy consumption.
But even these chips are limited. Because computation is tethered to hardware architecture, most chips resemble just one specific type of brain-inspired network called spiking neural networks (SNNs). Without doubt, neuromorphic chips are highly efficient setups with dynamics similar to biological networks. They also don’t play nicely with deep learning and other software-based AI.
Brain-AI Hybrid Core
Shi’s new Tianjic chip brought the two incompatibilities together onto a single piece of brainy hardware.
First was to bridge the deep learning and SNN divide. The two have very different computation philosophies and memory organizations, the team said. The biggest difference, however, is that artificial neural networks transform multidimensional data—image pixels, for example—into a single, continuous, multi-bit 0 and 1 stream. In contrast, neurons in SNNs activate using something called “binary spikes” that code for specific activation events in time.
Confused? Yeah, it’s hard to wrap my head around it too. That’s because SNNs act very similarly to our neural networks and nothing like computers. A particular neuron needs to generate an electrical signal (a “spike”) large enough to transfer down to the next one; little blips in signals don’t count. The way they transmit data also heavily depends on how they’re connected, or the network topology. The takeaway: SNNs work pretty differently than deep learning.
Shi’s team first recreated this firing quirk in the language of computers—0s and 1s—so that the coding mechanism would become compatible with deep learning algorithms. They then carefully aligned the step-by-step building blocks of the two models, which allowed them to tease out similarities into a common ground to further build on. “On the basis of this unified abstraction, we built a cross-paradigm neuron scheme,” they said.
In general, the design allowed both computational approaches to share the synapses, where neurons connect and store data, and the dendrites, the outgoing branches of the neurons. In contrast, the neuron body, where signals integrate, was left reconfigurable for each type of computation, as were the input branches. Each building block was combined into a single unified functional core (FCore), which acts like a deep learning/SNN converter depending on its specific setup. Translation: the chip can do both types of previously incompatible computation.
Using nanoscale fabrication, the team arranged 156 FCores, containing roughly 40,000 neurons and 10 million synapses, onto a chip less than a fifth of an inch in length and width. Initial tests showcased the chip’s versatility, in that it can run both SNNs and deep learning algorithms such as the popular convolutional neural network (CNNs) often used in machine vision.
Compared to IBM TrueNorth, the density of Tianjic’s cores increased by 20 percent, speeding up performance ten times and increasing bandwidth at least 100-fold, the team said. When pitted against GPUs, the current hardware darling of machine learning, the chip increased processing throughput up to 100 times, while using just a sliver (1/10,000) of energy.
Although these stats are great, real-life performance is even better as a demo. Here’s where the authors gave their Tianjic brain a body. The team combined one chip with multiple specialized networks to process vision, balance, voice commands, and decision-making in real time. Object detection and target tracking, for example, relied on a deep neural net CNN, whereas voice commands and balance data were recognized using an SNN. The inputs were then integrated inside a neural state machine, which churned out decisions to downstream output modules—for example, controlling the handle bar to turn left.
Thanks to the chip’s brain-like architecture and bilingual ability, Tianjic “allowed all of the neural network models to operate in parallel and realized seamless communication across the models,” the team said. The result is an autonomous bike that rolls after its human, balances across speed bumps, avoids crashing into roadblocks, and answers to voice commands.
“It’s a wonderful demonstration and quite impressive,” said the editorial team at Nature, which published the study on its cover last week.
However, they cautioned, when comparing Tianjic with state-of-the-art chips designed for a single problem toe-to-toe on that particular problem, Tianjic falls behind. But building these jack-of-all-trades hybrid chips is definitely worth the effort. Compared to today’s limited AI, what people really want is artificial general intelligence, which will require new architectures that aren’t designed to solve one particular problem.
Until people start to explore, innovate, and play around with different designs, it’s not clear how we can further progress in the pursuit of general AI. A self-driving bike might not be much to look at, but its hybrid brain is a pretty neat place to start.
*The name, in Chinese, means “heavenly machine,” “unknowable mystery of nature,” or “confidentiality.” Go figure.
Image Credit: Alexander Ryabintsev / Shutterstock.com Continue reading