Tag Archives: training
#435528 The Time for AI Is Now. Here’s Why
You hear a lot these days about the sheer transformative power of AI.
There’s pure intelligence: DeepMind’s algorithms readily beat humans at Go and StarCraft, and DeepStack triumphs over humans at no-limit hold’em poker. Often, these silicon brains generate gameplay strategies that don’t resemble anything from a human mind.
There’s astonishing speed: algorithms routinely surpass radiologists in diagnosing breast cancer, eye disease, and other ailments visible from medical imaging, essentially collapsing decades of expert training down to a few months.
Although AI’s silent touch is mainly felt today in the technological, financial, and health sectors, its impact across industries is rapidly spreading. At the Singularity University Global Summit in San Francisco this week Neil Jacobstein, Chair of AI and Robotics, painted a picture of a better AI-powered future for humanity that is already here.
Thanks to cloud-based cognitive platforms, sophisticated AI tools like deep learning are no longer relegated to academic labs. For startups looking to tackle humanity’s grand challenges, the tools to efficiently integrate AI into their missions are readily available. The progress of AI is massively accelerating—to the point you need help from AI to track its progress, joked Jacobstein.
Now is the time to consider how AI can impact your industry, and in the process, begin to envision a beneficial relationship with our machine coworkers. As Jacobstein stressed in his talk, the future of a brain-machine mindmeld is a collaborative intelligence that augments our own. “AI is reinventing the way we invent,” he said.
AI’s Rapid Revolution
Machine learning and other AI-based methods may seem academic and abstruse. But Jacobstein pointed out that there are already plenty of real-world AI application frameworks.
Their secret? Rather than coding from scratch, smaller companies—with big visions—are tapping into cloud-based solutions such as Google’s TensorFlow, Microsoft’s Azure, or Amazon’s AWS to kick off their AI journey. These platforms act as all-in-one solutions that not only clean and organize data, but also contain built-in security and drag-and-drop coding that allow anyone to experiment with complicated machine learning algorithms.
Google Cloud’s Anthos, for example, lets anyone migrate data from other servers—IBM Watson or AWS, for example—so users can leverage different computing platforms and algorithms to transform data into insights and solutions.
Rather than coding from scratch, it’s already possible to hop onto a platform and play around with it, said Jacobstein. That’s key: this democratization of AI is how anyone can begin exploring solutions to problems we didn’t even know we had, or those long thought improbable.
The acceleration is only continuing. Much of AI’s mind-bending pace is thanks to a massive infusion of funding. Microsoft recently injected $1 billion into OpenAI, the Elon Musk venture that engineers socially responsible artificial general intelligence (AGI).
The other revolution is in hardware, and Google, IBM, and NVIDIA—among others—are racing to manufacture computing chips tailored to machine learning.
Democratizing AI is like the birth of the printing press. Mechanical printing allowed anyone to become an author; today, an iPhone lets anyone film a movie masterpiece.
However, this diffusion of AI into the fabric of our lives means tech explorers need to bring skepticism to their AI solutions, giving them a dose of empathy, nuance, and humanity.
A Path Towards Ethical AI
The democratization of AI is a double-edged sword: as more people wield the technology’s power in real-world applications, problems embedded in deep learning threaten to disrupt those very judgment calls.
Much of the press on the dangers of AI focuses on superintelligence—AI that’s more adept at learning than humans—taking over the world, said Jacobstein. But the near-term threat, and far more insidious, is in humans misusing the technology.
Deepfakes, for example, allow AI rookies to paste one person’s head on a different body or put words into a person’s mouth. As the panel said, it pays to think of AI as a cybersecurity problem, one with currently shaky accountability and complexity, and one that fails at diversity and bias.
Take bias. Thanks to progress in natural language processing, Google Translate works nearly perfectly today, so much so that many consider the translation problem solved. Not true, the panel said. One famous example is how the algorithm translates gender-neutral terms like “doctor” into “he” and “nurse” into “she.”
These biases reflect our own, and it’s not just a data problem. To truly engineer objective AI systems, ones stripped of our society’s biases, we need to ask who is developing these systems, and consult those who will be impacted by the products. In addition to gender, racial bias is also rampant. For example, one recent report found that a supposedly objective crime-predicting system was trained on falsified data, resulting in outputs that further perpetuate corrupt police practices. Another study from Google just this month found that their hate speech detector more often labeled innocuous tweets from African-Americans as “obscene” compared to tweets from people of other ethnicities.
We often think of building AI as purely an engineering job, the panelists agreed. But similar to gene drives, germ-line genome editing, and other transformative—but dangerous—tools, AI needs to grow under the consultation of policymakers and other stakeholders. It pays to start young: educating newer generations on AI biases will mold malleable minds early, alerting them to the problem of bias and potentially mitigating risks.
As panelist Tess Posner from AI4ALL said, AI is rocket fuel for ambition. If young minds set out using the tools of AI to tackle their chosen problems, while fully aware of its inherent weaknesses, we can begin to build an AI-embedded future that is widely accessible and inclusive.
The bottom line: people who will be impacted by AI need to be in the room at the conception of an AI solution. People will be displaced by the new technology, and ethical AI has to consider how to mitigate human suffering during the transition. Just because AI looks like “magic fairy dust doesn’t mean that you’re home free,” the panelists said. You, the sentient human, bear the burden of being responsible for how you decide to approach the technology.
The time for AI is now. Let’s make it ethical.
Image Credit: GrAI / Shutterstock.com Continue reading →
Posted in Human Robots
#435423 Moving Beyond Mind-Controlled Limbs to ...
Brain-machine interface enthusiasts often gush about “closing the loop.” It’s for good reason. On the implant level, it means engineering smarter probes that only activate when they detect faulty electrical signals in brain circuits. Elon Musk’s Neuralink—among other players—are readily pursuing these bi-directional implants that both measure and zap the brain.
But to scientists laboring to restore functionality to paralyzed patients or amputees, “closing the loop” has broader connotations. Building smart mind-controlled robotic limbs isn’t enough; the next frontier is restoring sensation in offline body parts. To truly meld biology with machine, the robotic appendage has to “feel one” with the body.
This month, two studies from Science Robotics describe complementary ways forward. In one, scientists from the University of Utah paired a state-of-the-art robotic arm—the DEKA LUKE—with electrically stimulating remaining nerves above the attachment point. Using artificial zaps to mimic the skin’s natural response patterns to touch, the team dramatically increased the patient’s ability to identify objects. Without much training, he could easily discriminate between the small and large and the soft and hard while blindfolded and wearing headphones.
In another, a team based at the National University of Singapore took inspiration from our largest organ, the skin. Mimicking the neural architecture of biological skin, the engineered “electronic skin” not only senses temperature, pressure, and humidity, but continues to function even when scraped or otherwise damaged. Thanks to artificial nerves that transmit signals far faster than our biological ones, the flexible e-skin shoots electrical data 1,000 times quicker than human nerves.
Together, the studies marry neuroscience and robotics. Representing the latest push towards closing the loop, they show that integrating biological sensibilities with robotic efficiency isn’t impossible (super-human touch, anyone?). But more immediately—and more importantly—they’re beacons of hope for patients who hope to regain their sense of touch.
For one of the participants, a late middle-aged man with speckled white hair who lost his forearm 13 years ago, superpowers, cyborgs, or razzle-dazzle brain implants are the last thing on his mind. After a barrage of emotionally-neutral scientific tests, he grasped his wife’s hand and felt her warmth for the first time in over a decade. His face lit up in a blinding smile.
That’s what scientists are working towards.
Biomimetic Feedback
The human skin is a marvelous thing. Not only does it rapidly detect a multitude of sensations—pressure, temperature, itch, pain, humidity—its wiring “binds” disparate signals together into a sensory fingerprint that helps the brain identify what it’s feeling at any moment. Thanks to over 45 miles of nerves that connect the skin, muscles, and brain, you can pick up a half-full coffee cup, knowing that it’s hot and sloshing, while staring at your computer screen. Unfortunately, this complexity is also why restoring sensation is so hard.
The sensory electrode array implanted in the participant’s arm. Image Credit: George et al., Sci. Robot. 4, eaax2352 (2019)..
However, complex neural patterns can also be a source of inspiration. Previous cyborg arms are often paired with so-called “standard” sensory algorithms to induce a basic sense of touch in the missing limb. Here, electrodes zap residual nerves with intensities proportional to the contact force: the harder the grip, the stronger the electrical feedback. Although seemingly logical, that’s not how our skin works. Every time the skin touches or leaves an object, its nerves shoot strong bursts of activity to the brain; while in full contact, the signal is much lower. The resulting electrical strength curve resembles a “U.”
The LUKE hand. Image Credit: George et al., Sci. Robot. 4, eaax2352 (2019).
The team decided to directly compare standard algorithms with one that better mimics the skin’s natural response. They fitted a volunteer with a robotic LUKE arm and implanted an array of electrodes into his forearm—right above the amputation—to stimulate the remaining nerves. When the team activated different combinations of electrodes, the man reported sensations of vibration, pressure, tapping, or a sort of “tightening” in his missing hand. Some combinations of zaps also made him feel as if he were moving the robotic arm’s joints.
In all, the team was able to carefully map nearly 120 sensations to different locations on the phantom hand, which they then overlapped with contact sensors embedded in the LUKE arm. For example, when the patient touched something with his robotic index finger, the relevant electrodes sent signals that made him feel as if he were brushing something with his own missing index fingertip.
Standard sensory feedback already helped: even with simple electrical stimulation, the man could tell apart size (golf versus lacrosse ball) and texture (foam versus plastic) while blindfolded and wearing noise-canceling headphones. But when the team implemented two types of neuromimetic feedback—electrical zaps that resembled the skin’s natural response—his performance dramatically improved. He was able to identify objects much faster and more accurately under their guidance. Outside the lab, he also found it easier to cook, feed, and dress himself. He could even text on his phone and complete routine chores that were previously too difficult, such as stuffing an insert into a pillowcase, hammering a nail, or eating hard-to-grab foods like eggs and grapes.
The study shows that the brain more readily accepts biologically-inspired electrical patterns, making it a relatively easy—but enormously powerful—upgrade that seamlessly integrates the robotic arms with the host. “The functional and emotional benefits…are likely to be further enhanced with long-term use, and efforts are underway to develop a portable take-home system,” the team said.
E-Skin Revolution: Asynchronous Coded Electronic Skin (ACES)
Flexible electronic skins also aren’t new, but the second team presented an upgrade in both speed and durability while retaining multiplexed sensory capabilities.
Starting from a combination of rubber, plastic, and silicon, the team embedded over 200 sensors onto the e-skin, each capable of discerning contact, pressure, temperature, and humidity. They then looked to the skin’s nervous system for inspiration. Our skin is embedded with a dense array of nerve endings that individually transmit different types of sensations, which are integrated inside hubs called ganglia. Compared to having every single nerve ending directly ping data to the brain, this “gather, process, and transmit” architecture rapidly speeds things up.
The team tapped into this biological architecture. Rather than pairing each sensor with a dedicated receiver, ACES sends all sensory data to a single receiver—an artificial ganglion. This setup lets the e-skin’s wiring work as a whole system, as opposed to individual electrodes. Every sensor transmits its data using a characteristic pulse, which allows it to be uniquely identified by the receiver.
The gains were immediate. First was speed. Normally, sensory data from multiple individual electrodes need to be periodically combined into a map of pressure points. Here, data from thousands of distributed sensors can independently go to a single receiver for further processing, massively increasing efficiency—the new e-skin’s transmission rate is roughly 1,000 times faster than that of human skin.
Second was redundancy. Because data from individual sensors are aggregated, the system still functioned even when any individual receptors are damaged, making it far more resilient than previous attempts. Finally, the setup could easily scale up. Although the team only tested the idea with 240 sensors, theoretically the system should work with up to 10,000.
The team is now exploring ways to combine their invention with other material layers to make it water-resistant and self-repairable. As you might’ve guessed, an immediate application is to give robots something similar to complex touch. A sensory upgrade not only lets robots more easily manipulate tools, doorknobs, and other objects in hectic real-world environments, it could also make it easier for machines to work collaboratively with humans in the future (hey Wall-E, care to pass the salt?).
Dexterous robots aside, the team also envisions engineering better prosthetics. When coated onto cyborg limbs, for example, ACES may give them a better sense of touch that begins to rival the human skin—or perhaps even exceed it.
Regardless, efforts that adapt the functionality of the human nervous system to machines are finally paying off, and more are sure to come. Neuromimetic ideas may very well be the link that finally closes the loop.
Image Credit: Dan Hixson/University of Utah College of Engineering.. Continue reading →
Posted in Human Robots
#435224 Can AI Save the Internet from Fake News?
There’s an old proverb that says “seeing is believing.” But in the age of artificial intelligence, it’s becoming increasingly difficult to take anything at face value—literally.
The rise of so-called “deepfakes,” in which different types of AI-based techniques are used to manipulate video content, has reached the point where Congress held its first hearing last month on the potential abuses of the technology. The congressional investigation coincided with the release of a doctored video of Facebook CEO Mark Zuckerberg delivering what appeared to be a sinister speech.
View this post on Instagram
‘Imagine this…’ (2019) Mark Zuckerberg reveals the truth about Facebook and who really owns the future… see more @sheffdocfest VDR technology by @cannyai #spectreknows #privacy #democracy #surveillancecapitalism #dataism #deepfake #deepfakes #contemporaryartwork #digitalart #generativeart #newmediaart #codeart #markzuckerberg #artivism #contemporaryart
A post shared by Bill Posters (@bill_posters_uk) on Jun 7, 2019 at 7:15am PDT
Scientists are scrambling for solutions on how to combat deepfakes, while at the same time others are continuing to refine the techniques for less nefarious purposes, such as automating video content for the film industry.
At one end of the spectrum, for example, researchers at New York University’s Tandon School of Engineering have proposed implanting a type of digital watermark using a neural network that can spot manipulated photos and videos.
The idea is to embed the system directly into a digital camera. Many smartphone cameras and other digital devices already use AI to boost image quality and make other corrections. The authors of the study out of NYU say their prototype platform increased the chances of detecting manipulation from about 45 percent to more than 90 percent without sacrificing image quality.
On the other hand, researchers at Carnegie Mellon University recently hit on a technique for automatically and rapidly converting large amounts of video content from one source into the style of another. In one example, the scientists transferred the facial expressions of comedian John Oliver onto the bespectacled face of late night show host Stephen Colbert.
The CMU team says the method could be a boon to the movie industry, such as by converting black and white films to color, though it also conceded that the technology could be used to develop deepfakes.
Words Matter with Fake News
While the current spotlight is on how to combat video and image manipulation, a prolonged trench warfare on fake news is being fought by academia, nonprofits, and the tech industry.
This isn’t the fake news that some have come to use as a knee-jerk reaction to fact-based information that might be less than flattering to the subject of the report. Rather, fake news is deliberately-created misinformation that is spread via the internet.
In a recent Pew Research Center poll, Americans said fake news is a bigger problem than violent crime, racism, and terrorism. Fortunately, many of the linguistic tools that have been applied to determine when people are being deliberately deceitful can be baked into algorithms for spotting fake news.
That’s the approach taken by a team at the University of Michigan (U-M) to develop an algorithm that was better than humans at identifying fake news—76 percent versus 70 percent—by focusing on linguistic cues like grammatical structure, word choice, and punctuation.
For example, fake news tends to be filled with hyperbole and exaggeration, using terms like “overwhelming” or “extraordinary.”
“I think that’s a way to make up for the fact that the news is not quite true, so trying to compensate with the language that’s being used,” Rada Mihalcea, a computer science and engineering professor at U-M, told Singularity Hub.
The paper “Automatic Detection of Fake News” was based on the team’s previous studies on how people lie in general, without necessarily having the intention of spreading fake news, she said.
“Deception is a complicated and complex phenomenon that requires brain power,” Mihalcea noted. “That often results in simpler language, where you have shorter sentences or shorter documents.”
AI Versus AI
While most fake news is still churned out by humans with identifiable patterns of lying, according to Mihalcea, other researchers are already anticipating how to detect misinformation manufactured by machines.
A group led by Yejin Choi, with the Allen Institute of Artificial Intelligence and the University of Washington in Seattle, is one such team. The researchers recently introduced the world to Grover, an AI platform that is particularly good at catching autonomously-generated fake news because it’s equally good at creating it.
“This is due to a finding that is perhaps counterintuitive: strong generators for neural fake news are themselves strong detectors of it,” wrote Rowan Zellers, a PhD student and team member, in a Medium blog post. “A generator of fake news will be most familiar with its own peculiarities, such as using overly common or predictable words, as well as the peculiarities of similar generators.”
The team found that the best current discriminators can classify neural fake news from real, human-created text with 73 percent accuracy. Grover clocks in with 92 percent accuracy based on a training set of 5,000 neural network-generated fake news samples. Zellers wrote that Grover got better at scale, identifying 97.5 percent of made-up machine mumbo jumbo when trained on 80,000 articles.
It performed almost as well against fake news created by a powerful new text-generation system called GPT-2 built by OpenAI, a nonprofit research lab founded by Elon Musk, classifying 96.1 percent of the machine-written articles.
OpenAI had so feared that the platform could be abused that it has only released limited versions of the software. The public can play with a scaled-down version posted by a machine learning engineer named Adam King, where the user types in a short prompt and GPT-2 bangs out a short story or poem based on the snippet of text.
No Silver AI Bullet
While real progress is being made against fake news, the challenges of using AI to detect and correct misinformation are abundant, according to Hugo Williams, outreach manager for Logically, a UK-based startup that is developing different detectors using elements of deep learning and natural language processing, among others. He explained that the Logically models analyze information based on a three-pronged approach.
Publisher metadata: Is the article from a known, reliable, and trustworthy publisher with a history of credible journalism?
Network behavior: Is the article proliferating through social platforms and networks in ways typically associated with misinformation?
Content: The AI scans articles for hundreds of known indicators typically found in misinformation.
“There is no single algorithm which is capable of doing this,” Williams wrote in an email to Singularity Hub. “Even when you have a collection of different algorithms which—when combined—can give you relatively decent indications of what is unreliable or outright false, there will always need to be a human layer in the pipeline.”
The company released a consumer app in India back in February just before that country’s election cycle that was a “great testing ground” to refine its technology for the next app release, which is scheduled in the UK later this year. Users can submit articles for further scrutiny by a real person.
“We see our technology not as replacing traditional verification work, but as a method of simplifying and streamlining a very manual process,” Williams said. “In doing so, we’re able to publish more fact checks at a far quicker pace than other organizations.”
“With heightened analysis and the addition of more contextual information around the stories that our users are reading, we are not telling our users what they should or should not believe, but encouraging critical thinking based upon reliable, credible, and verified content,” he added.
AI may never be able to detect fake news entirely on its own, but it can help us be smarter about what we read on the internet.
Image Credit: Dennis Lytyagin / Shutterstock.com Continue reading →
Posted in Human Robots
#435199 The Rise of AI Art—and What It Means ...
Artificially intelligent systems are slowly taking over tasks previously done by humans, and many processes involving repetitive, simple movements have already been fully automated. In the meantime, humans continue to be superior when it comes to abstract and creative tasks.
However, it seems like even when it comes to creativity, we’re now being challenged by our own creations.
In the last few years, we’ve seen the emergence of hundreds of “AI artists.” These complex algorithms are creating unique (and sometimes eerie) works of art. They’re generating stunning visuals, profound poetry, transcendent music, and even realistic movie scripts. The works of these AI artists are raising questions about the nature of art and the role of human creativity in future societies.
Here are a few works of art created by non-human entities.
Unsecured Futures
by Ai.Da
Ai-Da Robot with Painting. Image Credit: Ai-Da portraits by Nicky Johnston. Published with permission from Midas Public Relations.
Earlier this month we saw the announcement of Ai.Da, considered the first ultra-realistic drawing robot artist. Her mechanical abilities, combined with AI-based algorithms, allow her to draw, paint, and even sculpt. She is able to draw people using her artificial eye and a pencil in her hand. Ai.Da’s artwork and first solo exhibition, Unsecured Futures, will be showcased at Oxford University in July.
Ai-Da Cartesian Painting. Image Credit: Ai-Da Artworks. Published with permission from Midas Public Relations.
Obviously Ai.Da has no true consciousness, thoughts, or feelings. Despite that, the (human) organizers of the exhibition believe that Ai.Da serves as a basis for crucial conversations about the ethics of emerging technologies. The exhibition will serve as a stimulant for engaging with critical questions about what kind of future we ought to create via such technologies.
The exhibition’s creators wrote, “Humans are confident in their position as the most powerful species on the planet, but how far do we actually want to take this power? To a Brave New World (Nightmare)? And if we use new technologies to enhance the power of the few, we had better start safeguarding the future of the many.”
Google’s PoemPortraits
Our transcendence adorns,
That society of the stars seem to be the secret.
The two lines of poetry above aren’t like any poetry you’ve come across before. They are generated by an algorithm that was trained via deep learning neural networks trained on 20 million words of 19th-century poetry.
Google’s latest art project, named PoemPortraits, takes a word of your suggestion and generates a unique poem (once again, a collaboration of man and machine). You can even add a selfie in the final “PoemPortrait.” Artist Es Devlin, the project’s creator, explains that the AI “doesn’t copy or rework existing phrases, but uses its training material to build a complex statistical model. As a result, the algorithm generates original phrases emulating the style of what it’s been trained on.”
The generated poetry can sometimes be profound, and sometimes completely meaningless.But what makes the PoemPortraits project even more interesting is that it’s a collaborative project. All of the generated lines of poetry are combined to form a consistently growing collective poem, which you can view after your lines are generated. In many ways, the final collective poem is a collaboration of people from around the world working with algorithms.
Faceless Portraits Transcending Time
AICAN + Ahmed Elgammal
Image Credit: AICAN + Ahmed Elgammal | Faceless Portrait #2 (2019) | Artsy.
In March of this year, an AI artist called AICAN and its creator Ahmed Elgammal took over a New York gallery. The exhibition at HG Commentary showed two series of canvas works portraying harrowing, dream-like faceless portraits.
The exhibition was not simply credited to a machine, but rather attributed to the collaboration between a human and machine. Ahmed Elgammal is the founder and director of the Art and Artificial Intelligence Laboratory at Rutgers University. He considers AICAN to not only be an autonomous AI artist, but also a collaborator for artistic endeavors.
How did AICAN create these eerie faceless portraits? The system was presented with 100,000 photos of Western art from over five centuries, allowing it to learn the aesthetics of art via machine learning. It then drew from this historical knowledge and the mandate to create something new to create an artwork without human intervention.
Genesis
by AIVA Technologies
Listen to the score above. While you do, reflect on the fact that it was generated by an AI.
AIVA is an AI that composes soundtrack music for movies, commercials, games, and trailers. Its creative works span a wide range of emotions and moods. The scores it generates are indistinguishable from those created by the most talented human composers.
The AIVA music engine allows users to generate original scores in multiple ways. One is to upload an existing human-generated score and select the temp track to base the composition process on. Another method involves using preset algorithms to compose music in pre-defined styles, including everything from classical to Middle Eastern.
Currently, the platform is promoted as an opportunity for filmmakers and producers. But in the future, perhaps every individual will have personalized music generated for them based on their interests, tastes, and evolving moods. We already have algorithms on streaming websites recommending novel music to us based on our interests and history. Soon, algorithms may be used to generate music and other works of art that are tailored to impact our unique psyches.
The Future of Art: Pushing Our Creative Limitations
These works of art are just a glimpse into the breadth of the creative works being generated by algorithms and machines. Many of us will rightly fear these developments. We have to ask ourselves what our role will be in an era where machines are able to perform what we consider complex, abstract, creative tasks. The implications on the future of work, education, and human societies are profound.
At the same time, some of these works demonstrate that AI artists may not necessarily represent a threat to human artists, but rather an opportunity for us to push our creative boundaries. The most exciting artistic creations involve collaborations between humans and machines.
We have always used our technological scaffolding to push ourselves beyond our biological limitations. We use the telescope to extend our line of sight, planes to fly, and smartphones to connect with others. Our machines are not always working against us, but rather working as an extension of our minds. Similarly, we could use our machines to expand on our creativity and push the boundaries of art.
Image Credit: Ai-Da portraits by Nicky Johnston. Published with permission from Midas Public Relations. Continue reading →
Posted in Human Robots
#435127 Teaching AI the Concept of ‘Similar, ...
As a human you instinctively know that a leopard is closer to a cat than a motorbike, but the way we train most AI makes them oblivious to these kinds of relations. Building the concept of similarity into our algorithms could make them far more capable, writes the author of a new paper in Science Robotics.
Convolutional neural networks have revolutionized the field of computer vision to the point that machines are now outperforming humans on some of the most challenging visual tasks. But the way we train them to analyze images is very different from the way humans learn, says Atsuto Maki, an associate professor at KTH Royal Institute of Technology.
“Imagine that you are two years old and being quizzed on what you see in a photo of a leopard,” he writes. “You might answer ‘a cat’ and your parents might say, ‘yeah, not quite but similar’.”
In contrast, the way we train neural networks rarely gives that kind of partial credit. They are typically trained to have very high confidence in the correct label and consider all incorrect labels, whether ”cat” or “motorbike,” equally wrong. That’s a mistake, says Maki, because ignoring the fact that something can be “less wrong” means you’re not exploiting all of the information in the training data.
Even when models are trained this way, there will be small differences in the probabilities assigned to incorrect labels that can tell you a lot about how well the model can generalize what it has learned to unseen data.
If you show a model a picture of a leopard and it gives “cat” a probability of five percent and “motorbike” one percent, that suggests it picked up on the fact that a cat is closer to a leopard than a motorbike. In contrast, if the figures are the other way around it means the model hasn’t learned the broad features that make cats and leopards similar, something that could potentially be helpful when analyzing new data.
If we could boost this ability to identify similarities between classes we should be able to create more flexible models better able to generalize, says Maki. And recent research has demonstrated how variations of an approach called regularization might help us achieve that goal.
Neural networks are prone to a problem called “overfitting,” which refers to a tendency to pay too much attention to tiny details and noise specific to their training set. When that happens, models will perform excellently on their training data but poorly when applied to unseen test data without these particular quirks.
Regularization is used to circumvent this problem, typically by reducing the network’s capacity to learn all this unnecessary information and therefore boost its ability to generalize to new data. Techniques are varied, but generally involve modifying the network’s structure or the strength of the weights between artificial neurons.
More recently, though, researchers have suggested new regularization approaches that work by encouraging a broader spread of probabilities across all classes. This essentially helps them capture more of the class similarities, says Maki, and therefore boosts their ability to generalize.
One such approach was devised in 2017 by Google Brain researchers, led by deep learning pioneer Geoffrey Hinton. They introduced a penalty to their training process that directly punished overconfident predictions in the model’s outputs, and a technique called label smoothing that prevents the largest probability becoming much larger than all others. This meant the probabilities were lower for correct labels and higher for incorrect ones, which was found to boost performance of models on varied tasks from image classification to speech recognition.
Another came from Maki himself in 2017 and achieves the same goal, but by suppressing high values in the model’s feature vector—the mathematical construct that describes all of an object’s important characteristics. This has a knock-on effect on the spread of output probabilities and also helped boost performance on various image classification tasks.
While it’s still early days for the approach, the fact that humans are able to exploit these kinds of similarities to learn more efficiently suggests that models that incorporate them hold promise. Maki points out that it could be particularly useful in applications such as robotic grasping, where distinguishing various similar objects is important.
Image Credit: Marianna Kalashnyk / Shutterstock.com Continue reading →
Posted in Human Robots