Tag Archives: turning
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
A new technique using artificial intelligence to manipulate video content gives new meaning to the expression “talking head.”
An international team of researchers showcased the latest advancement in synthesizing facial expressions—including mouth, eyes, eyebrows, and even head position—in video at this month’s 2018 SIGGRAPH, a conference on innovations in computer graphics, animation, virtual reality, and other forms of digital wizardry.
The project is called Deep Video Portraits. It relies on a type of AI called generative adversarial networks (GANs) to modify a “target” actor based on the facial and head movement of a “source” actor. As the name implies, GANs pit two opposing neural networks against one another to create a realistic talking head, right down to the sneer or raised eyebrow.
In this case, the adversaries are actually working together: One neural network generates content, while the other rejects or approves each effort. The back-and-forth interplay between the two eventually produces a realistic result that can easily fool the human eye, including reproducing a static scene behind the head as it bobs back and forth.
The researchers say the technique can be used by the film industry for a variety of purposes, from editing facial expressions of actors for matching dubbed voices to repositioning an actor’s head in post-production. AI can not only produce highly realistic results, but much quicker ones compared to the manual processes used today, according to the researchers. You can read the full paper of their work here.
“Deep Video Portraits shows how such a visual effect could be created with less effort in the future,” said Christian Richardt, from the University of Bath’s motion capture research center CAMERA, in a press release. “With our approach, even the positioning of an actor’s head and their facial expression could be easily edited to change camera angles or subtly change the framing of a scene to tell the story better.”
AI Tech Different Than So-Called “Deepfakes”
The work is far from the first to employ AI to manipulate video and audio. At last year’s SIGGRAPH conference, researchers from the University of Washington showcased their work using algorithms that inserted audio recordings from a person in one instance into a separate video of the same person in a different context.
In this case, they “faked” a video using a speech from former President Barack Obama addressing a mass shooting incident during his presidency. The AI-doctored video injects the audio into an unrelated video of the president while also blending the facial and mouth movements, creating a pretty credible job of lip synching.
A previous paper by many of the same scientists on the Deep Video Portraits project detailed how they were first able to manipulate a video in real time of a talking head (in this case, actor and former California governor Arnold Schwarzenegger). The Face2Face system pulled off this bit of digital trickery using a depth-sensing camera that tracked the facial expressions of an Asian female source actor.
A less sophisticated method of swapping faces using a machine learning software dubbed FakeApp emerged earlier this year. Predictably, the tech—requiring numerous photos of the source actor in order to train the neural network—was used for more juvenile pursuits, such as injecting a person’s face onto a porn star.
The application gave rise to the term “deepfakes,” which is now used somewhat ubiquitously to describe all such instances of AI-manipulated video—much to the chagrin of some of the researchers involved in more legitimate uses.
Fighting AI-Created Video Forgeries
However, the researchers are keenly aware that their work—intended for benign uses such as in the film industry or even to correct gaze and head positions for more natural interactions through video teleconferencing—could be used for nefarious purposes. Fake news is the most obvious concern.
“With ever-improving video editing technology, we must also start being more critical about the video content we consume every day, especially if there is no proof of origin,” said Michael Zollhöfer, a visiting assistant professor at Stanford University and member of the Deep Video Portraits team, in the press release.
Toward that end, the research team is training the same adversarial neural networks to spot video forgeries. They also strongly recommend that developers clearly watermark videos that are edited through AI or otherwise, and denote clearly what part and element of the scene was modified.
To catch less ethical users, the US Department of Defense, through the Defense Advanced Research Projects Agency (DARPA), is supporting a program called Media Forensics. This latest DARPA challenge enlists researchers to develop technologies to automatically assess the integrity of an image or video, as part of an end-to-end media forensics platform.
The DARPA official in charge of the program, Matthew Turek, did tell MIT Technology Review that so far the program has “discovered subtle cues in current GAN-manipulated images and videos that allow us to detect the presence of alterations.” In one reported example, researchers have targeted eyes, which rarely blink in the case of “deepfakes” like those created by FakeApp, because the AI is trained on still pictures. That method would seem to be less effective to spot the sort of forgeries created by Deep Video Portraits, which appears to flawlessly match the entire facial and head movements between the source and target actors.
“We believe that the field of digital forensics should and will receive a lot more attention in the future to develop approaches that can automatically prove the authenticity of a video clip,” Zollhöfer said. “This will lead to ever-better approaches that can spot such modifications even if we humans might not be able to spot them with our own eyes.
Image Credit: Tancha / Shutterstock.com Continue reading
Earth’s oceans are having a rough go of it these days. On top of being the repository for millions of tons of plastic waste, global warming is affecting the oceans and upsetting marine ecosystems in potentially irreversible ways.
Coral bleaching, for example, occurs when warming water temperatures or other stress factors cause coral to cast off the algae that live on them. The coral goes from lush and colorful to white and bare, and sometimes dies off altogether. This has a ripple effect on the surrounding ecosystem.
Warmer water temperatures have also prompted many species of fish to move closer to the north or south poles, disrupting fisheries and altering undersea environments.
To keep these issues in check or, better yet, try to address and improve them, it’s crucial for scientists to monitor what’s going on in the water. A paper released last week by a team from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) unveiled a new tool for studying marine life: a biomimetic soft robotic fish, dubbed SoFi, that can swim with, observe, and interact with real fish.
SoFi isn’t the first robotic fish to hit the water, but it is the most advanced robot of its kind. Here’s what sets it apart.
It swims in three dimensions
Up until now, most robotic fish could only swim forward at a given water depth, advancing at a steady speed. SoFi blows older models out of the water. It’s equipped with side fins called dive planes, which move to adjust its angle and allow it to turn, dive downward, or head closer to the surface. Its density and thus its buoyancy can also be adjusted by compressing or decompressing air in an inner compartment.
“To our knowledge, this is the first robotic fish that can swim untethered in three dimensions for extended periods of time,” said CSAIL PhD candidate Robert Katzschmann, lead author of the study. “We are excited about the possibility of being able to use a system like this to get closer to marine life than humans can get on their own.”
The team took SoFi to the Rainbow Reef in Fiji to test out its swimming skills, and the robo fish didn’t disappoint—it was able to swim at depths of over 50 feet for 40 continuous minutes. What keeps it swimming? A lithium polymer battery just like the one that powers our smartphones.
It’s remote-controlled… by Super Nintendo
SoFi has sensors to help it see what’s around it, but it doesn’t have a mind of its own yet. Rather, it’s controlled by a nearby scuba-diving human, who can send it commands related to speed, diving, and turning. The best part? The commands come from an actual repurposed (and waterproofed) Super Nintendo controller. What’s not to love?
Image Credit: MIT CSAIL
Previous robotic fish built by this team had to be tethered to a boat, so the fact that SoFi can swim independently is a pretty big deal. Communication between the fish and the diver was most successful when the two were less than 10 meters apart.
It looks real, sort of
SoFi’s side fins are a bit stiff, and its camera may not pass for natural—but otherwise, it looks a lot like a real fish. This is mostly thanks to the way its tail moves; a motor pumps water between two chambers in the tail, and as one chamber fills, the tail bends towards that side, then towards the other side as water is pumped into the other chamber. The result is a motion that closely mimics the way fish swim. Not only that, the hydraulic system can change the water flow to get different tail movements that let SoFi swim at varying speeds; its average speed is around half a body length (21.7 centimeters) per second.
Besides looking neat, it’s important SoFi look lifelike so it can blend in with marine life and not scare real fish away, so it can get close to them and observe them.
“A robot like this can help explore the reef more closely than current robots, both because it can get closer more safely for the reef and because it can be better accepted by the marine species.” said Cecilia Laschi, a biorobotics professor at the Sant’Anna School of Advanced Studies in Pisa, Italy.
Just keep swimming
It sounds like this fish is nothing short of a regular Nemo. But its creators aren’t quite finished yet.
They’d like SoFi to be able to swim faster, so they’ll work on improving the robo fish’s pump system and streamlining its body and tail design. They also plan to tweak SoFi’s camera to help it follow real fish.
“We view SoFi as a first step toward developing almost an underwater observatory of sorts,” said CSAIL director Daniela Rus. “It has the potential to be a new type of tool for ocean exploration and to open up new avenues for uncovering the mysteries of marine life.”
The CSAIL team plans to make a whole school of SoFis to help biologists learn more about how marine life is reacting to environmental changes.
Image Credit: MIT CSAIL Continue reading