MIT researchers have identified a method to help AI systems avoid adversarial attacks, reports Matthew Hutson for Science. When the researchers “trained an algorithm on images without the subtle features, their image recognition software was fooled by adversarial attacks only 50% of the time,” Hutson explains. “That compares with a 95% rate of vulnerability when the AI was trained on images with both obvious and subtle patterns.”
Researchers at MIT have found that adversarial examples, a kind of optical illusion for AI that makes the system incorrectly identify an image, may not actually impact AI in the ways computer scientists have previously thought. “When algorithms fall for an adversarial example, they’re not hallucinating—they’re seeing something that people don’t,” Louise Matsakis writes for Wired.
A study by MIT researchers examining adversarial images finds that AI systems pick up on tiny details in images that are imperceptible to the human eye, which can lead to misidentification of objects, reports Louise Matsakis for Wired. “It’s not something that the model is doing weird, it’s just that you don’t see these things that are really predictive,” says graduate student Shibani Santurkar.
Boston Herald reporter Jordan Graham writes that MIT researchers have developed an autonomous system that allows fleets of drones to navigate without GPS and could be used to help find missing hikers. “What we’re trying to do is automate the search part of the search-and-rescue problem with a fleet of drones,” explains graduate student Yulun Tian.
Graduate students Ziv Epstein and Matt Groh have developed an AI system that adds spooky figures to photos, reports Mark Wilson for Fast Company. Wilson writes that the system “works so well because it places ghostly figures exactly where your brain naturally thinks they could be–on a path in the middle of a forest, rather than, say, floating randomly through the air.”
In this Reuters video, Jim Drury highlights how MIT researchers have developed an activity simulator that could one day help teach robots how to complete household chores. The simulator, VirtualHome, could train robots to “help the elderly or disabled in their homes,” Drury explains.
CSAIL researchers have developed a technique that makes it possible to create 3-D motion sculptures from 2-D video, reports Jennifer Kite-Powell for Forbes. The new technique could “open up the possibility to study social disorders, interpersonal interactions and team dynamics,” Kite-Powell explains.
BBC Click reports on a system developed by CSAIL researchers that creates 3-D motion sculptures based off of 2-D video. The technique, say the researchers, “could help dancers and athletes learn more about how they move.”
BBC Click reports on an AI system developed by CSAIL researchers that simplifies image editing. “Instead of requiring the user to select the pixels very accurately, our system can just detect it and give the opacities for every object in the image automatically, which can then be used for editing the images in a realistic way,” explains visiting researcher Yagiz Aksoy.
CSAIL researchers have created a system that allows robots to see and pick up objects they have never encountered without assistance from humans, writes Jason Murdock for Newsweek. The researchers are now working on teaching the system to “move objects with a specific goal in mind, such as cleaning a desk,” reports Murdock.
CSAIL researchers have developed a new system that gives robots a greater visual understanding of the world around them, reports Heather Kelly for CNN. “We want robots to learn by themselves how to very richly and visually understand lots of objects that are useful for lots of tasks,” explains graduate student Pete Florence.
Wired reporter Matt Simon writes that MIT researchers have developed a new system that allows robots to be able to visually inspect and then pick up new objects, all without human guidance. Graduate student Lucas Manuelli explains that the system is “all about letting the robot supervise itself, rather than humans going in and doing annotations.”
Quanta Magazine reporter Natalie Wolchover spotlights how the work of Profs. William Freeman, Antonio Torralba and Ramesh Raskar is shedding light on how visual signals can be used to uncover information on hidden objects. Freeman explains that he is thrilled by the idea that “the world is rich with lots of things yet to be discovered.”
CSAIL researchers have created a deep learning system that can isolate individual musical instruments in a video by clicking on the specific instrument, writes Andrew Liszewski for Gizmodo. The researchers suggest the system, “could be a vital tool when it comes to remixing and remastering older performances where the original recordings no longer exist,” explains Liszewski.
In this video, BBC Click spotlights VirtualHome, a simulator developed by CSAIL researchers that could be used to teach robots to perform household chores. The researchers hope the system could one day allow for seamless human-robot collaboration by allowing robots to, “cooperate with [humans] in finishing their activity,” explains graduate student Xavier Puig.