Moving beyond Big Data

Source: University of Maryland Language Science Center

“The National Science Foundation has announced that the University of Maryland’s Language Science Center (LSC) will (again!) receive a $3M grant for innovative research and graduate training, this time as part of the first cohort of awards made through its new NSF Research Traineeship (NRT) program. This 5-year award will support a model of interdisciplinary graduate training that prepares students to be adaptable scientists in multiple settings and career paths. The project will connect research on humans and machines, via a focus on how to succeed when Big Data is not available. The project is led by faculty and students from 10 departments across the entire university.”

“I’m excited by the research theme, which takes a “Beyond Big Data” approach,” says Colin Phillips, program PI and LSC Director. “We’re interested in how humans and machines can learn more efficiently from ‘multi-scale data’. Everybody’s talking these days about Big Data, but the current frontier in language science involves how to do more with less, you could call it ‘medium data’ or ‘small data’. It’s important for building better language technology, and it’s important for improving language learning outcomes in children and adults. Current language technologies like Google Translate and Apple’s Siri rely on a Big Data approach that stores billions of utterances. But that approach won’t generalize to the vast majority of the world’s 7000 languages. And human children easily outperform the best current technology, though they learn language from far less data. Child brains somehow learn the language around them more efficiently. But we all know that learning a new language as an adult is much harder. And we’re learning more and more about how children who experience ‘language poverty’, growing up with smaller amounts of quality language interaction, face negative consequences that last a lifetime. We want to understand why some learners fare better than others.” A key venue for exploring the program’s research goals will be yearly “Summer Camps”, intensive research-only workshops that will bring together students and faculty from UMD and beyond.”
Colin Phillips video.

Fooling Deep Neural Networks

A video summary of the paper: Nguyen, Anh, Jason Yosinski, and Jeff Clune. “Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images.” arXiv preprint arXiv:1412.1897 (2014). The paper is available here.

From MIT Technology Review: “A technique called deep learning has enabled Google and other companies to make breakthroughs in getting computers to understand the content of photos. Now researchers at Cornell University and the University of Wyoming have shown how to make images that fool such software into seeing things that aren’t there. The researchers can create images that appear to a human as scrambled nonsense or simple geometric patterns, but are identified by the software as an everyday object such as a school bus. The trick images offer new insight into the differences between how real brains and the simple simulated neurons used in deep learning process images” (December 24, 2014).

Connections: Learning everything about anything. Also: Google researchers have developed software that can match complex images with simple sentences describing whole scenes, rather than just objects, e.g. “a group of young people playing a game of frisbee.”  

Radhika Nagpal: One of 10 people who mattered this year in science

Radhika-Nagpal

Source: Reflection Films

From Nature, volume 516, issue 7531, December 17 2014.
“When Radhika Nagpal was a high-school student in India, she hated biology: it was the subject that girls were supposed to study so that they could become doctors. Never being one to follow tradition, Nagpal was determined to become an engineer. Now she is — leading an engineering research team at Harvard University in Cambridge, Massachusetts. But she also has a new appreciation for the subject she once disliked. This year, her group garnered great acclaim for passing a milestone in biology-inspired robotics. Taking their cue from the way in which ants, bees and termites build complex nests and other structures with no central direction, Nagpal’s group devised a swarm of 1,024 very simple ‘Kilobots’. Each Kilobot was just a few centimetres wide and tall, moved by shuffling about on three spindly legs and communicated with its immediate neighbours using infrared light. But the team showed that when the Kilobots worked together, they could organize themselves into stars and other two-dimensional shapes.”
Earlier entry: Inferring simple rules from complex structures.

Subash Khot: A grand vision for the impossible

Sources: Quanta Magazine and Simons Foundation

Subash Khot is this year’s winner of the Nevanlinna Prize. If his “Unique Games Conjecture is correct, then for many of the problems people would most like to solve, it’s hard not only to find an exact solution—finding even a good approximation of the solution is beyond the reach of a computer. This conjecture may seem, on the face of it, like a pretty parochial statement (though possibly hard to prove, or even false). Over the past decade, however, computer scientists have found to their surprise that if the conjecture is true, then a host of problems—some of them seeming to bear little resemblance to the original Unique Games problem—are also hard to approximate.” Simons Foundation

http://www.simonsfoundation.org/quanta/20140812-a-grand-vision-for-the-impossible/

Source: Quanta Magazine

“For the problem of coloring the nodes of a network that has a collection of constraints about which color combinations are allowed (top left), it is sometimes possible to find a coloring that satisfies all the constraints (top right). But for some networks and constraints (bottom left), it is impossible to satisfy all the constraints simultaneously. The Unique Games Conjecture concerns the problem of finding a coloring that satisfies as many constraints as possible, such as the bottom right coloring, which satisfies all the constraints except the one for the thick edge.” Quanta Magazine.