Category: ai

The agency ultimately chose to use HireVue, an online platform that allows employers to review asynchronously recorded video interviews and have recruits play video games as part of their application process. Over the years the platform has also offered a variety of AI features to automatically score candidates. HireVue, controversially, used to offer facial analysis to predict whether an applicant would be a good fit for an open job. In recent years, research has shown that facial recognition software is racially biased. In 2019, the company’s continued use of the technique led one member of its scientific advisory board to resign. It has since stopped using facial recognition. The Markup used GovSpend, a database of procurement records for U.S. agencies at the state, local, and federal levels, to identify agencies that use HireVue. We also searched for agencies using Teramind and ActivTrak, both another kind of controversial software that allows employers to remotely monitor their workers’ browsing activities through screenshots and logs. The Markup contacted and filed public records requests with those 24 agencies to understand how they were using the software. Eleven public agencies, including the FDA, replied to The Markup with documents or confirmations that they had bought HireVue at some point since 2017. Of the six public agencies that replied to The Markup’s questions confirming that they actually used the software, all but one — Lake Travis Independent School District in Texas — confirmed they did not make use of the AI scoring features of the software. Documents and responses from 13 agencies confirmed that they purchased Teramind or ActivTrak at some point during the same time frame.

The Bucheon official said the system should reduce the strain on overworked tracing teams in a city with a population of more than 800,000 people, and help use the teams more efficiently and accurately. […] The Ministry of Science and ICT said it has no current plans to expand the project to the national level. It said the purpose of the system was to digitize some of the manual labour that contact tracers currently have to carry out. The Bucheon system can simultaneously track up to ten people in five to ten minutes, cutting the time spent on manual work that takes around half an hour to one hour to trace one person, the plan said.

Initially, the AI was fed data from a sensor that looked for a single surface-to-surface missile at an oblique angle, Simpson said. Then it was fed data from another sensor that looked for multiple missiles at a near-vertical angle. “What a surprise: the algorithm did not perform well. It actually was accurate maybe about 25 percent of the time,” he said.

That’s an example of what’s sometimes called brittle AI, which “occurs when any algorithm cannot generalize or adapt to conditions outside a narrow set of assumptions,” according to a 2020 report by researcher and former Navy aviator Missy Cummings. When the data used to train the algorithm consists of too much of one type of image or sensor data from a unique vantage point, and not enough from other vantages, distances, or conditions, you get brittleness, Cummings said. In settings like driverless-car experiments, researchers will just collect more data for training. But that can be very difficult in military settings where there might be a whole lot of data of one type — say overhead satellite or drone imagery — but very little of any other type because it wasn’t useful on the battlefield…

Simpson said the low accuracy rate of the algorithm wasn’t the most worrying part of the exercise. While the algorithm was only right 25 percent of the time, he said, “It was confident that it was right 90 percent of the time, so it was confidently wrong. And that’s not the algorithm’s fault. It’s because we fed it the wrong training data.”

In understanding knots, mathematicians rely on something called invariants, which are algebraic, geometric or numerical quantities that are the same. In this case, they looked at invariants that were the same in equivalent knots; equivalence can be defined in several ways, but knots can be considered equivalent if you can distort one into another without breaking the knot. Geometric invariants are essentially measurements of a knot’s overall shape, whereas algebraic invariants describe how the knots twist in and around each other. “Up until now, there was no proven connection between those two things,” [said Alex Davies, a machine-learning specialist at DeepMind and one of the authors of the new paper], referring to geometric and algebraic invariants. But mathematicians thought there might be some kind of relationship between the two, so the researchers decided to use DeepMind to find it. With the help of the AI program, they were able to identify a new geometric measurement, which they dubbed the “natural slope” of a knot. This measurement was mathematically related to a known algebraic invariant called the signature, which describes certain surfaces on knots.

In the second case, DeepMind took a conjecture generated by mathematicians in the late 1970s and helped reveal why that conjecture works. For 40 years, mathematicians have conjectured that it’s possible to look at a specific kind of very complex, multidimensional graph and figure out a particular kind of equation to represent it. But they haven’t quite worked out how to do it. Now, DeepMind has come closer by linking specific features of the graphs to predictions about these equations, which are called Kazhdan-Lusztig (KL) polynomials, named after the mathematicians who first proposed them. “What we were able to do is train some machine-learning models that were able to predict what the polynomial was, very accurately, from the graph,” Davies said. The team also analyzed what features of the graph DeepMind was using to make those predictions, which got them closer to a general rule about how the two map to each other. This means DeepMind has made significant progress on solving this conjecture, known as the combinatorial invariance conjecture.