Archivio di tutti i clip:
(Notebook di Evernote).
Software that swaps out words can now fool the AI behind Alexa and SiriThe news: Software called TextFooler can trick natural-language processing (NLP) systems into misunderstanding text just by replacing certain words in a sentence with synonyms. In tests, it was able to drop the accuracy of three state-of-the-art NLP systems dramatically. For example, Google’s powerful BERT neural net was worse by a factor of five to seven at identifying whether reviews on Yelp were positive or negative.How it works: The software, developed by a team at MIT, looks for the words in a sentence that are most important to an NLP classifier and replaces them with a synonym that a human would find natural. For example, changing the sentence “The characters, cast in impossibly contrived situations, are totally estranged from reality” to “The characters, cast in impossibly engineered circumstances, are fully estranged from reality” makes no real difference to how we read it. But the tweaks made an AI interpret the sentences completely differently. Why it matters: We have seen many examples of such adversarial attacks, most often with image recognition systems, where tiny alterations to the input can flummox an AI and make it misclassify what it sees. TextFooler shows that this style of attack also breaks NLP, the AI behind virtual assistants—such as Siri, Alexa and Google Home—as well as other language classifiers like spam filters and hate-speech detectors. The researchers say that tools like TextFooler can help make NLP systems more robust, by revealing their weaknesses.ShareLinkTaggedNeural NetworksAuthorDouglas HeavenImageJon Fife | FlickrShareLinkTaggedNeural NetworksAuthorDouglas HeavenImageJon Fife | FlickrComputingFeb 8A dark web tycoon pleads guilty. But how was he caught?The FBI found Eric Marques by breaking the famed anonymity service Tor, and officials won’t reveal if a vulnerability was used. That has activists and lawyers concerned.Read moreSpaceFeb 7NASA: A second, unreported glitch could have wrecked Boeing’s StarlinerNASA safety officials have revealed that the Boeing Starliner’s failed attempt to dock with the International Space Station last December—the vehicle’s first ever flight into space—was marked by…The first glitch: We already knew that Starliner’s December 20 mission, an uncrewed test flight, ended early when the spacecraft’s engines did not burn as intended. This was due to a faulty internal timer, as previously reported. The issue meant it was unable to rendezvous with the ISS, and returned to Earth just 48 hours later. NASA now says a coding error caused Starliner to pull its time from the mission’s Atlas V launch vehicle, creating an 11-hour mismatch. The second glitch: But now NASA has revealed there was a second coding error, this time in the sequence that governs Starliner’s preparation for reentry into Earth’s atmosphere. If it had not been addressed, Starliner would have erroneously fired its thrusters, resulting in uncontrolled motions that could have made the service module collide with the crew module—potentially damaging the crew module’s heat shield, or causing it to tumble dangerously during descent.The glitch was caught and fixed by ground crews the night before the spacecraft returned to Earth. Jim Chilton, a senior vice president at Boeing, says the glitch wouldn’t have been found if the timer error hadn’t prompted mission control to look for more faults in Starliner’s software.Rising concerns: Neither NASA nor Boeing had given any previous indication of more than one issue, although both organizations became aware of the problems during the mission. The coding errors should have been noticed in multiple checks conduced by both parties, officials said. “Our NASA oversight was insufficient,” says Doug Loverro, the head of human spaceflight at NASA.What’s next: NASA’s independent review panel is recommending a complete review of Boeing’s software verification process, and Boeing plans to re-review all of Starliner’s software—more than one million lines of code. The panel is still finishing its own review of the mission and craft. NASA has still not given any indication of how these new revelations will affect Starliner’s 2020 time line, and whether it will require another uncrewed test flight to the ISS before Starliner can take people into space. A decision is expected by the end of February.ExpandEmTech Digital 2020Building AI you can trustAt MIT Technology Review’s AI conference, come find ideas, inspiration, and practical guidance on deploying AI in organizations in 2020.01.AI strategy studioAre you confident in your AI strategy? Start with an expert-driven workshop full of real-life case studies on how companies have implemented AI.02.Are you AI ready?Trust is the theme for this year’s event: algorithms we can trust, data we can trust, decisions we can trust. 03.Artificial common sense: what it takes to get AI to actually understandIs artificial general intelligence possible? How far off is it? And what’s standing in the way?04.Solving biasSome practical approaches to dealing with problems of bias in algorithms and data.SponsoredHow artificial intelligence is making health care more humanMIT Technology Review Insights surveyed more than 900 health-care professionals on their use of artificial intelligence today. This new report explores the major findings.Read moreBlockchainFeb 7A crypto project to make internet names censorship-proof is now liveThe Handshake Network, an ambitious public blockchain project whose developers want to reinvent how internet domain names are assigned, has finally launched its main network after the project was…Meet the DNS: When you enter a website name into your browser, you make a request of a network of computers called the domain name system, or DNS, which keeps track of all the names on the internet. The DNS converts the text name you entered (e.g., www.technologyreview.com) into a string of numbers called an IP address. This number lets your browser locate and connect to the server for the website you are trying to visit.ICANN haz centralized authority: The DNS is a hierarchical global network, and at the top of the hierarchy is the so-called DNS root. A Los Angeles–based nonprofit called the Internet Corporation for Assigned Names and Numbers (ICANN) oversees the DNS root and is responsible for allocating new “top-level domains,” which include .com, .org, .net, and most two-letter country codes. Internet freedom advocates have argued that relying on a single organization to do this makes the internet more vulnerable to censorship and hacking. Governments have been known to use the DNS to block access to certain sites. The Handshake Network’s backers say control over the root can be decentralized, using a blockchain. Bitcoin-esque: Handshake’s network will be similar to the Bitcoin network. Computers will compete to add new transactions to the blockchain and earn cryptocurrency. But the blockchain will also keep track of registered domain names, and the top 100,000 of the internet’s most popular names are already in the chain. If a name isn’t on the blockchain, the software will redirect your request to regular DNS servers, Steven McKie, a developer and investor in the project, told me in June. I’d like to try. How do I do it? It’s possible to change your computer’s DNS settings to point to a publicly available Handshake name resolver and start using it to look up names today. You won’t need cryptocurrency or any special software to do that. You can also participate more directly in the network by installing and running the Handshake software, a “light” version of which can be embedded in your browser. To register a name, you’ll need to participate in an online auction for it using the network’s cryptocurrency, called HNS, which you can buy here.Now that it’s built, will they come? To work, Handshake will need to build up a large community of “miners” willing to run the software in pursuit of new coins, entice developers to build applications on top of the network, and convince regular users to switch from the traditional DNS. Can the project succeed where a number of other blockchain-based internet naming systems have already failed? Now that it’s finally in the wild, we’re going to find out.Keep up with the fast-moving and sometimes baffling world of cryptocurrencies and blockchains with our weekly newsletter Chain Letter. Subscribe here. It’s free!ShareLinkImageChris Liverani on UnsplashExpandCRISPRFeb 7The first US trial of CRISPR gene editing in cancer patients suggests the technique is safeThe news: Three people with advanced cancer have received injections of immune cells gene-edited using CRISPR without any serious side effects. It’s the first US clinical trial of the technique and…The trial: The aim was solely to discern if the technique was safe and feasible, not to find a cure. On that basis, it was a success. The three patients, all in their 60s, all had tumors that hadn’t responded to other treatments and that produced the same protein. Last year, they were each given a dose of CRISPR-altered versions of their own T cells, tweaked to make them more efficient at recognizing that protein and killing the cancer. These cells successfully integrated with each patient’s immune system and were still found in their blood nine months later. The results of the trial were published in Science.Dangers: One concern was that the introduction of the cells might trigger an immune response because the protein used in the CRISPR process is derived from bacteria. That didn’t happen. A second worry was that CRISPR gene editing creates so-called off-target effects in which unintended deletions happen in cells, which could cause them to turn cancerous. Some such changes were seen, but nothing serious.What about the patients? Although the patients saw their conditions stabilize during treatment, and one had a reduction in tumor size, the cell injection did not work long term. One has since died; the other two have seen their cancer worsen and are now being given other treatments. What’s next: The trial won’t continue because the technique it uses is outdated (gene-editing technology has progressed rapidly since 2016, when the study was approved), but the encouraging results will pave the way for many other tests of gene-editing technology in patients. Fyodor Urnov, at the University of California, Berkeley, who was not involved in the work, told Science that the study answers “questions that have frankly haunted the field.”ShareLinkImageAPExpandHe Jiankui faces three years in prison for CRISPR babiesHere are some actual facts about George Church’s DNA dating companySign up for The Download — your daily dose of what’s up in emerging technologyAlso stay updated on MIT Technology Review initiatives and events?YesNoSponsoredSupercharging with converged infrastructureAs analytics becomes core to decision-making, companies are pushing the boundaries of their IT systems to take advantage of data-driven technologies.Read more