Researchers at @AuburnEngineers and @AdobeResearch tried to get an natural language processing system to generate explanations for its behavior, such as why it claimed different sentences meant the same thing.
When they tested their approach they realized that shuffling words in a sentence made no difference to the explanations.
For example, the systems correctly spotted that the sentences “Does marijuana cause cancer?” and “How can smoking marijuana give you lung cancer?” were paraphrases.
But they were even more certain that “You smoking cancer how marijuana lung can give?” and “Lung can give marijuana smoking how you cancer?” meant the same thing too.
“This is a general problem to all [natural-language processing] models,” says @anh_ng8 at Auburn University, who led the work.
The models appear to pick up on a few key words in a sentence, whatever order they come in. They do not understand language as we do.
Moreover, GLUE—a very popular benchmark—does not measure true language use and teaches NLP models to jump through hoops. Many researchers have started to use a harder set of tests called SuperGLUE, but Nguyen suspects it will have similar problems.
The good news is that it might not be too hard to fix.
The researchers found that forcing a model to focus on word order, by training it to do a task where word order mattered, such as spotting grammatical errors, also made the model perform better on other tasks.
These results are yet another example of how artificial intelligence often falls far short of what people believe its capable of. technologyreview.com/2020/11/18/101…
There’s been a lot of discussion since the Capitol insurrection about the role social media companies, their site designs, and their business models have played in the spread of online conspiracy theories. This is a thread about whether what’s being done goes far enough.
First, let’s start with President Trump, the unapologetic, chief author of the false claim that the election had been stolen from him. That conspiracy theory was an inspiration for the mob at the Capitol.
Last week, social media platforms finally blocked the president’s accounts. Previous cases show that deplatforming works... but only if the costs are high enough. technologyreview.com/2021/01/08/101…
That not only means that countries face a huge logistical challenge to distribute them—which is complicated by the fact the two most promising vaccines require ultra-cold temperatures—but they also have to grapple with hard choices over who gets them first.
Here’s how the US, China, the UK, and other countries are planning to distribute covid-19 vaccines to their populations. technologyreview.com/2020/12/04/101…
For the last session of #CyberSecureMIT, we’re speaking with @JamilFarshchi, Equifax’s CISO who was brought on after its data was breached in 2017. The Equifax hack was one of the biggest thefts of sensitive personal information of all time. technologyreview.com/2020/02/10/349…
“In the security industry today, we don’t have enough data to measure risk," says Farshchi. Most organizations have a dataset of one, which is their company. #CyberSecureMIT
To build a cyber-resilient organization, he says he asks:
-What are the predominant threat factors for any organization?
-What are the core controls that help you to be able to defend and minimize a particular threat?
Bilateral agreements and global accords can be the beginnings of an international counter-strategy to cyberattacks, says Choucri. #CyberSecureMIT#TechReviewEvents