How the tables have turned: Bloke says he trained facial recognition algorithm to identify police officers

Plus: Facebook won't get lost in translation, Cruise all set to build fleet of driverless taxis, and more


In brief An activist in Portland, Oregon, claims to have trained a facial recognition model that is able to identify one-fifth of the police working in the US city.

Christopher Howell, described as a "self-taught coder", told The New York Times that his model isn't publicly available, and that he'd used it to help a friend identify a police officer. Unsurprisingly, the most difficult part of the project is collecting enough images to amass a database large enough to accurately match photographs.

Howell said he looks for the names and faces of police officers in news articles and public websites. He then searches for them on social media and downloads their photographs, a similar technique used by the controversial facial recognition startup Clearview.

Portland has some of the strictest rules on facial recognition. Law enforcement and companies are not allowed to deploy AI cameras in any public or private places. But there's nothing stopping individuals, like Howell, developing their own algorithms.

New open-source machine translation model from Facebook

Researchers from the social media giant have published the code to their latest language model, capable of translating over 100 languages.

The most interesting thing about the system, known as M2M-100, is that it directly translates between two languages without having to map to English first.

"When translating, say, Chinese to French, previous best multilingual models train on Chinese to English and English to French, because English training data is the most widely available," Facebook explained this week.

By skipping the middle step of translating to English first, the meaning of sentences are less likely to be lost in translation. The researchers hope to keep tweaking the model so that it may one day be practically deployed on the social media platform. The code is all open-source, but beware the model is large and has over 15 billion parameters.

Cruise applies for official approval for self-driving car

San Francisco startup Cruise says it's ready to build a fleet of autonomous robo-taxi vehicles.

In January it unveiled Cruise Origins, a futuristic electric boxy-looking car with sliding doors. These vehicles contain no steering wheel or pedals for human drivers. Instead, there are just four seats to pick up passengers.

But before Cruise can begin testing its self-driving technology using these cars, it has to get explicit approval from the US National Highway Traffic Safety Administration (NHTSA) to start manufacturing the cars in the first place.

Cruise did not answer The Register's questions on when it hoped to start building its Origin cars. The NHTSA also did not tell us how long it typically takes to process such a request.

Nvidia appears to dominate latest inference hardware tests

MLPerf, an industry-led effort to benchmark the performance of AI hardware, has released its latest results testing how fast different systems and chips run specific machine learning models.

A quick glance at the results – spread over different categories like data centre, the edge, mobile phones, and mobile notebooks – shows that the most common entries come from Nvidia. Unfortunately, there were no entries from competing entities like Groq, Graphcore, Intel's Habana, and so on.

Google is also absent so this year's results don't seem all that competitive really. But there are some interesting numbers like Intel's stats for its 10nm Xe-LP Graphics chip (formally known as Tiger Lake) in mobile notebooks and from Raspberry Pi devices running Arm-based CPUs. You can see all the numbers here.

Kite has added support for 11 languages

If you're interested in a tool that autocompletes your code as you write it then consider Kite.

The startup spent years building a model that employs various machine learning methods to analyse software. It behaves like popular language models that generate text by predicting the next word in a given sentence. But instead of natural language, it's code.

It has updated the free version of its autocomplete tool and it now supports 11 languages including Java, C/C++, Python, Golang, Scala, and more. You can try it out here. ®


Biting the hand that feeds IT © 1998–2020