Clearview AI accused over free trials to US police that were plausibly deniable
- Reference: 1618228984
- News link: https://www.theregister.co.uk/2021/04/12/in_brief_ai/
- Source link:
The data collected by BuzzFeed News [1]showed just how haphazardly the machine learning software was used. In an attempt to win customer contracts, Clearview gave out free trials to public agencies, including law enforcement and even places, like the Department of Fish and Wildlife in Washington and Minnesota’s Commerce Fraud Bureau.
[2]
Employees could apparently use the technology on whomever they wanted, whether they were trying to identify a suspect in a criminal case or students at universities. In one case that was particularly disturbing, police officers in Alameda, California continued to use Clearview's tools although the local City Council voted to ban the use of public facial recognition tools in 2019.
AI algorithms aren’t perfect, and particularly struggle with correctly identifying women and people of colour. The data has been compiled into a handy [3]searchable database .
[4]
Google AI Research manager resigns after org is reshuffled
The manager that oversaw Google’s AI ethics unit, which has seen two researchers pushed out, has resigned.
Samy Bengio, a well-known name in the academic world of machine learning, has become the most senior member of the Chocolate Factory to leave after it controversially ousted [5]Timnit Gebru and [6]Margaret Mitchell .
Although Bengio did not explicitly say why he decided to leave in an email to his colleagues, he hinted at the recent fiasco, where Google fired its Ethical AI team co-leads over a paper that was critical about massive language models.
“I learned so much with all of you, in terms of machine learning research of course, but also on how difficult yet important it is to organize a large team of researchers so as to promote long term ambitious research, exploration, rigor, diversity and inclusion,” he wrote, [7]according to Bloomberg.
Google has since [8]reshuffled the management of its AI research teams.
Intel’s AI chips are going into a new academic supercomputer
Chipzilla's own-brand machine learning chips will be used to build Voyager, a new supercomputer for the University of California, San Diego, and it's expected to be up and running later this year.
Intel has been trying to give Nvidia a run for its money by developing its own training and inference chips to challenge the GPU. But it fell woefully behind and abandoned previous attempts led by Nervana, a startup it acquired in 2016.
It later snapped up Habana, another AI hardware startup in 2019, and it was [9]out with the old and in with the new . Now, it appears some of Intel’s or, rather, Habana’s hard work is paying off.
“The Voyager supercomputer will use Habana’s unique interconnectivity technology to efficiently scale AI capacity with 336 Gaudi processors for training and 16 Habana Goya processors for AI inference,” the company [10]said in a statement.
It’s difficult to work out just how good these chips really are, however, an Intel representative declined to comment on the performance of the supercomputer or provide a detailed breakdown of the chips’ specs.
Listen to AI-generated rock music
A non-profit organisation focused on mental health and music published a series of AI-generated songs stylized after artists, who tragically died at the age of 27 by suicide or drug-related incidents.
The project, named Lost Tapes of the 27 Club, has songs based on artists like Jimi Hendrix, Janis Joplin, Kurt Cobain, Amy Winehouse, and more. The group hasn't released details of exactly how they've managed to mix the tracks, but there are some interesting mashups.
[11]
You can listen to the tracks [12]here . The non-profit, Over The Bridge, hopes that this will raise awareness of mental health issues. ®
Get our [13]Tech Resources
[1] https://www.buzzfeednews.com/article/ryanmac/clearview-ai-local-police-facial-recognition
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2YHRunLF3x3EwopidZyG8yQAAAEg&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://www.buzzfeednews.com/article/ryanmac/facial-recognition-local-police-clearview-ai-table
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33YHRunLF3x3EwopidZyG8yQAAAEg&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[5] https://www.theregister.com/2020/12/04/google_nlrb_complaint_ai_ethics/
[6] https://www.theregister.com/2021/01/25/in_brief_ai/
[7] https://www.bloomberg.com/news/articles/2021-04-06/google-ai-research-manager-samy-bengio-resigns-in-email-to-staff
[8] https://www.theregister.com/2021/02/19/google_ai_reorganization/
[9] https://www.theregister.com/2020/01/31/intel_kills_spring_crest/
[10] https://newsroom.intel.com/news/sd-supercomputer-center-selects-habana-intel-efficient-ai/
[11] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44YHRunLF3x3EwopidZyG8yQAAAEg&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[12] https://losttapesofthe27club.com/#the-album
[13] https://whitepapers.theregister.com/
Re: So why is the facial recognition software so bad for non-whites?
Because that database will tell the AI that black people are criminals, which is a problem they already have.
Not to mention the human rights of those prisoners, who may not want to be part of mammon's all seeing technopanopticon.
Re: So why is the facial recognition software so bad for non-whites?
Facial recognition algorithms do not (yet) do what they are supposed to do for a variety of reasons. They are reasonbly good at matching facial images that are taken in a similar way - e.g. a passport photo and a frontal face image taken when the person deliberately looks into a fixed camera (e.g. at an e-gate in an airport), although false negatives are quite frequent even then (false positives in that situation would be undetected), so it's best used to alert a human (e.g. immigration officer) to do a double-check. And this is the simplest case of matching a single face to a single image, not trying to find a match within a database of thousands of images, where the probability of a face having similar biometrics to at least one face in the database and flagging a false positive increases greatly in proportion to the number of faces in the database.
Matching a CCTV image to a single face in a database of mugshots has to compensate for a different face position in sub-optimal lighting, and which might have additional features such as a hat, scarf, beard, long hair, makeup, glasses, completely different facial expression, age difference between sample and real-time face and/or (especially these days) a mask that obliterates many of the key biometric measurements. Trials in the UK showed that the algorithms used were so bad that it's essentially useless. Heck, humans often have great difficulty matching faces in a large set of random photograhs that include the same person in completely different conditions. There are many stories of siblings accidentally using each others' passports and going through multiple border checks undetected.
My own experience at an airport e-gate showed that I was initially denied entry because I was smiling when looking into the camera - it worked only if I deliberately held the same sombre expression as I had in my passport photo.
"[he] did not explicitly say why he decided to leave"
Of course not. One needs to demonstrate loyalty and discretion if one wants to stay part of the team and remain a high-flyer.
Blurting out the truth would be sure-fire way of not ever getting that level of position any more.
So why is the facial recognition software so bad for non-whites?
According to the Pew Research center there were 476K black and 437K white prisoners in federal and state prisons in the United States as of the end of 2017. I presume that each of these people have at least one booking picture. With a sample size that large, why on earth is their recognition software so crap? For that matter, since you have a captive (!) sample, why not take one picture a month and get a really huge sample database?